AI and science frontiers · 2026-11-13

Policy optimization in language-model post-training starts from a capable base model, yet REINFORCE, the basis of PPO and GRPO, can drive the probability of optimal responses toward zero even as reward improves, and convergence can take exponentially many steps even with exact gradients and KL regularization to the base model. Periodically resetting the KL reference to the most recent policy, formalized as Approximate Policy Mirror Descent (APMD), lets each subproblem be solved locally under strong regularization while successive resets accumulate into global optimality. For softmax linear policies, this yields the first global convergence and coverage-preserving guarantees for PPO and GRPO variants with reference resets and for Kimi-style squared-loss algorithms.

Policy optimization in language-model post-training starts from a capable base model, yet REINFORCE, the basis of PPO and GRPO, can drive the probability of optimal responses toward zero even as reward improves, and convergence can take exponentially many steps even with exact gradients and KL regularization to the base model. Periodically resetting the KL reference to the most recent policy, formalized as Approximate Policy Mirror Descent (APMD), lets each subproblem be solved locally under strong regularization while successive resets accumulate into global optimality. For softmax linear policies, this yields the first global convergence and coverage-preserving guarantees for PPO and GRPO variants with reference resets and for Kimi-style squared-loss algorithms.

Robot manipulation has relied on specialist policies trained on demonstrations, such as vision-language-action models (VLAs) and world-action models (WAMs). AMP1 keeps a generalist AI model in the control loop and has it command arm motion through code, reaching 85.6% zero-shot success on RoboLab-120 against 49.2% and 42.9% for leading VLA and WAM policies. It succeeded in 104 of 108 real-world trials across twelve tasks and three robot instances, with no demonstrations or task-specific training.

Robot manipulation has relied on specialist policies trained on demonstrations, such as vision-language-action models (VLAs) and world-action models (WAMs). AMP1 keeps a generalist AI model in the control loop and has it command arm motion through code, reaching 85.6% zero-shot success on RoboLab-120 against 49.2% and 42.9% for leading VLA and WAM policies. It succeeded in 104 of 108 real-world trials across twelve tasks and three robot instances, with no demonstrations or task-specific training.

A team at Seoul National University Hospital released CARMEN, a single encoder pretrained without labels on ten cardiorespiratory waveforms totaling about 2.27 million modality-hours, which accepts any combination of co-recorded signals and inputs from 10 s to 5 min. With the encoder frozen and only a linear layer trained, it was significantly better than end-to-end supervised models on 12 of the 15 tasks that had supervised comparators, and better than three pretrained foundation models on 14 to 17 of 19 tasks, with none exceeding it.

A team at Seoul National University Hospital released CARMEN, a single encoder pretrained without labels on ten cardiorespiratory waveforms totaling about 2.27 million modality-hours, which accepts any combination of co-recorded signals and inputs from 10 s to 5 min. With the encoder frozen and only a linear layer trained, it was significantly better than end-to-end supervised models on 12 of the 15 tasks that had supervised comparators, and better than three pretrained foundation models on 14 to 17 of 19 tasks, with none exceeding it.

EEG foundation models transfer poorly as frozen feature extractors, so SeeEEG takes another route: it renders EEG as images using the montage, filtering and contrast settings of a clinical EEG viewer and trains only those display parameters, leaving the vision backbone frozen. With a frozen DINOv3 ViT-B as encoder, SeeEEG exceeds seven EEG foundation models in balanced accuracy on all eight clinical, sleep, affective and brain-computer interface datasets, reaching a mean rank of 1.3 by Cohen or AUROC against 2.9 for the strongest baseline, REVE.

EEG foundation models transfer poorly as frozen feature extractors, so SeeEEG takes another route: it renders EEG as images using the montage, filtering and contrast settings of a clinical EEG viewer and trains only those display parameters, leaving the vision backbone frozen. With a frozen DINOv3 ViT-B as encoder, SeeEEG exceeds seven EEG foundation models in balanced accuracy on all eight clinical, sleep, affective and brain-computer interface datasets, reaching a mean rank of 1.3 by Cohen or AUROC against 2.9 for the strongest baseline, REVE.

Online reinforcement-learning adaptation of long-chunk robot policies previously optimized high-dimensional action sequences directly, which demanded many samples and made real-robot adaptation costly. ExRL instead adapts a scalar decision, how many steps of each predicted action chunk to execute, training a lightweight adapter with off-policy reinforcement learning while the base policy stays frozen. Across simulation benchmarks and four real manipulation tasks it improves diffusion policies, ACT, several vision-language-action models and a world action model with limited online interaction, approaching perfect success on real cube stacking after about ten minutes of physical interaction.

Online reinforcement-learning adaptation of long-chunk robot policies previously optimized high-dimensional action sequences directly, which demanded many samples and made real-robot adaptation costly. ExRL instead adapts a scalar decision, how many steps of each predicted action chunk to execute, training a lightweight adapter with off-policy reinforcement learning while the base policy stays frozen. Across simulation benchmarks and four real manipulation tasks it improves diffusion policies, ACT, several vision-language-action models and a world action model with limited online interaction, approaching perfect success on real cube stacking after about ten minutes of physical interaction.

Linear atomic cluster expansion (ACE) potentials are misspecified when trained on deterministic density-functional labels, so Bayesian posterior variances understate their errors. A delete-one-configuration CR3 jackknife covariance supplies the shape and one conformal quantile from held-out data supplies the scale, bringing per-atom force regions close to nominal coverage on held-out silicon and propagating the same parameter distribution through implicit differentiation to 21 silicon properties; of 20 scored properties, 17 fall inside 95% intervals for the smaller model and 15 for the larger, whose intervals are about 4 times tighter.

Linear atomic cluster expansion (ACE) potentials are misspecified when trained on deterministic density-functional labels, so Bayesian posterior variances understate their errors. A delete-one-configuration CR3 jackknife covariance supplies the shape and one conformal quantile from held-out data supplies the scale, bringing per-atom force regions close to nominal coverage on held-out silicon and propagating the same parameter distribution through implicit differentiation to 21 silicon properties; of 20 scored properties, 17 fall inside 95% intervals for the smaller model and 15 for the larger, whose intervals are about 4 times tighter.

A signature proves who published a tool, not that the tool will not leak: a validly signed, fully functional MCP tool can fold sensitive items from earlier calls into a request to a policy-allowed destination, and Meta's MCPGuard-Dynamic failed to stop it under all seven configurations. Germline execution instead leaves tool code unexamined and runs each call in disposable copies forked from a data-free germline process, each receiving only the data its destination may see. No unauthorised disclosure was detected across 153 attack runs, against 5.9–74.5% for baseline defences and an ablation, and single-copy calls took 10–49 ms.

A signature proves who published a tool, not that the tool will not leak: a validly signed, fully functional MCP tool can fold sensitive items from earlier calls into a request to a policy-allowed destination, and Meta's MCPGuard-Dynamic failed to stop it under all seven configurations. Germline execution instead leaves tool code unexamined and runs each call in disposable copies forked from a data-free germline process, each receiving only the data its destination may see. No unauthorised disclosure was detected across 153 attack runs, against 5.9–74.5% for baseline defences and an ablation, and single-copy calls took 10–49 ms.

Classical neural scaling laws write excess risk as a power law in capacity and data and assume the compute-optimal allocation exponent stays fixed across scales. This analysis uses Mellin–Barnes contour integration to split finite-scale departures along the capacity axis into a three-term expansion, reduces multi-domain mixing to scalar water-filling via KKT duality, and shows the allocation exponent drifts in proportion to capacity and data exponent variances. Across 45 operator spectra, Pythia, Chinchilla's 245 runs and RegMix's 17 domains, the estimators give holdout error and cut high-compute allocation regret.

Classical neural scaling laws write excess risk as a power law in capacity and data and assume the compute-optimal allocation exponent stays fixed across scales. This analysis uses Mellin–Barnes contour integration to split finite-scale departures along the capacity axis into a three-term expansion, reduces multi-domain mixing to scalar water-filling via KKT duality, and shows the allocation exponent drifts in proportion to capacity and data exponent variances. Across 45 operator spectra, Pythia, Chinchilla's 245 runs and RegMix's 17 domains, the estimators give holdout error and cut high-compute allocation regret.

Track reconstruction at the High-Luminosity LHC usually keeps only charged particles above a transverse-momentum threshold, and that cut discards more than ninety percent of the particles. AlphaHits-TrackFormer instead processes all pixel and strip hits in complete TrackML events without a transverse-momentum selection on target tracks, reaching a TrackML score of 0.959 on 820 test events. The same framework, pretrained on tracking and then jointly fine-tuned, uses a shared hit-level representation to output tracks, jets, jet flavour and event-level observables together, with a Higgs classification AUC of 0.944.

Track reconstruction at the High-Luminosity LHC usually keeps only charged particles above a transverse-momentum threshold, and that cut discards more than ninety percent of the particles. AlphaHits-TrackFormer instead processes all pixel and strip hits in complete TrackML events without a transverse-momentum selection on target tracks, reaching a TrackML score of 0.959 on 820 test events. The same framework, pretrained on tracking and then jointly fine-tuned, uses a shared hit-level representation to output tracks, jets, jet flavour and event-level observables together, with a Higgs classification AUC of 0.944.

RawMod, released by a University of Maryland-led team, stacks the raw nanopore signal of many reads at each site and compares it base by base with the expected signal of the reference, so a single 0.23M-parameter model can say whether a site carries a modification without naming its type. Across 15 datasets spanning 5mC, 5hmC, 4mC, 6mA and 5hmU in bacterial, phage, plant, mouse, human and synthetic DNA, it reaches a mean AUROC of 0.964, against 0.799 for the best released model, Dorado 4mC+5mC. It needs no unmodified control from the same sample, but a modification type absent from training transfers zero-shot only when another modification of the same base was seen.

RawMod, released by a University of Maryland-led team, stacks the raw nanopore signal of many reads at each site and compares it base by base with the expected signal of the reference, so a single 0.23M-parameter model can say whether a site carries a modification without naming its type. Across 15 datasets spanning 5mC, 5hmC, 4mC, 6mA and 5hmU in bacterial, phage, plant, mouse, human and synthetic DNA, it reaches a mean AUROC of 0.964, against 0.799 for the best released model, Dorado 4mC+5mC. It needs no unmodified control from the same sample, but a modification type absent from training transfers zero-shot only when another modification of the same base was seen.

Regional air-quality forecasting has relied on online-coupled numerical chemistry-weather models whose skill is sensitive to meteorological initial conditions, emission inventories and lateral boundary conditions. CWAI treats meteorology, particles and reactive gases as one joint atmospheric state, driven by two consecutive states, emissions and static surface descriptors through a hierarchical Swin Transformer, and rolls out 40 six-hour steps to a 10-day forecast. Against WRF-Chem over 62 January and July 2020 starts, CWAI had lower initialization-weighted RMSE for all six surface pollutants, from 15.8% for PM2.5 to 25.8% for NO2.

Regional air-quality forecasting has relied on online-coupled numerical chemistry-weather models whose skill is sensitive to meteorological initial conditions, emission inventories and lateral boundary conditions. CWAI treats meteorology, particles and reactive gases as one joint atmospheric state, driven by two consecutive states, emissions and static surface descriptors through a hierarchical Swin Transformer, and rolls out 40 six-hour steps to a 10-day forecast. Against WRF-Chem over 62 January and July 2020 starts, CWAI had lower initialization-weighted RMSE for all six surface pollutants, from 15.8% for PM2.5 to 25.8% for NO2.

On the Aharonov-Casher continuum model, neural quantum state variational Monte Carlo accommodates both fractional Chern insulators and generalized Wigner crystals and yields phase diagrams at fractional filling for up to 36 electrons with strong band mixing. Two wave functions, a flux-attachment form and a Slater-Jastrow form, describe topological order and crystalline order respectively, and their energies match multiband exact diagonalization on small systems. As the quantum geometry becomes more nonuniform, the Laughlin-like state at 1/3 filling keeps nearly constant energy while the Jain-like state at 2/5 filling drops, so the two respond differently to discrete translation symmetry.

On the Aharonov-Casher continuum model, neural quantum state variational Monte Carlo accommodates both fractional Chern insulators and generalized Wigner crystals and yields phase diagrams at fractional filling for up to 36 electrons with strong band mixing. Two wave functions, a flux-attachment form and a Slater-Jastrow form, describe topological order and crystalline order respectively, and their energies match multiband exact diagonalization on small systems. As the quantum geometry becomes more nonuniform, the Laughlin-like state at 1/3 filling keeps nearly constant energy while the Jain-like state at 2/5 filling drops, so the two respond differently to discrete translation symmetry.

For monotone dynamical systems with unknown transition maps, safety verification previously relied on dense sampling or Lipschitz bounds, with sample counts growing exponentially in dimension. Max-Monotone Networks build monotonicity into the network structure, reducing global safety conditions to checks over finitely many boundary cells, so a zero training loss yields a formally valid barrier certificate. The method certifies a 9,000-state Kuramoto oscillator network and a 1,000-state freeway traffic network, where two baselines fail to converge or exceed memory limits.

For monotone dynamical systems with unknown transition maps, safety verification previously relied on dense sampling or Lipschitz bounds, with sample counts growing exponentially in dimension. Max-Monotone Networks build monotonicity into the network structure, reducing global safety conditions to checks over finitely many boundary cells, so a zero training loss yields a formally valid barrier certificate. The method certifies a 9,000-state Kuramoto oscillator network and a 1,000-state freeway traffic network, where two baselines fail to converge or exceed memory limits.

BioXArena-Pro replaces biomedical modeling evaluation with 64 research-level tasks rebuilt from published studies, most holding out an entire biological unit such as a species, cohort or cell line, and audited in several rounds to close label leaks. Across 22 agent configurations the best scores 55.1 of 100 and no task saturates; on the nine tasks close to the original studies the strongest agent ties the reproduced published methods (54.9 versus 54.7), while on the 48 rebuilt tasks the methods trail by 8.2 points.

BioXArena-Pro replaces biomedical modeling evaluation with 64 research-level tasks rebuilt from published studies, most holding out an entire biological unit such as a species, cohort or cell line, and audited in several rounds to close label leaks. Across 22 agent configurations the best scores 55.1 of 100 and no task saturates; on the nine tasks close to the original studies the strongest agent ties the reproduced published methods (54.9 versus 54.7), while on the 48 rebuilt tasks the methods trail by 8.2 points.

f-LENNS super-resolves conventional-resolution fMRI magnitude images to 1 mm isotropic, letting mesoscale functional organization be recovered without an ultra-high-field scanner. The model is trained only on domain-randomized synthetic MRI, with no adversarial or generative components, so it applies to experimental data across field strengths, sequences and resolutions without retraining. In visual cortex, color-, motion- and disparity-selective sites localized from super-resolved maps came closer to the original high-resolution acquisitions than interpolation upsampling and sharpening did.

f-LENNS super-resolves conventional-resolution fMRI magnitude images to 1 mm isotropic, letting mesoscale functional organization be recovered without an ultra-high-field scanner. The model is trained only on domain-randomized synthetic MRI, with no adversarial or generative components, so it applies to experimental data across field strengths, sequences and resolutions without retraining. In visual cortex, color-, motion- and disparity-selective sites localized from super-resolved maps came closer to the original high-resolution acquisitions than interpolation upsampling and sharpening did.

Existing secure fine-tuning protocols hand the data owner hidden representations or logits from every step, and from those the data owner can copy the provider's LoRA adapter: from the hidden representations of a single mini-batch of 32 samples, the copy agrees with the original on 99.2% of test inputs, and logits alone give about 94%. VeilTune instead runs the forward pass, backpropagation and the Adam update entirely in secure multi-party computation and reveals only the trained adapter to the model owner, averaging 5.3 points above SecP-Tuning on four few-shot tasks with RoBERTa-large and staying within 0.3 to 1.7 points of plaintext training.

Existing secure fine-tuning protocols hand the data owner hidden representations or logits from every step, and from those the data owner can copy the provider's LoRA adapter: from the hidden representations of a single mini-batch of 32 samples, the copy agrees with the original on 99.2% of test inputs, and logits alone give about 94%. VeilTune instead runs the forward pass, backpropagation and the Adam update entirely in secure multi-party computation and reveals only the trained adapter to the model owner, averaging 5.3 points above SecP-Tuning on four few-shot tasks with RoBERTa-large and staying within 0.3 to 1.7 points of plaintext training.

Fully homomorphic encryption lets a cloud evaluate a Transformer directly on ciphertext, but rotations and format conversions have kept it slow. MOAI chains column packing and diagonal packing into one fixed flow across all layers and evaluates Softmax and LayerNorm without rotations, needing only 9,648 homomorphic rotations across all matrix multiplications of BERT-base. On a single A100 GPU its end-to-end evaluation time is 52.8% lower than THOR, amortized to 2.36 minutes per input.

Fully homomorphic encryption lets a cloud evaluate a Transformer directly on ciphertext, but rotations and format conversions have kept it slow. MOAI chains column packing and diagonal packing into one fixed flow across all layers and evaluates Softmax and LayerNorm without rotations, needing only 9,648 homomorphic rotations across all matrix multiplications of BERT-base. On a single A100 GPU its end-to-end evaluation time is 52.8% lower than THOR, amortized to 2.36 minutes per input.

Shanghai Jiao Tong University and collaborators assembled a roughly 47,000-hour multisource pool and extended a pretrained video model's language–video priors into language–video–action learning, yielding the world action model SWAM. On the 50 manipulation tasks of RoboTwin 2.0, SWAM reaches 61.56% success under the randomized Hard setting, 12.86 percentage points above the strongest baseline OpenWAM, and 88.10% under the clean Easy setting, slightly below OpenWAM's 89.30%. On a real bimanual platform, its three-task average Hard success rate is 40.00% against OpenWAM's 23.33%.

Shanghai Jiao Tong University and collaborators assembled a roughly 47,000-hour multisource pool and extended a pretrained video model's language–video priors into language–video–action learning, yielding the world action model SWAM. On the 50 manipulation tasks of RoboTwin 2.0, SWAM reaches 61.56% success under the randomized Hard setting, 12.86 percentage points above the strongest baseline OpenWAM, and 88.10% under the clean Easy setting, slightly below OpenWAM's 89.30%. On a real bimanual platform, its three-task average Hard success rate is 40.00% against OpenWAM's 23.33%.

BridgeKV lets a large model in a model family run prefill once, then a lightweight KV-stitching module translates the large model's prompt cache into the small model's native cache layout, so every output token is generated by the small model alone. Across six model pairs and five datasets, it scores above fine-tuned small-model baselines with matched trainable parameter budgets in all 30 settings, closing up to 80% of the quality gap to running the large model for both phases. On one A40, Llama-3.1-Instruct-8B→Llama-3.2-Instruct-1B serves a 1K-token prompt with 256 output tokens in 2.96 s, close to the 2.80 s of the small model alone and 3.0× faster than the 8.83 s of the large model alone.

BridgeKV lets a large model in a model family run prefill once, then a lightweight KV-stitching module translates the large model's prompt cache into the small model's native cache layout, so every output token is generated by the small model alone. Across six model pairs and five datasets, it scores above fine-tuned small-model baselines with matched trainable parameter budgets in all 30 settings, closing up to 80% of the quality gap to running the large model for both phases. On one A40, Llama-3.1-Instruct-8B→Llama-3.2-Instruct-1B serves a 1K-token prompt with 256 output tokens in 2.96 s, close to the 2.80 s of the small model alone and 3.0× faster than the 8.83 s of the large model alone.

Stochastic diffusion samplers have lagged deterministic ones, and raising approximation order does not necessarily improve their strong convergence order. F-DDPM evaluates the data-prediction model at a cleaner lookahead point built from the previous prediction and the previously injected noise, still with one model evaluation per iteration and a first-order strong convergence guarantee. On CIFAR-10, ImageNet, FFHQ and LSUN it reaches lower FID than the evaluated stochastic baselines at matched NFE budgets, and with 1.5 times the NFE it approaches or beats deterministic baselines in many settings.

Stochastic diffusion samplers have lagged deterministic ones, and raising approximation order does not necessarily improve their strong convergence order. F-DDPM evaluates the data-prediction model at a cleaner lookahead point built from the previous prediction and the previously injected noise, still with one model evaluation per iteration and a first-order strong convergence guarantee. On CIFAR-10, ImageNet, FFHQ and LSUN it reaches lower FID than the evaluated stochastic baselines at matched NFE budgets, and with 1.5 times the NFE it approaches or beats deterministic baselines in many settings.

AudioSphere extends self-supervised audio pretraining from single-channel to multichannel: it takes four-channel Ambisonics recordings and trains a masked autoencoder to reconstruct both spectrograms and intensity vectors, so a frozen backbone with a shallow probe performs sound event detection and localization in real scenes. Prior single-channel models were blind to spatial structure, while spatial models relied on clip-level supervision. It reaches an aggregate HEAR score of 83.2 and also supports frame-level localization.

AudioSphere extends self-supervised audio pretraining from single-channel to multichannel: it takes four-channel Ambisonics recordings and trains a masked autoencoder to reconstruct both spectrograms and intensity vectors, so a frozen backbone with a shallow probe performs sound event detection and localization in real scenes. Prior single-channel models were blind to spatial structure, while spatial models relied on clip-level supervision. It reaches an aggregate HEAR score of 83.2 and also supports frame-level localization.

A trained physics surrogate can be made to respect a known symmetry at test time by transforming the input over the symmetry group, transforming the prediction back, and averaging, without changing the architecture or retraining. For compact groups this averaging carries a known non-increasing-risk guarantee and lowered errors on several The Well surrogates; for non-compact groups such as unit rescalings and the Lorentz group, normalized averaging no longer exists, and the Weyl trick and frame averaging gave clear gains mainly out of distribution or in low-data regimes, while models trained directly on invariant features matched or beat them.

A trained physics surrogate can be made to respect a known symmetry at test time by transforming the input over the symmetry group, transforming the prediction back, and averaging, without changing the architecture or retraining. For compact groups this averaging carries a known non-increasing-risk guarantee and lowered errors on several The Well surrogates; for non-compact groups such as unit rescalings and the Lorentz group, normalized averaging no longer exists, and the Weyl trick and frame averaging gave clear gains mainly out of distribution or in low-data regimes, while models trained directly on invariant features matched or beat them.

Hierarchical Joint-Embedding Predictive Architectures (H-JEPA) previously planned by running gradient descent on actions at test time, and the iterative solver consumed 96.6%–97.7% of episode runtime on long-horizon tasks. WARP-JEPA keeps the world models frozen and replaces online optimization with learned multi-candidate trajectory proposers, scoring candidates with the frozen predictor inside the generation loop. Across 30 frozen two-, three- and four-level models on four benchmarks, it speeds long-horizon episode inference by 9.6–18.8× (2.9–3.0× on Cube) and raises task success by up to 33.3 percentage points.

Hierarchical Joint-Embedding Predictive Architectures (H-JEPA) previously planned by running gradient descent on actions at test time, and the iterative solver consumed 96.6%–97.7% of episode runtime on long-horizon tasks. WARP-JEPA keeps the world models frozen and replaces online optimization with learned multi-candidate trajectory proposers, scoring candidates with the frozen predictor inside the generation loop. Across 30 frozen two-, three- and four-level models on four benchmarks, it speeds long-horizon episode inference by 9.6–18.8× (2.9–3.0× on Cube) and raises task success by up to 33.3 percentage points.

With phasor measurement units installed at only a few generator buses and reporting intermittently, a physics-informed graph neural operator maps a measurement window directly to the rotor-angle and speed trajectories of all machines, including those without sensors, and trains without state labels. It imposes the swing equations as a hard constraint, or as a soft one with a learned correction that absorbs model error; on the 9-bus, Kundur, 39-bus and 118-bus systems it is the most accurate estimator that neither optimizes per scenario nor uses state labels, except on the 9-bus system at large model error, where a tuned Kalman smoother is better.

With phasor measurement units installed at only a few generator buses and reporting intermittently, a physics-informed graph neural operator maps a measurement window directly to the rotor-angle and speed trajectories of all machines, including those without sensors, and trains without state labels. It imposes the swing equations as a hard constraint, or as a soft one with a learned correction that absorbs model error; on the 9-bus, Kundur, 39-bus and 118-bus systems it is the most accurate estimator that neither optimizes per scenario nor uses state labels, except on the 9-bus system at large model error, where a tuned Kalman smoother is better.

Electrode caps differ in channel count and layout across laboratories, so recordings do not even share an observation space, and cross-dataset EEG decoding has stayed limited. hTSMNet places a linear map generated from electrode coordinates in front of the decoding backbone, projecting any montage onto 15 fixed scalp anchors; the map has no subject-specific parameters, so unseen subjects and datasets align zero-shot. On a leave-one-dataset-out motor-imagery benchmark of three datasets and 172 subjects it matches the strongest comparator at strict zero-shot and reaches the highest accuracy on every dataset once a few calibration trials refit only the final classifier.

Electrode caps differ in channel count and layout across laboratories, so recordings do not even share an observation space, and cross-dataset EEG decoding has stayed limited. hTSMNet places a linear map generated from electrode coordinates in front of the decoding backbone, projecting any montage onto 15 fixed scalp anchors; the map has no subject-specific parameters, so unseen subjects and datasets align zero-shot. On a leave-one-dataset-out motor-imagery benchmark of three datasets and 172 subjects it matches the strongest comparator at strict zero-shot and reaches the highest accuracy on every dataset once a few calibration trials refit only the final classifier.

Judging where quantum dynamics simulation wins depends on two method portfolios that keep moving, so a crossover fixed in a single paper ages quickly. QSim-Arena turns that assessment into a closed loop maintained by AI agents: agents convert published techniques into executable components, admit a method only after it reproduces a published number, and have results checked by a lane that did not produce them. Starting from the fault-tolerant crossover route of Sun et al., agents introduced three published techniques within a week, lowering certified runtime on four Hamiltonian families at sixteen qubits (twelve for the hydrogen chain) by factors of 21, 23, 25 and 185, and measured runtime fourfold on every family.

Judging where quantum dynamics simulation wins depends on two method portfolios that keep moving, so a crossover fixed in a single paper ages quickly. QSim-Arena turns that assessment into a closed loop maintained by AI agents: agents convert published techniques into executable components, admit a method only after it reproduces a published number, and have results checked by a lane that did not produce them. Starting from the fault-tolerant crossover route of Sun et al., agents introduced three published techniques within a week, lowering certified runtime on four Hamiltonian families at sixteen qubits (twelve for the hydrogen chain) by factors of 21, 23, 25 and 185, and measured runtime fourfold on every family.

A team at Eindhoven University of Technology organized LLM-assisted Dafny verification into a ten-step workflow in which Claude Code acts as a verification construction agent, producing specifications, implementations, invariants and proofs under human supervision. The workflow yielded machine-checked Dafny implementations for 37 algorithms from graph theory, labelled transition systems and computational geometry, including the Hopcroft–Tarjan triconnectivity test; the authors report many verifications within hours to a few days, but ran no controlled comparison against human-led verification.

A team at Eindhoven University of Technology organized LLM-assisted Dafny verification into a ten-step workflow in which Claude Code acts as a verification construction agent, producing specifications, implementations, invariants and proofs under human supervision. The workflow yielded machine-checked Dafny implementations for 37 algorithms from graph theory, labelled transition systems and computational geometry, including the Hopcroft–Tarjan triconnectivity test; the authors report many verifications within hours to a few days, but ran no controlled comparison against human-led verification.

A global one-to-one renaming of the amino-acid alphabet leaves alignment-level statistics, column entropies and pairwise edit distances unchanged, yet contact maps from ESM-2, MSA Transformer and AMPLIFY collapse, while an inverse-covariance Potts model stays bit-identical. Across nine proteins and three seeds, ESM-2 650M keeps an overlap of only 0.015 with its own original contact set after random relabeling of all twenty residues, versus 0.655 when relabeling stays inside BLOSUM62 similarity groups. The contact readout therefore depends on canonical residue identity and cannot be cited as proof that these models have internalized coevolutionary statistics.

A global one-to-one renaming of the amino-acid alphabet leaves alignment-level statistics, column entropies and pairwise edit distances unchanged, yet contact maps from ESM-2, MSA Transformer and AMPLIFY collapse, while an inverse-covariance Potts model stays bit-identical. Across nine proteins and three seeds, ESM-2 650M keeps an overlap of only 0.015 with its own original contact set after random relabeling of all twenty residues, versus 0.655 when relabeling stays inside BLOSUM62 similarity groups. The contact readout therefore depends on canonical residue identity and cannot be cited as proof that these models have internalized coevolutionary statistics.

A single time-resolved micro-CT experiment can yield tens to hundreds of three-dimensional images, and segmenting each one independently lets errors drift between frames and be mistaken for physical change. PoreTimeNet feeds several images from one temporal window through a shared spatial encoder and borrows information across time steps with temporal attention and decoder fusion. On a carbonate dissolution sequence its mean absolute porosity error was 0.28 percentage points against 0.97 for a 3D model that ignores temporal information, and rock boundary F1 rose from 0.833 to 0.941.

A single time-resolved micro-CT experiment can yield tens to hundreds of three-dimensional images, and segmenting each one independently lets errors drift between frames and be mistaken for physical change. PoreTimeNet feeds several images from one temporal window through a shared spatial encoder and borrows information across time steps with temporal attention and decoder fusion. On a carbonate dissolution sequence its mean absolute porosity error was 0.28 percentage points against 0.97 for a 3D model that ignores temporal information, and rock boundary F1 rose from 0.833 to 0.941.

PharmaDiff turns a 3D pharmacophore hypothesis into an atom-level subgraph and injects it as a condition into diffusion-based generation, producing 3D molecules that follow the prescribed spatial arrangement. Across 250 pharmacophore conditions drawn from the GEOM-Drugs test set it reaches a mean match score of 0.8964 and a perfect match rate of 0.6990, above PGMG, TransPharmer, REINVENT4, ShEPhERD and PhoreGen; its average docking scores on four targets also beat DiffSBDD, which needs the protein structure, though its PoseBusters pass rate is below ShEPhERD and PhoreGen.

PharmaDiff turns a 3D pharmacophore hypothesis into an atom-level subgraph and injects it as a condition into diffusion-based generation, producing 3D molecules that follow the prescribed spatial arrangement. Across 250 pharmacophore conditions drawn from the GEOM-Drugs test set it reaches a mean match score of 0.8964 and a perfect match rate of 0.6990, above PGMG, TransPharmer, REINVENT4, ShEPhERD and PhoreGen; its average docking scores on four targets also beat DiffSBDD, which needs the protein structure, though its PoseBusters pass rate is below ShEPhERD and PhoreGen.

Majority-vote test-time reinforcement learning makes each question lock onto one answer during training, often a wrong one on hard questions, and until now which answer that is became known only after training. The work treats this as a short race driven by sampling noise between the answers that are close at the start, and gives a closed-form probability for how the network kernel shifts the gap of the two leaders through updates of the trailing answers. In LoRA training on MMLU-Pro multiple-choice questions with three models from two families, predictions from a frozen-kernel chain computed before training forecast which answer each question locks into better than a logistic model refitted on the outcomes, using no fitted constants.

Majority-vote test-time reinforcement learning makes each question lock onto one answer during training, often a wrong one on hard questions, and until now which answer that is became known only after training. The work treats this as a short race driven by sampling noise between the answers that are close at the start, and gives a closed-form probability for how the network kernel shifts the gap of the two leaders through updates of the trailing answers. In LoRA training on MMLU-Pro multiple-choice questions with three models from two families, predictions from a frozen-kernel chain computed before training forecast which answer each question locks into better than a logistic model refitted on the outcomes, using no fitted constants.

Language agents can store experience as natural-language skill documents, but rewriting a skill from a single trajectory after every deployed task lets its quality drift. EvoGuard turns each update into a two-way choice: offline, skills are scored on a private test set, rejection sampling keeps pairs with a clear utility gap, and a selector trained on the logits of two label tokens decides which skill to keep; at deployment, one forward pass settles which of the old and new skills stays. On ALFWorld, with Qwen3-4B and Qwen3-8B executors, it averages 56.5% success, against 45.5% for accepting every rewrite and 36.7% with no skill.

Language agents can store experience as natural-language skill documents, but rewriting a skill from a single trajectory after every deployed task lets its quality drift. EvoGuard turns each update into a two-way choice: offline, skills are scored on a private test set, rejection sampling keeps pairs with a clear utility gap, and a selector trained on the logits of two label tokens decides which skill to keep; at deployment, one forward pass settles which of the old and new skills stays. On ALFWorld, with Qwen3-4B and Qwen3-8B executors, it averages 56.5% success, against 45.5% for accepting every rewrite and 36.7% with no skill.

Non-prefix KV-cache reuse encodes retrieved chunks separately and concatenates their caches, losing 19% to 100% of context-dependent questions that full prefill answers, with same-kind distractor chunks adding to the loss. Comparing whether a small repair budget should go to distractors or to the answer-bearing chunk, the study finds that restoring all distractors costs nearly a full prefill, while recomputing the answer-bearing chunk beats token-level selectors at a 5% budget in all nine open-domain QA and API-documentation settings.

Non-prefix KV-cache reuse encodes retrieved chunks separately and concatenates their caches, losing 19% to 100% of context-dependent questions that full prefill answers, with same-kind distractor chunks adding to the loss. Comparing whether a small repair budget should go to distractors or to the answer-bearing chunk, the study finds that restoring all distractors costs nearly a full prefill, while recomputing the answer-bearing chunk beats token-level selectors at a 5% budget in all nine open-domain QA and API-documentation settings.

Offline imitation learning usually demands demonstrations that are both safe and high-performing, or a hand-written cost function, and neither is easy to obtain in real settings. DexDICE separates the data into safe demonstrations and performant demonstrations, uses the safe set to define a safe support region in which the policy may act, and then matches the performant behavior's distribution inside that region, training without reward or cost labels. On 38 re-curated DSRL tasks it keeps group-averaged cost below the constraint and attains the highest mean return among safe methods; on a SUMMIT-XL mobile robot trained from 20 human teleoperation episodes, it reached the goal in all 10 trials with zero safety violations.

Offline imitation learning usually demands demonstrations that are both safe and high-performing, or a hand-written cost function, and neither is easy to obtain in real settings. DexDICE separates the data into safe demonstrations and performant demonstrations, uses the safe set to define a safe support region in which the policy may act, and then matches the performant behavior's distribution inside that region, training without reward or cost labels. On 38 re-curated DSRL tasks it keeps group-averaged cost below the constraint and attains the highest mean return among safe methods; on a SUMMIT-XL mobile robot trained from 20 human teleoperation episodes, it reached the goal in all 10 trials with zero safety violations.

Long-running LLM agents compact their context by writing a handoff summary that the next instance continues from, and that summary can state permissions or requirements the record does not support. Using synthetic tool-use tasks with a ground-truth permission ledger and same-snapshot replay, two AI readers confirmed unsupported restrictions in 1.2% and 2.4% of readable summaries from two open-weight models on held-out instances, and unsupported authorizations in 0.1% and 1.3%. Replacing the system prompt's "Some actions need approval." with a sentence saying where requirements come from cut unsupported restrictions from 5.0% to 1.8% of development summaries and from 1.8% to 0.3% on held-out templates, while requests for approvals the ledger requires barely changed.

Long-running LLM agents compact their context by writing a handoff summary that the next instance continues from, and that summary can state permissions or requirements the record does not support. Using synthetic tool-use tasks with a ground-truth permission ledger and same-snapshot replay, two AI readers confirmed unsupported restrictions in 1.2% and 2.4% of readable summaries from two open-weight models on held-out instances, and unsupported authorizations in 0.1% and 1.3%. Replacing the system prompt's "Some actions need approval." with a sentence saying where requirements come from cut unsupported restrictions from 5.0% to 1.8% of development summaries and from 1.8% to 0.3% on held-out templates, while requests for approvals the ledger requires barely changed.

In few-shot CLIP adaptation, support-set gradient magnitudes are badly miscalibrated, yet the signs of a small top-magnitude subset still point toward descent. SignSparseSGD and DynSignSparseSGD therefore use gradient magnitude only to choose where to update and fixed-magnitude signs to decide how, keeping no first- or second-moment state. Across 11 datasets, three CLIP backbones and five shot levels, the dynamic variant has the highest mean accuracy in all 15 backbone–shot settings, reaching 72.7% on ViT-B/16 at one shot, 1.3 points above SO.

In few-shot CLIP adaptation, support-set gradient magnitudes are badly miscalibrated, yet the signs of a small top-magnitude subset still point toward descent. SignSparseSGD and DynSignSparseSGD therefore use gradient magnitude only to choose where to update and fixed-magnitude signs to decide how, keeping no first- or second-moment state. Across 11 datasets, three CLIP backbones and five shot levels, the dynamic variant has the highest mean accuracy in all 15 backbone–shot settings, reaching 72.7% on ViT-B/16 at one shot, 1.3 points above SO.

Diffusion models for spatiotemporal prediction usually add independent noise to each future frame and leave temporal dependence to the denoiser. DSSD instead builds the temporal operator of the corruption process from a learnable diagonal state-space model whose decay rates and frequencies train jointly with the denoiser, and uses the Cholesky factor of that model's stationary covariance so the operator stays causal, exactly invertible and unit-variance per frame. On SEVIR, Turbulence, BAIR and RoboNet it beats diffusion baselines sharing the same denoiser, with raw CRPS 3.975 versus 4.174 for the strongest SEVIR baseline, DPM.

Diffusion models for spatiotemporal prediction usually add independent noise to each future frame and leave temporal dependence to the denoiser. DSSD instead builds the temporal operator of the corruption process from a learnable diagonal state-space model whose decay rates and frequencies train jointly with the denoiser, and uses the Cholesky factor of that model's stationary covariance so the operator stays causal, exactly invertible and unit-variance per frame. On SEVIR, Turbulence, BAIR and RoboNet it beats diffusion baselines sharing the same denoiser, with raw CRPS 3.975 versus 4.174 for the strongest SEVIR baseline, DPM.

Each time a web-search agent calls a tool, the long returned text is appended verbatim to its context, so after several steps the context balloons and dilutes attention. Coca-Cola instead compresses those observations into a few continuous latent tokens and embeds them directly in the agent's input embedding space, with action-preserving training aligning the next-action distributions under full and compressed observations. Across four QA benchmarks and two model families, total token consumption falls by 70.01%–89.24% and end-to-end runtime by 24.45%–55.97%, while EM and F1 exceed the uncompressed agent in every setting.

Each time a web-search agent calls a tool, the long returned text is appended verbatim to its context, so after several steps the context balloons and dilutes attention. Coca-Cola instead compresses those observations into a few continuous latent tokens and embeds them directly in the agent's input embedding space, with action-preserving training aligning the next-action distributions under full and compressed observations. Across four QA benchmarks and two model families, total token consumption falls by 70.01%–89.24% and end-to-end runtime by 24.45%–55.97%, while EM and F1 exceed the uncompressed agent in every setting.

Translating parallel scientific code among CUDA, OpenMP and SYCL has relied on fixed sequential pipelines that apply one repair strategy to every error and judge correctness only by compilation and execution. A verdict-driven orchestrator splits translation, compilation diagnosis and execution validation into three agents and routes each failure to a patch, semantic, sync or memory repair skill. On 30 HeCBench kernels, three open-source models and six translation directions, the workflow reaches a higher verified correctness rate than LASSI in all 18 model-and-direction combinations while using up to 52.4% fewer tokens.

Translating parallel scientific code among CUDA, OpenMP and SYCL has relied on fixed sequential pipelines that apply one repair strategy to every error and judge correctness only by compilation and execution. A verdict-driven orchestrator splits translation, compilation diagnosis and execution validation into three agents and routes each failure to a patch, semantic, sync or memory repair skill. On 30 HeCBench kernels, three open-source models and six translation directions, the workflow reaches a higher verified correctness rate than LASSI in all 18 model-and-direction combinations while using up to 52.4% fewer tokens.

Feed-forward stereo 3D reconstruction predicts dense pointmaps in one pass and offers no way to correct a wrong result. Click3R inserts a lightweight point-correspondence adapter ahead of DUSt3R's decoder, injecting a handful of user-supplied cross-view point pairs as geometric constraints without test-time fine-tuning or optimization. On the curated Hard3DBench, a single correspondence click reduces reconstruction error; in a 100-pair human-labeling test, median rotation error fell from 17.30° to 8.78° and mAA@30 rose from 36.2 to 56.2.

Feed-forward stereo 3D reconstruction predicts dense pointmaps in one pass and offers no way to correct a wrong result. Click3R inserts a lightweight point-correspondence adapter ahead of DUSt3R's decoder, injecting a handful of user-supplied cross-view point pairs as geometric constraints without test-time fine-tuning or optimization. On the curated Hard3DBench, a single correspondence click reduces reconstruction error; in a 100-pair human-labeling test, median rotation error fell from 17.30° to 8.78° and mAA@30 rose from 36.2 to 56.2.

TactileLab moves tangential motion out of soft-body deformation and generates a dense tactile flow field directly from rigid-body simulation, feeding it to policies alongside contact depth images. Rigid-body tactile simulators previously offered mainly depth-like signals that could not express shear or slip, while soft-body simulators captured deformation but could not sustain large-scale reinforcement learning. On dual-arm following and dexterous in-hand rotation, policies with tactile flow converged faster, and on real hardware the depth-plus-flow policy completed the most rotations across cylindrical, round and cube-shaped objects.

TactileLab moves tangential motion out of soft-body deformation and generates a dense tactile flow field directly from rigid-body simulation, feeding it to policies alongside contact depth images. Rigid-body tactile simulators previously offered mainly depth-like signals that could not express shear or slip, while soft-body simulators captured deformation but could not sustain large-scale reinforcement learning. On dual-arm following and dexterous in-hand rotation, policies with tactile flow converged faster, and on real hardware the depth-plus-flow policy completed the most rotations across cylindrical, round and cube-shaped objects.

Web platforms must classify new pages before anyone labels them, large language model labels per page are noisy, and links mean opposite things across sites: online stores often join same-category products, while university sites mostly join pages of different classes. ZeroPolar calls the sign with which links enter a prediction structural polarity, reads that sign without true labels from the excess agreement of LLM labels and text-prototype predictions along edges, and applies one signed propagation step. With one GPT-4o label per page, ZeroPolar-T reaches the highest eleven-graph mean, 73.99%, 6.01 points above LLM-BP.

Web platforms must classify new pages before anyone labels them, large language model labels per page are noisy, and links mean opposite things across sites: online stores often join same-category products, while university sites mostly join pages of different classes. ZeroPolar calls the sign with which links enter a prediction structural polarity, reads that sign without true labels from the excess agreement of LLM labels and text-prototype predictions along edges, and applies one signed propagation step. With one GPT-4o label per page, ZeroPolar-T reaches the highest eleven-graph mean, 73.99%, 6.01 points above LLM-BP.

Automated design of multi-agent systems has mostly worked at the task level, optimizing one fixed workflow per benchmark and applying it to every query. OOPMAS moves design to the individual query: a frozen LLM as meta-agent first generates object-style agents with roles, tools and persistent state, then an executable main-function workflow, while a dynamic skill library accumulates structured lessons from execution feedback without gradient updates. On a benchmark mixing code, math and QA, OOPMAS reaches 89.6% overall accuracy, 18.1 percentage points above the strongest baseline, ADAS.

Automated design of multi-agent systems has mostly worked at the task level, optimizing one fixed workflow per benchmark and applying it to every query. OOPMAS moves design to the individual query: a frozen LLM as meta-agent first generates object-style agents with roles, tools and persistent state, then an executable main-function workflow, while a dynamic skill library accumulates structured lessons from execution feedback without gradient updates. On a benchmark mixing code, math and QA, OOPMAS reaches 89.6% overall accuracy, 18.1 percentage points above the strongest baseline, ADAS.

After a failed tool call, whether an agent retries, fixes the arguments or stops to report depends not only on the error text but on how that feedback enters the conversation. Holding the error content fixed and varying only delivery across 13 open models from 3B to 70B parameters, keeping the failed exchange and appending a runtime note with guidance in a new user turn avoided a severe stopping failure in Llama models; implemented as an opt-in recovery layer inside an inference engine, this form raised ToolMaze success on explicit errors by 14.2 points with 7–8B models and 10.5 points with 32–70B models, with no measurable gain on Loud/Quiet.

After a failed tool call, whether an agent retries, fixes the arguments or stops to report depends not only on the error text but on how that feedback enters the conversation. Holding the error content fixed and varying only delivery across 13 open models from 3B to 70B parameters, keeping the failed exchange and appending a runtime note with guidance in a new user turn avoided a severe stopping failure in Llama models; implemented as an opt-in recovery layer inside an inference engine, this form raised ToolMaze success on explicit errors by 14.2 points with 7–8B models and 10.5 points with 32–70B models, with no measurable gain on Loud/Quiet.

DexForesight injects predictive features from an action-conditioned spatial world model into a dexterous policy's latent space through query alignment, and at deployment keeps only the policy, which acts from RGB, language and robot state. On 11 DexJoCo simulation tasks it raises the GR00T N1.7 backbone's overall success from 49.4% to 60.5%, and on three real-robot tasks it reaches 58.3%, 8.3 points above the strongest baseline. The world model, depth encoder and future-prediction heads are all removed at inference.

DexForesight injects predictive features from an action-conditioned spatial world model into a dexterous policy's latent space through query alignment, and at deployment keeps only the policy, which acts from RGB, language and robot state. On 11 DexJoCo simulation tasks it raises the GR00T N1.7 backbone's overall success from 49.4% to 60.5%, and on three real-robot tasks it reaches 58.3%, 8.3 points above the strongest baseline. The world model, depth encoder and future-prediction heads are all removed at inference.

In humanoid manipulation, a large vision-language-action backbone is slow, so synchronous execution acts on stale observations and control frequency falls as the backbone grows. DASH splits the policy into a slow vision-language model and a fast flow-matching action head running as three concurrent processes over LCM, so the control rate is set by the action head alone. On a RoboTwin benchmark that charges inference latency in simulated time, DASH holds 20 Hz from 0.8B to 8B backbones while the synchronous rate falls from 9.7 Hz to 4.1 Hz, with higher success at every size; on a real dexterous humanoid, pick-and-place succeeds in 23 of 24 trials versus 16 of 24 for the synchronous pipeline.

In humanoid manipulation, a large vision-language-action backbone is slow, so synchronous execution acts on stale observations and control frequency falls as the backbone grows. DASH splits the policy into a slow vision-language model and a fast flow-matching action head running as three concurrent processes over LCM, so the control rate is set by the action head alone. On a RoboTwin benchmark that charges inference latency in simulated time, DASH holds 20 Hz from 0.8B to 8B backbones while the synchronous rate falls from 9.7 Hz to 4.1 Hz, with higher success at every size; on a real dexterous humanoid, pick-and-place succeeds in 23 of 24 trials versus 16 of 24 for the synchronous pipeline.

High-stakes regression decisions need a point estimate, a risk interval and an option to abstain, yet interval methods and selective prediction each supply only part of that triad. SelPI trains interval, value and two abstention heads jointly under one loss, with a Fréchet penalty tying the two gates' budgets to the joint acceptance rate they deliver. Across 18 tabular benchmarks, interval width predicts large point errors well but is anti-correlated with interval misses, so abstaining on wide intervals discards inputs the interval got right; with the base predictor held fixed, the learned gate beats width thresholding on a proper interval score on all 18 tabular datasets, and the advantage grows with task heteroscedasticity.

High-stakes regression decisions need a point estimate, a risk interval and an option to abstain, yet interval methods and selective prediction each supply only part of that triad. SelPI trains interval, value and two abstention heads jointly under one loss, with a Fréchet penalty tying the two gates' budgets to the joint acceptance rate they deliver. Across 18 tabular benchmarks, interval width predicts large point errors well but is anti-correlated with interval misses, so abstaining on wide intervals discards inputs the interval got right; with the base predictor held fixed, the learned gate beats width thresholding on a proper interval score on all 18 tabular datasets, and the advantage grows with task heteroscedasticity.

Medical and astronomical images are orders of magnitude larger than natural images, yet tumors and galaxies occupy a tiny fraction of the data, and uniform compression such as quantization or pruning can erode precision exactly there. Conformal Adaptation first screens candidates with a low-resolution model, calibrates the selection threshold with conformal risk control so the discarded cells carry at most a user-set expected fraction of true signal mass, then refines only the selected regions. Across brain-tumor segmentation, head-and-neck segmentation, calcium spike-rate inference and galaxy flux super-resolution, the pipeline cuts computation while keeping accuracy close, and VISTA3D reaches an 82.10x FLOPs speedup in CPU-only inference.

Medical and astronomical images are orders of magnitude larger than natural images, yet tumors and galaxies occupy a tiny fraction of the data, and uniform compression such as quantization or pruning can erode precision exactly there. Conformal Adaptation first screens candidates with a low-resolution model, calibrates the selection threshold with conformal risk control so the discarded cells carry at most a user-set expected fraction of true signal mass, then refines only the selected regions. Across brain-tumor segmentation, head-and-neck segmentation, calcium spike-rate inference and galaxy flux super-resolution, the pipeline cuts computation while keeping accuracy close, and VISTA3D reaches an 82.10x FLOPs speedup in CPU-only inference.

Möbius places the entire denoising process of motion generation inside the same multimodal representation space used for motion understanding, so clean motion, noisy motion and text share attention without a discrete tokenizer or a pretrained motion VAE. On HumanML3D, the 0.6B variant outperforms diffusion-based, language-model-based and unified motion-language baselines on both text-to-motion generation and motion captioning; scaling to 1.7B parameters and 1,700 hours of motion data improves all motion-language tasks further.

Möbius places the entire denoising process of motion generation inside the same multimodal representation space used for motion understanding, so clean motion, noisy motion and text share attention without a discrete tokenizer or a pretrained motion VAE. On HumanML3D, the 0.6B variant outperforms diffusion-based, language-model-based and unified motion-language baselines on both text-to-motion generation and motion captioning; scaling to 1.7B parameters and 1,700 hours of motion data improves all motion-language tasks further.

FunFormer compresses up to 20 kb of sequence into a fixed token set for binary regulatory-activity prediction, paired with an architecture- and parameter-matched control that reads only the central 1 kb, and both are transferred among six vertebrates. Within species the 20 kb version has the highest mean AUROC in all six genomes, 1.0 to 3.8 percentage points above three published 1 kb baselines; across species that context gain decays with divergence, nears zero at the amniote split, and reverses between 320 and 435 Mya, while the 1 kb model still transfers into fish.

FunFormer compresses up to 20 kb of sequence into a fixed token set for binary regulatory-activity prediction, paired with an architecture- and parameter-matched control that reads only the central 1 kb, and both are transferred among six vertebrates. Within species the 20 kb version has the highest mean AUROC in all six genomes, 1.0 to 3.8 percentage points above three published 1 kb baselines; across species that context gain decays with divergence, nears zero at the amniote split, and reverses between 320 and 435 Mya, while the 1 kb model still transfers into fish.

IBM released LLMON (LLM Object Notation), a markup language, together with two experimental preview models trained from Granite 4.1 3B and 8B base checkpoints, granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview, which use explicit tags to separate instructions, data and requirements. On the OPI benchmark the LLMON models avoided 47% more attacks than the best frontier model tested and 95% more than the best open-source model, while general capability stayed on par with Granite 4.1 Instruct. IBM describes the release as an experimental preview for exploration and evaluation, not for production.

IBM released LLMON (LLM Object Notation), a markup language, together with two experimental preview models trained from Granite 4.1 3B and 8B base checkpoints, granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview, which use explicit tags to separate instructions, data and requirements. On the OPI benchmark the LLMON models avoided 47% more attacks than the best frontier model tested and 95% more than the best open-source model, while general capability stayed on par with Granite 4.1 Instruct. IBM describes the release as an experimental preview for exploration and evaluation, not for production.

Training five operator models on the one-dimensional discrete Poisson problem shows that error is not learned wavenumber by wavenumber but is retained along small-eigenvalue directions of the kernel that governs the training update. Sine-aligned spectral and linear Fourier neural operators give each wavenumber its own rate, so error collects at high wavenumbers; dense networks have eigendirections close to random in the sine basis, so error is nearly equal at every wavenumber and relative error grows only because the true solution spectrum decays. Removing the minibatch cross terms leaves the GELU dense network with under 0.1% of squared error below ηλ≈8×10⁻⁵ under either forcing.

Training five operator models on the one-dimensional discrete Poisson problem shows that error is not learned wavenumber by wavenumber but is retained along small-eigenvalue directions of the kernel that governs the training update. Sine-aligned spectral and linear Fourier neural operators give each wavenumber its own rate, so error collects at high wavenumbers; dense networks have eigendirections close to random in the sine basis, so error is nearly equal at every wavenumber and relative error grows only because the true solution spectrum decays. Removing the minibatch cross terms leaves the GELU dense network with under 0.1% of squared error below ηλ≈8×10⁻⁵ under either forcing.

When a pretrained flow-matching model must satisfy an equality constraint supplied only at inference time, common practice repeatedly corrects predicted endpoints, and those discrete corrections can disturb the learned physical structure. MAFM instead evolves a persistent endpoint anchor continuously under an ordinary differential equation, separating model-guided tangential motion from normal attraction toward the constraint. On five PDE benchmarks—Poisson, Burgers, reaction–diffusion, Euler and Navier–Stokes—MAFM meets the prescribed constraint tolerance and achieves lower sample-level relative error than baselines that meet the same tolerance.

When a pretrained flow-matching model must satisfy an equality constraint supplied only at inference time, common practice repeatedly corrects predicted endpoints, and those discrete corrections can disturb the learned physical structure. MAFM instead evolves a persistent endpoint anchor continuously under an ordinary differential equation, separating model-guided tangential motion from normal attraction toward the constraint. On five PDE benchmarks—Poisson, Burgers, reaction–diffusion, Euler and Navier–Stokes—MAFM meets the prescribed constraint tolerance and achieves lower sample-level relative error than baselines that meet the same tolerance.

Remote sensing change detection usually depends on large collections of temporally aligned, pixel-level annotations, which limits scaling across regions and platforms. SpotDiff composes DINOv3 semantic features, AnyUp pixel-resolution feature reconstruction and SAM 3 text-guided region proposals into a training-free pipeline, then compares the cosine similarity of the two dates' reconstructed features inside the proposed regions. On LEVIR-CD the ViT-7B version reaches 82.46 F1, 2.06 points above the strongest prior training-free method; on SECOND it raises building-change IoU by 10.3.

Remote sensing change detection usually depends on large collections of temporally aligned, pixel-level annotations, which limits scaling across regions and platforms. SpotDiff composes DINOv3 semantic features, AnyUp pixel-resolution feature reconstruction and SAM 3 text-guided region proposals into a training-free pipeline, then compares the cosine similarity of the two dates' reconstructed features inside the proposed regions. On LEVIR-CD the ViT-7B version reaches 82.46 F1, 2.06 points above the strongest prior training-free method; on SECOND it raises building-change IoU by 10.3.

Unpaired medical image translation previously packed structure preservation and appearance generation into one source-to-target mapping, which without paired supervision can drop vessels or distort the foveal avascular zone. This work separates the two: a cross-modality shared structure latent space is learned first, then the structure encoder is frozen and a conditional flow model renders only target-modality appearance. On OCT-to-OCTA and brain MR-to-CT, it outperforms all unpaired baselines and, despite no paired supervision, surpasses paired baselines, improving PSNR by up to 11.2% and SSIM by up to 26.4% on OCT-to-OCTA.

Unpaired medical image translation previously packed structure preservation and appearance generation into one source-to-target mapping, which without paired supervision can drop vessels or distort the foveal avascular zone. This work separates the two: a cross-modality shared structure latent space is learned first, then the structure encoder is frozen and a conditional flow model renders only target-modality appearance. On OCT-to-OCTA and brain MR-to-CT, it outperforms all unpaired baselines and, despite no paired supervision, surpasses paired baselines, improving PSNR by up to 11.2% and SSIM by up to 26.4% on OCT-to-OCTA.

Generative models for target-based drug design have long compressed protein–ligand interactions into a single docking score, hiding functional differences between binding modes on the same target. BindMol clusters the binding pocket into hotspots from reference complexes, assigns each hotspot its own reward dimension, and optimizes the conflicting objectives with preference-conditioned multi-objective reinforcement learning. On parp1, fa7, 5ht1b, braf and jak2 its novel hit ratios are 42.617, 24.580, 42.149, 30.767 and 44.864, above GEAM at 40.635, 17.903, 36.667, 27.887 and 42.025, and it recovers the crystallographic type-I and type-II binding modes more often on braf and jak2.

Generative models for target-based drug design have long compressed protein–ligand interactions into a single docking score, hiding functional differences between binding modes on the same target. BindMol clusters the binding pocket into hotspots from reference complexes, assigns each hotspot its own reward dimension, and optimizes the conflicting objectives with preference-conditioned multi-objective reinforcement learning. On parp1, fa7, 5ht1b, braf and jak2 its novel hit ratios are 42.617, 24.580, 42.149, 30.767 and 44.864, above GEAM at 40.635, 17.903, 36.667, 27.887 and 42.025, and it recovers the crystallographic type-I and type-II binding modes more often on braf and jak2.

Quasar variability is usually tested for periodic signals against a damped random walk (DRW) null, and a mis-specified noise model biases the significance estimates. Nullasar uses flow matching to learn the statistical structure of stochastic variability from an ensemble of synthetic light curves without assuming a noise model, then generates null realizations conditioned on a single observed curve. On synthetic AR(2) light curves its periodicity-detection AUC is 0.9648, averaged over three independently trained models, above 0.9180 for fitting AR(2) per curve; in the AR(1) case it matches fitting the correct DRW model.

Quasar variability is usually tested for periodic signals against a damped random walk (DRW) null, and a mis-specified noise model biases the significance estimates. Nullasar uses flow matching to learn the statistical structure of stochastic variability from an ensemble of synthetic light curves without assuming a noise model, then generates null realizations conditioned on a single observed curve. On synthetic AR(2) light curves its periodicity-detection AUC is 0.9648, averaged over three independently trained models, above 0.9180 for fitting AR(2) per curve; in the AR(1) case it matches fitting the correct DRW model.

Curiosity-0826-4B extends a generative medical event foundation model trained on Epic Cosmos records from adults to pediatric patients and adds explicit representations of death and right-censoring. The model grows from 1B to 4.3B parameters and from an 8K to a 16K context window, generally matching or exceeding Curiosity-0226-1B on tasks both support, with mean absolute error on laboratory value prediction 18–46% lower. Pediatric AUROCs for incident asthma and asthma exacerbation are comparable to an adult reference cohort, while sickle cell crisis is lower in children.

Curiosity-0826-4B extends a generative medical event foundation model trained on Epic Cosmos records from adults to pediatric patients and adds explicit representations of death and right-censoring. The model grows from 1B to 4.3B parameters and from an 8K to a 16K context window, generally matching or exceeding Curiosity-0226-1B on tasks both support, with mean absolute error on laboratory value prediction 18–46% lower. Pediatric AUROCs for incident asthma and asthma exacerbation are comparable to an adult reference cohort, while sickle cell crisis is lower in children.

Planning and control of differentially flat systems rests on a flat output, yet no general procedure writes one down for arbitrary nonlinear dynamics. FlatSeek reverses the reconstruction relation in the definition of flatness: fix an output and its reconstruction maps, close the dynamics differentially, and obtain system–output pairs with known certificates at scale. A Qwen3.5-2B model fine-tuned with LoRA proposes candidates that a symbolic verifier accepts or rejects. FlatSeek solves 74 literature and 171 synthetic systems, against 58 and 127 for the strongest traditional symbolic method.

Planning and control of differentially flat systems rests on a flat output, yet no general procedure writes one down for arbitrary nonlinear dynamics. FlatSeek reverses the reconstruction relation in the definition of flatness: fix an output and its reconstruction maps, close the dynamics differentially, and obtain system–output pairs with known certificates at scale. A Qwen3.5-2B model fine-tuned with LoRA proposes candidates that a symbolic verifier accepts or rejects. FlatSeek solves 74 literature and 171 synthetic systems, against 58 and 127 for the strongest traditional symbolic method.

Medical image synthesis is often treated as filling in a missing modality, yet a realistic image need not help the downstream task. Across five public datasets, a training framework ties synthesis to task learning: the translator is constrained during training by image-adaptive frequency consistency on structure and detail, and the downstream network receives teacher features only when a teacher using synthetic images beats the student by a task-specific margin on labelled samples. Translation fidelity improved for CycleGAN, CUT and SynDiff unpaired translators, with gains in brain-tumour segmentation and lesion centre-point detection, and the training-time components are removed at inference.

Medical image synthesis is often treated as filling in a missing modality, yet a realistic image need not help the downstream task. Across five public datasets, a training framework ties synthesis to task learning: the translator is constrained during training by image-adaptive frequency consistency on structure and detail, and the downstream network receives teacher features only when a teacher using synthetic images beats the student by a task-specific margin on labelled samples. Translation fidelity improved for CycleGAN, CUT and SynDiff unpaired translators, with gains in brain-tumour segmentation and lesion centre-point detection, and the training-time components are removed at inference.

Rebuilding a 128×128 vorticity field from 8×8 observations is a 16× upscaling in which the coarse grid resolves only wavenumbers up to 4, so fine vortical structure must be inferred rather than interpolated. SIMR-NO splits the reconstruction into 32→64 and 64→128 stages, each adding a Fourier-layer residual correction gated by radial wavenumber on top of a bicubic base, with local convolutions recovering features beyond the truncated Fourier basis. Its mean relative error is 0.2506 on Kolmogorov-forced turbulence, below LapSRN, FNO and EDSR, and about 3% on decaying homogeneous isotropic turbulence at Re = 1000, 4000 and 8000, with little change across Reynolds number.

Rebuilding a 128×128 vorticity field from 8×8 observations is a 16× upscaling in which the coarse grid resolves only wavenumbers up to 4, so fine vortical structure must be inferred rather than interpolated. SIMR-NO splits the reconstruction into 32→64 and 64→128 stages, each adding a Fourier-layer residual correction gated by radial wavenumber on top of a bicubic base, with local convolutions recovering features beyond the truncated Fourier basis. Its mean relative error is 0.2506 on Kolmogorov-forced turbulence, below LapSRN, FNO and EDSR, and about 3% on decaying homogeneous isotropic turbulence at Re = 1000, 4000 and 8000, with little change across Reynolds number.

Recovering a partial differential equation from sampled solution data demands expert work in equation discovery, coefficient estimation and numerical solver configuration. RSI-PDEbench turns that workflow into a benchmark of 1,001 cases over 42 equation templates, split into noise-free, 50% Gaussian-noise and unseen-template-with-shifted-coefficient settings; the accompanying RSI-PDE agent shortlists candidate equations by residual, refines coefficients and solvers with relative feedback, and keeps improvements behind a validation gate. Across four frontier models its mean pass rate rises 106.5–181.9% over single-run baselines and 3.0–19.6% over native agent harnesses.

Recovering a partial differential equation from sampled solution data demands expert work in equation discovery, coefficient estimation and numerical solver configuration. RSI-PDEbench turns that workflow into a benchmark of 1,001 cases over 42 equation templates, split into noise-free, 50% Gaussian-noise and unseen-template-with-shifted-coefficient settings; the accompanying RSI-PDE agent shortlists candidate equations by residual, refines coefficients and solvers with relative feedback, and keeps improvements behind a validation gate. Across four frontier models its mean pass rate rises 106.5–181.9% over single-run baselines and 3.0–19.6% over native agent harnesses.

AI safety policy needs diverse input, yet the public cannot join every discussion. The work builds pluralistic agents: each AI agent first interviews the person it represents, then speaks for them in a multi-agent deliberation, while a policy-writer agent drafts and revises a budget after each round. In a preregistered study with 746 participants, people rated this process above giving input to a single reasoning model, recovering 68% of the satisfaction gap and 82% of the procedural-justice gap between that baseline and direct participation.

AI safety policy needs diverse input, yet the public cannot join every discussion. The work builds pluralistic agents: each AI agent first interviews the person it represents, then speaks for them in a multi-agent deliberation, while a policy-writer agent drafts and revises a budget after each round. In a preregistered study with 746 participants, people rated this process above giving input to a single reasoning model, recovering 68% of the satisfaction gap and 82% of the procedural-justice gap between that baseline and direct participation.

Neural network verifiers usually treat the input domain as an axis-aligned box, yet reachability, obstacle-avoidance and neural Lyapunov certificates in cyber-physical systems often live on oblique or non-convex polytopes, where box bounds become very loose. Fit-LRP optimizes the final affine bound from LiRPA or α-CROWN exactly over the true polyhedral constraints inside branch-and-bound, replacing the box approximation with a sequence of small linear programs. On the ARCH-COMP spacecraft rendezvous benchmark and two neural Lyapunov benchmarks, branching on the hardest region drops, and single-core CPU verification runs faster than α-CROWN on a GPU.

Neural network verifiers usually treat the input domain as an axis-aligned box, yet reachability, obstacle-avoidance and neural Lyapunov certificates in cyber-physical systems often live on oblique or non-convex polytopes, where box bounds become very loose. Fit-LRP optimizes the final affine bound from LiRPA or α-CROWN exactly over the true polyhedral constraints inside branch-and-bound, replacing the box approximation with a sequence of small linear programs. On the ARCH-COMP spacecraft rendezvous benchmark and two neural Lyapunov benchmarks, branching on the hardest region drops, and single-core CPU verification runs faster than α-CROWN on a GPU.

Fluid surrogate models have long reused U-shaped multiscale designs from vision, but their resampling, skip fusion and normalization were not built for flow fields. SUMNO spatially unrolls a multigrid V-cycle into a feed-forward U-shaped network, using parameter-free downsampling and bilinear upsampling for restriction and prolongation, and shifted-window attention at the coarsest level, with window masks set per-axis periodicity, for global coupling. Across seven settings on four flow benchmarks it is best in five and second in two, cutting 10-step rollout MSE by about 30% on Pinball and RBC, with 9.4M parameters.

Fluid surrogate models have long reused U-shaped multiscale designs from vision, but their resampling, skip fusion and normalization were not built for flow fields. SUMNO spatially unrolls a multigrid V-cycle into a feed-forward U-shaped network, using parameter-free downsampling and bilinear upsampling for restriction and prolongation, and shifted-window attention at the coarsest level, with window masks set per-axis periodicity, for global coupling. Across seven settings on four flow benchmarks it is best in five and second in two, cutting 10-step rollout MSE by about 30% on Pinball and RBC, with 9.4M parameters.

Building a phase-diagram database has long depended on expert judgment that is rarely recorded, so published CALPHAD assessments are hard to regenerate or update as new data arrive. ORCHID-SAGE splits an assessment into a general protocol, a system-specific protocol and run cards for each execution, and distributes them with the database. For Cr–Ni, that record alone let a script on a second machine and an AI agent working only through an MCP server regenerate the database, with predictions inside the tolerances declared in advance.

Building a phase-diagram database has long depended on expert judgment that is rarely recorded, so published CALPHAD assessments are hard to regenerate or update as new data arrive. ORCHID-SAGE splits an assessment into a general protocol, a system-specific protocol and run cards for each execution, and distributes them with the database. For Cr–Ni, that record alone let a script on a second machine and an AI agent working only through an MCP server regenerate the database, with predictions inside the tolerances declared in advance.

Particle-resolved granular flow simulation has been limited by repeated information passing over the full particle graph. HCG-GNN moves 5 of its 7 message-passing steps onto two levels of coarse graphs built from occupied cells, matching the multi-step accuracy of a conventional particle-level GNN in inclined-plane granular flow benchmarks while training 6.47 times faster with 46.2% less GPU memory; a model trained on up to 5000 grains predicted a 2 s release of 10000 grains in 31 s, versus 2000 s for DEM under the same benchmark conditions.

Particle-resolved granular flow simulation has been limited by repeated information passing over the full particle graph. HCG-GNN moves 5 of its 7 message-passing steps onto two levels of coarse graphs built from occupied cells, matching the multi-step accuracy of a conventional particle-level GNN in inclined-plane granular flow benchmarks while training 6.47 times faster with 46.2% less GPU memory; a model trained on up to 5000 grains predicted a 2 s release of 10000 grains in 31 s, versus 2000 s for DEM under the same benchmark conditions.

Convolutional VocalTrax replaces the sample-by-sample tract recursion in differentiable articulatory synthesis with fixed-block convolution: it computes the exact frequency response of the same waveguide and applies it as a causal convolution in 5.8 ms blocks, so gradients pass through FFTs and a 43-step recursion along the tract. On 600 AudioMNIST recordings, the tract forward and backward pass runs 47 to 63 times faster on an A100, and a wav2vec 2.0 classifier trained only on real speech recognizes its reconstructions 14.5% of the time versus 10.0% for VocalTrax, although the log-mel distance is 11% higher.

Convolutional VocalTrax replaces the sample-by-sample tract recursion in differentiable articulatory synthesis with fixed-block convolution: it computes the exact frequency response of the same waveguide and applies it as a causal convolution in 5.8 ms blocks, so gradients pass through FFTs and a 43-step recursion along the tract. On 600 AudioMNIST recordings, the tract forward and backward pass runs 47 to 63 times faster on an A100, and a wav2vec 2.0 classifier trained only on real speech recognizes its reconstructions 14.5% of the time versus 10.0% for VocalTrax, although the log-mel distance is 11% higher.

Starting from noisy velocity images, unsteady three-dimensional blood flow and elastic vessel-wall motion can be reconstructed together without a prescribed inlet waveform or initial velocity. The solver alternates a fluid update, in which a physics-informed network fits the observations while enforcing the incompressible Navier–Stokes equations on the moving domain, with a geometry update, in which registration plus interface traction continuity and an elastic momentum residual constrain wall displacement. On a converging channel and a patient-like aorta, the PirateNets version reaches velocity relative errors of 0.0243 and 0.0372, below the 0.1066 and 0.1305 of a standard PINN of the same depth and width.

Starting from noisy velocity images, unsteady three-dimensional blood flow and elastic vessel-wall motion can be reconstructed together without a prescribed inlet waveform or initial velocity. The solver alternates a fluid update, in which a physics-informed network fits the observations while enforcing the incompressible Navier–Stokes equations on the moving domain, with a geometry update, in which registration plus interface traction continuity and an elastic momentum residual constrain wall displacement. On a converging channel and a patient-like aorta, the PirateNets version reaches velocity relative errors of 0.0243 and 0.0372, below the 0.1066 and 0.1305 of a standard PINN of the same depth and width.

Protein language models can be steered toward a target property without changing their weights, but existing activation steering builds the direction from the difference between the global means of property-positive and property-negative proteins, ignoring that the useful direction shifts with sequence background. PULSE first matches positive and negative proteins by representation similarity, then aggregates similarity-weighted, unit-normalized differences between matched pairs into a steering vector. On solubility, immunogenicity and toxicity optimization, PULSE beats global centroid steering in both final predicted property scores and optimization success rates, with the largest score gain on toxicity and the largest success gain on immunogenicity.

Protein language models can be steered toward a target property without changing their weights, but existing activation steering builds the direction from the difference between the global means of property-positive and property-negative proteins, ignoring that the useful direction shifts with sequence background. PULSE first matches positive and negative proteins by representation similarity, then aggregates similarity-weighted, unit-normalized differences between matched pairs into a steering vector. On solubility, immunogenicity and toxicity optimization, PULSE beats global centroid steering in both final predicted property scores and optimization success rates, with the largest score gain on toxicity and the largest success gain on immunogenicity.

Financial scenario pipelines often run diffusion samplers at very few function evaluations, where time-grid discretization error can quietly distort terminal reports such as covariances, tails and drawdown proxies. TAPES treats grid construction as a goal-oriented numerical problem: Richardson comparisons estimate each step's signed local defect, a terminal-report adjoint transports those defects through the remaining discrete flow, and the squared norm of their sum is minimized before a capped direct refinement. Across six budgets on Regime-SV and Mag7, TAPES gives the lowest design- and audit-report errors in all 24 dataset–budget–report rows within the reported five-pipeline panel, with cross-budget mean relative audit-error reductions of about 10–11% and about 1%.

Financial scenario pipelines often run diffusion samplers at very few function evaluations, where time-grid discretization error can quietly distort terminal reports such as covariances, tails and drawdown proxies. TAPES treats grid construction as a goal-oriented numerical problem: Richardson comparisons estimate each step's signed local defect, a terminal-report adjoint transports those defects through the remaining discrete flow, and the squared norm of their sum is minimized before a capped direct refinement. Across six budgets on Regime-SV and Mag7, TAPES gives the lowest design- and audit-report errors in all 24 dataset–budget–report rows within the reported five-pipeline panel, with cross-budget mean relative audit-error reductions of about 10–11% and about 1%.

Large-scale observations of the three-dimensional interior of low clouds have been missing because physics-based tomography repeats expensive 3D radiative-transfer calculations, a cost that has kept it from large, realistic cloud scenes. HybridCT splits the problem: a U-Net estimates cloud optical thickness projections from multi-angle passive radiances, and SART then explicitly inverts those projections into a 3D extinction field. On a synthetic low-cloud scene it reconstructs extinction with RMSE 13.3 km⁻¹ in 3.5 s, against 22.7 km⁻¹ in about 30 min for the physics baseline AT3D.

Large-scale observations of the three-dimensional interior of low clouds have been missing because physics-based tomography repeats expensive 3D radiative-transfer calculations, a cost that has kept it from large, realistic cloud scenes. HybridCT splits the problem: a U-Net estimates cloud optical thickness projections from multi-angle passive radiances, and SART then explicitly inverts those projections into a 3D extinction field. On a synthetic low-cloud scene it reconstructs extinction with RMSE 13.3 km⁻¹ in 3.5 s, against 22.7 km⁻¹ in about 30 min for the physics baseline AT3D.

High-dimensional PDE forecasting on unstructured meshes has run on superconducting quantum processors for the first time. A graph autoencoder compresses mesh fields into a low-dimensional latent state, a fixed, untrained quantum reservoir network supplies features, and a classical linear readout predicts the next state; on a 33,824-cell cylinder flow and a Gray–Scott reaction–diffusion problem, adding those features lowers normalized root mean square error by roughly 21–43% for one-step and pressure-observer prediction against a linear baseline on the same latent space, with the pressure observer demonstrated only on the 26-qubit cylinder configuration and the 52-qubit hardware used for Gray–Scott one-step forecasting.

High-dimensional PDE forecasting on unstructured meshes has run on superconducting quantum processors for the first time. A graph autoencoder compresses mesh fields into a low-dimensional latent state, a fixed, untrained quantum reservoir network supplies features, and a classical linear readout predicts the next state; on a 33,824-cell cylinder flow and a Gray–Scott reaction–diffusion problem, adding those features lowers normalized root mean square error by roughly 21–43% for one-step and pressure-observer prediction against a linear baseline on the same latent space, with the pressure observer demonstrated only on the 26-qubit cylinder configuration and the 52-qubit hardware used for Gray–Scott one-step forecasting.

Jet taggers in high-energy physics are judged mainly by accuracy, yet it has been hard to say whether their decisions track the radiation structure inside a jet. Researchers fit a Kolmogorov-Arnold Network (KAN) locally to a trained model's perturbation response, then read two-body and three-body interaction scores from the surrogate's mixed partial derivatives and rank-correlate them with the dominant contributions to two-point and three-point energy correlation functions. On JetClass, the Particle Transformer shows clear positive two-body alignment while a standard multi-head attention baseline does not; three-body alignment is weaker for both, with ParT still stronger and its error bars excluding null correlation.

Jet taggers in high-energy physics are judged mainly by accuracy, yet it has been hard to say whether their decisions track the radiation structure inside a jet. Researchers fit a Kolmogorov-Arnold Network (KAN) locally to a trained model's perturbation response, then read two-body and three-body interaction scores from the surrogate's mixed partial derivatives and rank-correlate them with the dominant contributions to two-point and three-point energy correlation functions. On JetClass, the Particle Transformer shows clear positive two-body alignment while a standard multi-head attention baseline does not; three-body alignment is weaker for both, with ParT still stronger and its error bars excluding null correlation.

Bifidelity modeling uses cheap low-fidelity simulations to supplement scarce high-fidelity experiments, but the usual product is a single deterministic fit that says little about how trustworthy a prediction is. This work extends a single-fidelity Langevin sampling hypernetwork to bifidelity data, using a shared-latent data model that predicts both the low-fidelity response and the high-fidelity correction while a learned drift network generates parameter samples from the posterior score. On a synthetic hyperelastic case, conditional staged training cut the SPD-drift RMSE from 0.075 to 0.024 and the convex-potential drift from 0.081 to 0.017; on experimental and DNS data, predicting the high-fidelity response directly gave lower error than an additive discrepancy.

Bifidelity modeling uses cheap low-fidelity simulations to supplement scarce high-fidelity experiments, but the usual product is a single deterministic fit that says little about how trustworthy a prediction is. This work extends a single-fidelity Langevin sampling hypernetwork to bifidelity data, using a shared-latent data model that predicts both the low-fidelity response and the high-fidelity correction while a learned drift network generates parameter samples from the posterior score. On a synthetic hyperelastic case, conditional staged training cut the SPD-drift RMSE from 0.075 to 0.024 and the convex-potential drift from 0.081 to 0.017; on experimental and DNS data, predicting the high-fidelity response directly gave lower error than an additive discrepancy.

Autoregressive DNA language models generate left to right, yet genome repair and editing usually condition on both flanks. The CreatiPhage team converted a 144.5M-parameter phage autoregressive model into CPDD, a 147.5M-parameter bidirectional discrete diffusion model; on 522 screened phage coding sequences it restores randomly masked bases more accurately than the comparable diffusion model PhaGen, and in a mixed-corruption test a single forward pass returns 42.4% of visible incorrect bases to the correct base.

Autoregressive DNA language models generate left to right, yet genome repair and editing usually condition on both flanks. The CreatiPhage team converted a 144.5M-parameter phage autoregressive model into CPDD, a 147.5M-parameter bidirectional discrete diffusion model; on 522 screened phage coding sequences it restores randomly masked bases more accurately than the comparable diffusion model PhaGen, and in a mixed-corruption test a single forward pass returns 42.4% of visible incorrect bases to the correct base.

OpenAI's Codex Security lets developers find, investigate and fix code vulnerabilities in one workflow, started from a Codex session or automated with the @openai/codex-security command line or TypeScript SDK. Scans cover whole repositories, selected paths and Git changes, and a deep scan repeats a standard scan and returns deduplicated findings. A completed scan saves a report and findings, and the report lists source evidence and exclusions for judging coverage.

OpenAI's Codex Security lets developers find, investigate and fix code vulnerabilities in one workflow, started from a Codex session or automated with the @openai/codex-security command line or TypeScript SDK. Scans cover whole repositories, selected paths and Git changes, and a deep scan repeats a standard scan and returns deduplicated findings. A completed scan saves a report and findings, and the report lists source evidence and exclusions for judging coverage.

Raysemble splits single-view bone reconstruction into shape and placement: it predicts each bone's shape on a normalized grid, then uses predicted depth and camera geometry to fix its center and scale in millimeters. On simulated frontal radiographs from 43 CT scans of 23 held-out patients, it averages 65.24% Dice, 9.09 mm HD95 and 3.32 mm ASSD before rigid alignment, leading 15 configurations; with identical inputs it beats VolumeNeRF by 5.86 Dice points.

Raysemble splits single-view bone reconstruction into shape and placement: it predicts each bone's shape on a normalized grid, then uses predicted depth and camera geometry to fix its center and scale in millimeters. On simulated frontal radiographs from 43 CT scans of 23 held-out patients, it averages 65.24% Dice, 9.09 mm HD95 and 3.32 mm ASSD before rigid alignment, leading 15 configurations; with identical inputs it beats VolumeNeRF by 5.86 Dice points.

In cross-subject neural-image retrieval, comparing the mean of a few target observations with the global source mean conflates subject-specific representation shifts with differences in which stimuli were sampled. SA-LBC instead estimates a first-order shift by averaging residuals against same-stimulus source-population anchors, then shrinks the correction using source-derived uncertainty. On THINGS-EEG2 with 10 subjects under leave-one-subject-out evaluation, the Attention-CLIP ViT-B/32 configuration rises from 0.1867 MRR without calibration to 0.2013 and 0.2336, though retrieval gains remain representation-dependent.

In cross-subject neural-image retrieval, comparing the mean of a few target observations with the global source mean conflates subject-specific representation shifts with differences in which stimuli were sampled. SA-LBC instead estimates a first-order shift by averaging residuals against same-stimulus source-population anchors, then shrinks the correction using source-derived uncertainty. On THINGS-EEG2 with 10 subjects under leave-one-subject-out evaluation, the Attention-CLIP ViT-B/32 configuration rises from 0.1867 MRR without calibration to 0.2013 and 0.2336, though retrieval gains remain representation-dependent.

Anthropic released Claude Meetings in beta, letting Claude join video calls as a visible participant that follows the conversation and starts work during the meeting. Teams can ask it questions out loud or in chat, it answers in writing by default, and with permission it can use each person's connected tools; it also builds docs and whiteboards alongside the team and writes the notes and action items. Support is currently limited to Google Meet, rolling out first on Team and Enterprise plans, with other platforms coming soon.

Anthropic released Claude Meetings in beta, letting Claude join video calls as a visible participant that follows the conversation and starts work during the meeting. Teams can ask it questions out loud or in chat, it answers in writing by default, and with permission it can use each person's connected tools; it also builds docs and whiteboards alongside the team and writes the notes and action items. Support is currently limited to Google Meet, rolling out first on Team and Enterprise plans, with other platforms coming soon.

Natural-language data requests in clinical research used to require staff to translate them into computable rules, pick data sources and write institution-specific queries. A large-language-model multi-agent pipeline decomposes each investigator request into source-specific tasks, generates and executes SQL, then combines structured records with clinical-note adjudication. Of 50 pediatric orthopedic requests, 43 completed manual review, with adjusted median positive predictive value 0.93, sensitivity within the observed union 0.92 and F1 0.90.

Natural-language data requests in clinical research used to require staff to translate them into computable rules, pick data sources and write institution-specific queries. A large-language-model multi-agent pipeline decomposes each investigator request into source-specific tasks, generates and executes SQL, then combines structured records with clinical-note adjudication. Of 50 pediatric orthopedic requests, 43 completed manual review, with adjusted median positive predictive value 0.93, sensitivity within the observed union 0.92 and F1 0.90.

In a retrospective cohort of 1,286 men with non-obstructive azoospermia at two reproductive medicine centers, intraoperative AI-SpermScope marking of likely spermatogenic tubules and rapid sperm confirmation raised sperm retrieval success to 68.1% versus 47.6% for conventional micro-TESE, a gap that persisted after propensity score matching. Among men with retrieved sperm, the subsequent ICSI live birth rate per transfer cycle was 38.5% in the AI group and 26.8% in the conventional group; group assignment was not randomized, so a causal effect still needs prospective randomized trials.

In a retrospective cohort of 1,286 men with non-obstructive azoospermia at two reproductive medicine centers, intraoperative AI-SpermScope marking of likely spermatogenic tubules and rapid sperm confirmation raised sperm retrieval success to 68.1% versus 47.6% for conventional micro-TESE, a gap that persisted after propensity score matching. Among men with retrieved sperm, the subsequent ICSI live birth rate per transfer cycle was 38.5% in the AI group and 26.8% in the conventional group; group assignment was not randomized, so a causal effect still needs prospective randomized trials.

Thinking Machines' Inkling 2 Small gained vLLM support on its release day. It is a text-and-image-in, text-out mixture-of-experts model with 594B total and 17B active parameters, up to 1M tokens of context, and a bundled DSpark speculative-decoding drafter. vLLM fits it with kernel fusion, programmatic dependent launch, sequence parallelism and a Gluon attention kernel rewritten for the 6:1 GQA ratio, reaching about 630 tok/s for a single user at TP4 on GB200 with acceptance length 3.5, and also supports LoRA and reinforcement learning via SkyRL.

Thinking Machines' Inkling 2 Small gained vLLM support on its release day. It is a text-and-image-in, text-out mixture-of-experts model with 594B total and 17B active parameters, up to 1M tokens of context, and a bundled DSpark speculative-decoding drafter. vLLM fits it with kernel fusion, programmatic dependent launch, sequence parallelism and a Gluon attention kernel rewritten for the 6:1 GQA ratio, reaching about 630 tok/s for a single user at TP4 on GB200 with acceptance length 3.5, and also supports LoRA and reinforcement learning via SkyRL.

Virtual metrology for chemical mechanical planarization (CMP) material removal rate (MRR) has relied on simplified linear assumptions or handcrafted feature engineering, which needs domain knowledge and can discard temporal information in raw sensor signals. GREP first pretrains self-supervised on unlabeled sensor sequences, having the encoder regress global proxy targets projected from complete and half sequences, then fine-tunes on labeled data. On the PHM 2016 CMP dataset it matches a feature-engineered supervised baseline across most labeled-data fractions, lowers error in selected limited-label settings, and cuts RMSE from 11.23 to 9.32 in the final chronological interval.

Virtual metrology for chemical mechanical planarization (CMP) material removal rate (MRR) has relied on simplified linear assumptions or handcrafted feature engineering, which needs domain knowledge and can discard temporal information in raw sensor signals. GREP first pretrains self-supervised on unlabeled sensor sequences, having the encoder regress global proxy targets projected from complete and half sequences, then fine-tunes on labeled data. On the PHM 2016 CMP dataset it matches a feature-engineered supervised baseline across most labeled-data fractions, lowers error in selected limited-label settings, and cuts RMSE from 11.23 to 9.32 in the final chronological interval.

Monocular Gaussian SLAM predicts a full set of Gaussians for every mapping keyframe, so regions already represented get re-predicted and the map keeps growing. Sparse-Mono encodes the current observation, the persistent map rendered from the current viewpoint, and their photometric and geometric discrepancies, then decides where additions are needed and how many before decoding complete Gaussians only at the selected locations. On Replica it reaches an average PSNR 1.05 dB above MonoGS and uses 89.3% fewer Gaussians than VINGS-Mono, at comparable online processing latency.

Monocular Gaussian SLAM predicts a full set of Gaussians for every mapping keyframe, so regions already represented get re-predicted and the map keeps growing. Sparse-Mono encodes the current observation, the persistent map rendered from the current viewpoint, and their photometric and geometric discrepancies, then decides where additions are needed and how many before decoding complete Gaussians only at the selected locations. On Replica it reaches an average PSNR 1.05 dB above MonoGS and uses 89.3% fewer Gaussians than VINGS-Mono, at comparable online processing latency.

In heterogeneous multi-task offline reinforcement learning, tasks differ in state dimension, action space and dynamics, so a directly shared backbone hits both interface mismatch and gradient conflict. HeteroGrad-DT gives each task its own input tokenizer and prediction head around a shared sequence backbone, and a Gradient Flow Transformer pools per-task gradients over a synchronization interval and learns fusion weights to coordinate the shared update. On the DeepMind Control Suite with up to 30 tasks, its average normalized return exceeds representative multi-task methods, with a 32.3% relative gain over standard shared-backbone training and 8.6% over a PCGrad baseline using the same task-aware interfaces.

In heterogeneous multi-task offline reinforcement learning, tasks differ in state dimension, action space and dynamics, so a directly shared backbone hits both interface mismatch and gradient conflict. HeteroGrad-DT gives each task its own input tokenizer and prediction head around a shared sequence backbone, and a Gradient Flow Transformer pools per-task gradients over a synchronization interval and learns fusion weights to coordinate the shared update. On the DeepMind Control Suite with up to 30 tasks, its average normalized return exceeds representative multi-task methods, with a 32.3% relative gain over standard shared-backbone training and 8.6% over a PCGrad baseline using the same task-aware interfaces.

Quantitative phenotype prediction has mostly relied on how single positions correlate with a trait, which cannot capture several residues acting together. The new scheme adds second-order epistasis and position-specific amino acid clustering on top of a Monte Carlo position search, then fits a regression from sequence distances to phenotype distances. Across 13 datasets spanning temperatures, enzyme inhibition, myoglobin concentration and rhodopsin absorption wavelength, the AC-MC-E variant using all three components reached the highest R² in 9 datasets and ranked in the top three in all 13.

Quantitative phenotype prediction has mostly relied on how single positions correlate with a trait, which cannot capture several residues acting together. The new scheme adds second-order epistasis and position-specific amino acid clustering on top of a Monte Carlo position search, then fits a regression from sequence distances to phenotype distances. Across 13 datasets spanning temperatures, enzyme inhibition, myoglobin concentration and rhodopsin absorption wavelength, the AC-MC-E variant using all three components reached the highest R² in 9 datasets and ranked in the top three in all 13.

Autism resting-state fMRI classifiers rarely generalize across acquisition sites, and many reported accuracies are inflated by data leakage. ASD Neuro Agentic (ANA) has language-model agents propose features from the literature and researcher input while fixed tools perform every calculation; across 1,028 participants from 17 ABIDE I centers it searched pipelines and selected a ComBat-harmonized ensemble of tangent-space connectivity models. In leave-one-center-out testing the ensemble reached an AUC of 0.814 (95% CI, 0.773–0.841), above ASD-DiagNet and ASD-SAENet.

Autism resting-state fMRI classifiers rarely generalize across acquisition sites, and many reported accuracies are inflated by data leakage. ASD Neuro Agentic (ANA) has language-model agents propose features from the literature and researcher input while fixed tools perform every calculation; across 1,028 participants from 17 ABIDE I centers it searched pipelines and selected a ComBat-harmonized ensemble of tangent-space connectivity models. In leave-one-center-out testing the ensemble reached an AUC of 0.814 (95% CI, 0.773–0.841), above ASD-DiagNet and ASD-SAENet.

A defined-contribution pension member invests in cash, two nominal bonds, an inflation-linked bond and a stock index in a two-factor market, and may also buy term life insurance before retirement, with trading restricted to a non-empty closed convex constraint set. The approach builds a family of unconstrained artificial markets, computes dual upper bounds with a reparameterized neural network, and recovers feasible lower bounds via nested Monte Carlo and a Hamiltonian quadratic program. Relative duality gaps under two risk-aversion settings are below the SAMS and deep pricing kernel benchmarks, one of them at 2.87%.

A defined-contribution pension member invests in cash, two nominal bonds, an inflation-linked bond and a stock index in a two-factor market, and may also buy term life insurance before retirement, with trading restricted to a non-empty closed convex constraint set. The approach builds a family of unconstrained artificial markets, computes dual upper bounds with a reparameterized neural network, and recovers feasible lower bounds via nested Monte Carlo and a Hamiltonian quadratic program. Relative duality gaps under two risk-aversion settings are below the SAMS and deep pricing kernel benchmarks, one of them at 2.87%.

DINGO-TGR infers binary parameters and six post-Newtonian phase-deformation parameters jointly with neural posterior estimation, then reweights the network's approximate posterior by importance sampling. For GW150914 the inference takes about 38 minutes against roughly 177 hours for bilby-tgr with dynesty, and about 4.4 hours against 238 hours for the higher signal-to-noise GW250114. GW150914 shows a mild offset from general relativity in the leading principal-component mode, but 115 of 1,000 general-relativity injections resembling GW150914 reach or exceed that offset, an empirical upper-tail p-value of about 0.115 that does not reject the general-relativity null at the 5% level.

DINGO-TGR infers binary parameters and six post-Newtonian phase-deformation parameters jointly with neural posterior estimation, then reweights the network's approximate posterior by importance sampling. For GW150914 the inference takes about 38 minutes against roughly 177 hours for bilby-tgr with dynesty, and about 4.4 hours against 238 hours for the higher signal-to-noise GW250114. GW150914 shows a mild offset from general relativity in the leading principal-component mode, but 115 of 1,000 general-relativity injections resembling GW150914 reach or exceed that offset, an empirical upper-tail p-value of about 0.115 that does not reject the general-relativity null at the 5% level.

Cross-modal retinal image rigid registration has largely relied on a two-step pipeline that segments vessels first and detects keypoints second, so registration quality is capped by segmentation accuracy. The MICRO benchmark gathers 1,365 cross-modal retinal image pairs spanning preoperative, intraoperative and postoperative surgery phases, each with expert-annotated control points. Built on that benchmark, the one-step OFS2Reg constructs keypoint topology constraints directly through a frequency-guided structure preservation descriptor and cross-modal vessel structure similarity learning, reaching a 0.784 success rate and 5.08-pixel RMSE on the MICRO-pre subset.

Cross-modal retinal image rigid registration has largely relied on a two-step pipeline that segments vessels first and detects keypoints second, so registration quality is capped by segmentation accuracy. The MICRO benchmark gathers 1,365 cross-modal retinal image pairs spanning preoperative, intraoperative and postoperative surgery phases, each with expert-annotated control points. Built on that benchmark, the one-step OFS2Reg constructs keypoint topology constraints directly through a frequency-guided structure preservation descriptor and cross-modal vessel structure similarity learning, reaching a 0.784 success rate and 5.08-pixel RMSE on the MICRO-pre subset.

F-box proteins are the substrate-recruiting modules of SCF ubiquitin ligases and have long been treated as relatively rigid structural adaptors. This work combines cryo-EM structures, cross-linking mass spectrometry distance restraints and deep-learning conformational ensemble prediction to compare the same proteins across different complexes. FBXL2's substrate-binding domain moves 43 Å relative to its anchor domain, FBXL17 differs by 17 to 33 Å across assembly contexts, and FBXL1 shifts only 4.5 Å; FBXL2 needs at least four states to satisfy its cross-links to CSN.

F-box proteins are the substrate-recruiting modules of SCF ubiquitin ligases and have long been treated as relatively rigid structural adaptors. This work combines cryo-EM structures, cross-linking mass spectrometry distance restraints and deep-learning conformational ensemble prediction to compare the same proteins across different complexes. FBXL2's substrate-binding domain moves 43 Å relative to its anchor domain, FBXL17 differs by 17 to 33 Å across assembly contexts, and FBXL1 shifts only 4.5 Å; FBXL2 needs at least four states to satisfy its cross-links to CSN.

Katachi-PRIME (形') feeds three-band gri galaxy images together with photometry into implicit likelihood inference to return full posteriors for stellar mass, specific star formation rate, redshift, age, dust attenuation and gas-phase metallicity. On roughly 232,000 GSWLC galaxies, image-only posteriors are tighter than those from 5-band photometry and combining both is best; the same image embeddings show that main-sequence and quiescent morphologies retain star formation history continuously for several Gyr, while starbursts and green-valley galaxies keep only the most recent activity.

Katachi-PRIME (形') feeds three-band gri galaxy images together with photometry into implicit likelihood inference to return full posteriors for stellar mass, specific star formation rate, redshift, age, dust attenuation and gas-phase metallicity. On roughly 232,000 GSWLC galaxies, image-only posteriors are tighter than those from 5-band photometry and combining both is best; the same image embeddings show that main-sequence and quiescent morphologies retain star formation history continuously for several Gyr, while starbursts and green-valley galaxies keep only the most recent activity.

Otoendoscopic image synthesis has been limited because text prompts state a diagnosis but not visual details such as hue and texture, leaving appearance variation within a diagnostic class hard to control. CLNA adds a style encoder and dual cross-attention to a Stable Diffusion 2.1 backbone, compressing such non-prompt attributes into a continuous latent whose sampling range is set by a style parameter. Across normal tympanic membrane, otitis media with effusion and chronic otitis media, FID against the held-out internal test set fell from 52.91 for baseline Stable Diffusion to 10.11, and feature-space recall rose from 57.67% to 90.22%.

Otoendoscopic image synthesis has been limited because text prompts state a diagnosis but not visual details such as hue and texture, leaving appearance variation within a diagnostic class hard to control. CLNA adds a style encoder and dual cross-attention to a Stable Diffusion 2.1 backbone, compressing such non-prompt attributes into a continuous latent whose sampling range is set by a style parameter. Across normal tympanic membrane, otitis media with effusion and chronic otitis media, FID against the held-out internal test set fell from 52.91 for baseline Stable Diffusion to 10.11, and feature-space recall rose from 57.67% to 90.22%.

A non-convex objective is written into the potential term of a photonic waveguide array, where light evolves by quantum tunneling and the output intensity distribution marks where the minima sit. Experiments ran on one- and two-dimensional waveguide arrays, preparing a low-energy input state with incomplete local annealing and then adding a varying Hamiltonian to speed the search and resolve the global minimum. The authors report quantum speedup over classical stochastic gradient descent, though the main text frames that comparison as an illustrative exploration-behavior baseline.

A non-convex objective is written into the potential term of a photonic waveguide array, where light evolves by quantum tunneling and the output intensity distribution marks where the minima sit. Experiments ran on one- and two-dimensional waveguide arrays, preparing a low-energy input state with incomplete local annealing and then adding a varying Hamiltonian to speed the search and resolve the global minimum. The authors report quantum speedup over classical stochastic gradient descent, though the main text frames that comparison as an illustrative exploration-behavior baseline.

Graph retrieval for Wikipedia multi-hop question answering no longer needs a large model to run open extraction over the whole corpus: a local GLiNER2 extractor builds the graph under a preset relation schema, then discovers corpus-specific relations from entity pairs the schema left unconnected and injects the resulting triples after the fact. On 1,000 questions each from MuSiQue, 2WikiMultihopQA and HotpotQA, the full pipeline reaches F1 of 0.4577, 0.6637 and 0.7154, above the BGE-based HippoRAG2 baseline; on MuSiQue, adaptive expansion adds 93 relation types and 1,691 triples and raises F1 by 0.0225 over the seed-only graph. Final answer generation still calls GPT-4o-mini.

Graph retrieval for Wikipedia multi-hop question answering no longer needs a large model to run open extraction over the whole corpus: a local GLiNER2 extractor builds the graph under a preset relation schema, then discovers corpus-specific relations from entity pairs the schema left unconnected and injects the resulting triples after the fact. On 1,000 questions each from MuSiQue, 2WikiMultihopQA and HotpotQA, the full pipeline reaches F1 of 0.4577, 0.6637 and 0.7154, above the BGE-based HippoRAG2 baseline; on MuSiQue, adaptive expansion adds 93 relation types and 1,691 triples and raises F1 by 0.0225 over the seed-only graph. Final answer generation still calls GPT-4o-mini.

Protein stability-change predictors often disagree with themselves: swapping the two sequences does not flip the sign of the prediction, and paths that detour through intermediate states at the same site fail to close. Without retraining, EvoPhi-DDG rewrites the family-alignment predictor MSAesm_ddG as a difference of an effective potential conditioned on the mutation site and alignment context, giving exact sign reversal and exact same-site additivity under a fixed alignment context. Averaged with equal weights with the published JanusDDG, it reaches a Pearson correlation of 0.59 on S669, against 0.55 for JanusDDG and 0.53 for EvoPhi-DDG.

Protein stability-change predictors often disagree with themselves: swapping the two sequences does not flip the sign of the prediction, and paths that detour through intermediate states at the same site fail to close. Without retraining, EvoPhi-DDG rewrites the family-alignment predictor MSAesm_ddG as a difference of an effective potential conditioned on the mutation site and alignment context, giving exact sign reversal and exact same-site additivity under a fixed alignment context. Averaged with equal weights with the published JanusDDG, it reaches a Pearson correlation of 0.59 on S669, against 0.55 for JanusDDG and 0.53 for EvoPhi-DDG.

At its New York Summit, ElevenLabs extended 1 Million Voices from speech to song: people with permanent voice loss can now use ElevenMusic free to create and release music in their own voices. The program had focused on speech and has helped more than 11,000 people with ALS, motor neurone disease, head and neck cancer and other conditions speak again; Songwriters of North America verifies eligibility and distributes access, while NYU Langone Health and Wounded Warrior Project join as partners.

At its New York Summit, ElevenLabs extended 1 Million Voices from speech to song: people with permanent voice loss can now use ElevenMusic free to create and release music in their own voices. The program had focused on speech and has helped more than 11,000 people with ALS, motor neurone disease, head and neck cancer and other conditions speak again; Songwriters of North America verifies eligibility and distributes access, while NYU Langone Health and Wounded Warrior Project join as partners.

Small language models for clinical decision support often answer at far greater length than the question requires after fine-tuning, and existing length control relies on preference data, reward models or proprietary teachers. PECA adds a differentiable conciseness term to the QLoRA supervised fine-tuning objective: once generation passes a token budget, it raises the end-of-sequence probability, with no preference data, no auxiliary model and a single commodity GPU. On the 1.7-billion-parameter Qwen3-1.7B-Instruct, PECA reduces mean response length from 256.0 to 69.7 tokens, a 72.8% cut, while raising MedMCQA multiple-choice diagnostic accuracy from 40.8% to 44.0%.

Small language models for clinical decision support often answer at far greater length than the question requires after fine-tuning, and existing length control relies on preference data, reward models or proprietary teachers. PECA adds a differentiable conciseness term to the QLoRA supervised fine-tuning objective: once generation passes a token budget, it raises the end-of-sequence probability, with no preference data, no auxiliary model and a single commodity GPU. On the 1.7-billion-parameter Qwen3-1.7B-Instruct, PECA reduces mean response length from 256.0 to 69.7 tokens, a 72.8% cut, while raising MedMCQA multiple-choice diagnostic accuracy from 40.8% to 44.0%.

Three large language models independently reproved all 465 propositions of Euclid's Elements, and their citation choices differ systematically from Heath's human-proof baseline. Mean pairwise Jaccard similarity among three Claude Opus 4.6 runs is 0.68, nearly twice the 0.37 between AI and human proofs, so the difference is stable and reproducible. Its concrete signature is that about 40% of AI citations point to foundational elements such as definitions, common notions and postulates, against only 13% in human proofs.

Three large language models independently reproved all 465 propositions of Euclid's Elements, and their citation choices differ systematically from Heath's human-proof baseline. Mean pairwise Jaccard similarity among three Claude Opus 4.6 runs is 0.68, nearly twice the 0.37 between AI and human proofs, so the difference is stable and reproducible. Its concrete signature is that about 40% of AI citations point to foundational elements such as definitions, common notions and postulates, against only 13% in human proofs.

InertiaMPDD extends depression screening beyond a single recording-level label to a measurable departure from a healthy reference: a latent predictor of the next state is trained on frozen multimodal representations, and three descriptors are read from its prediction residuals, namely reference deviation, residual decay rate and a Jacobian spectral radius, used either for standalone screening or as an individual correction to a population decision. Across five cohorts from four datasets, the fused pathway reaches 98.41 F1 on D-Vlog and 84.98 F1 on LMVD, while on AVEC2014 the reference descriptors discriminate weakly and fusion reaches 66.67 F1.

InertiaMPDD extends depression screening beyond a single recording-level label to a measurable departure from a healthy reference: a latent predictor of the next state is trained on frozen multimodal representations, and three descriptors are read from its prediction residuals, namely reference deviation, residual decay rate and a Jacobian spectral radius, used either for standalone screening or as an individual correction to a population decision. Across five cohorts from four datasets, the fused pathway reaches 98.41 F1 on D-Vlog and 84.98 F1 on LMVD, while on AVEC2014 the reference descriptors discriminate weakly and fusion reaches 66.67 F1.

Retail media networks usually compress clickstreams into click-through, conversion, recency and similar summaries, so two shoppers with identical event counts become indistinguishable. Across three campaign waves on an anonymized omni-channel retailer's platform covering 5.2 million active members, purchase-intention inference is framed as theory-coherent hidden semi-Markov dynamics: states are declared by the Theory of Planned Behavior, the Consumer Decision Journey and the Persuasion Knowledge Model, a retrieval-augmented language model supplies a reference posterior, and a KL regularizer in the expectation step stabilizes state assignment. On n=1,008 paired test consumers the model reaches one-step purchase AUC 0.717 against 0.635 for an action-only covariate baseline.

Retail media networks usually compress clickstreams into click-through, conversion, recency and similar summaries, so two shoppers with identical event counts become indistinguishable. Across three campaign waves on an anonymized omni-channel retailer's platform covering 5.2 million active members, purchase-intention inference is framed as theory-coherent hidden semi-Markov dynamics: states are declared by the Theory of Planned Behavior, the Consumer Decision Journey and the Persuasion Knowledge Model, a retrieval-augmented language model supplies a reference posterior, and a KL regularizer in the expectation step stabilizes state assignment. On n=1,008 paired test consumers the model reaches one-step purchase AUC 0.717 against 0.635 for an action-only covariate baseline.

J.Crew launched Style Advisor, an AI stylist on JCrew.com that lets customers shop by complete outfit rather than one product at a time. Online shopping usually handles a single item, so a customer who finds a sweater must search again for matching pants and shoes. Built by J.Crew's team with Claude Code and several Claude models, Style Advisor takes a described occasion and returns complete looks drawn from J.Crew's own photo shoots, shown as photos and open to further refinement; it is available now for women's and men's styles in the U.S.

J.Crew launched Style Advisor, an AI stylist on JCrew.com that lets customers shop by complete outfit rather than one product at a time. Online shopping usually handles a single item, so a customer who finds a sweater must search again for matching pants and shoes. Built by J.Crew's team with Claude Code and several Claude models, Style Advisor takes a described occasion and returns complete looks drawn from J.Crew's own photo shoots, shown as photos and open to further refinement; it is available now for women's and men's styles in the U.S.

RGB-X tracking pairs visible frames with thermal, depth or event data, but paired annotations are far scarcer than RGB pretraining data, so only a small set of parameters is usually tuned. Earlier bidirectional cross-modal corrections computed both directions side by side from uncorrected input states, so the second correction could not use the first one's result. CCLT inserts low-rank exchange modules at blocks 6 and 10 of a frozen backbone, letting one direction correct first and the other read the freshly corrected state; against a parameter-matched parallel version, mean LasHeR SR rises by 1.980 points (95% interval [0.360, 2.690]), with 196.6K trainable parameters, 0.212% of the 92.7M-parameter tracker.

RGB-X tracking pairs visible frames with thermal, depth or event data, but paired annotations are far scarcer than RGB pretraining data, so only a small set of parameters is usually tuned. Earlier bidirectional cross-modal corrections computed both directions side by side from uncorrected input states, so the second correction could not use the first one's result. CCLT inserts low-rank exchange modules at blocks 6 and 10 of a frozen backbone, letting one direction correct first and the other read the freshly corrected state; against a parameter-matched parallel version, mean LasHeR SR rises by 1.980 points (95% interval [0.360, 2.690]), with 196.6K trainable parameters, 0.212% of the 92.7M-parameter tracker.