AI and science frontiers · 2026-11-14

Whether multimodal training changes how a model reasons over text alone had been supported mainly by behavior. Across 10 multimodal configurations from 6 model families, this work builds a vision–language direction and removes each hidden state's projection onto it at inference; in Chameleon-7B, BBH navigate drops while boolean expressions stay largely intact, a double dissociation. The effect is direction-specific and layer-localized, and extends to mental-simulation tasks such as mental rotation, paper folding and perspective taking.

Whether multimodal training changes how a model reasons over text alone had been supported mainly by behavior. Across 10 multimodal configurations from 6 model families, this work builds a vision–language direction and removes each hidden state's projection onto it at inference; in Chameleon-7B, BBH navigate drops while boolean expressions stay largely intact, a double dissociation. The effect is direction-specific and layer-localized, and extends to mental-simulation tasks such as mental rotation, paper folding and perspective taking.

High-accuracy gravitational-wave source modeling has long avoided machine learning because neural networks carry a noise floor. Researchers trained a graph neural network as a preconditioner for an iterative solver, leaving the traditional GMRES to enforce final accuracy while the network only accelerates convergence. On the high-frequency elliptic equations for asymmetric-mass black hole binaries, the hardest parameter regions need 2 to 15 fewer GMRES iterations than basic ILU and 2 to 4 times less wall time than the next best method.

High-accuracy gravitational-wave source modeling has long avoided machine learning because neural networks carry a noise floor. Researchers trained a graph neural network as a preconditioner for an iterative solver, leaving the traditional GMRES to enforce final accuracy while the network only accelerates convergence. On the high-frequency elliptic equations for asymmetric-mass black hole binaries, the hardest parameter regions need 2 to 15 fewer GMRES iterations than basic ILU and 2 to 4 times less wall time than the next best method.

An engineering design study holds only tens to hundreds of simulations, each on its own mesh, and conventional surrogates are retrained for every study. PhysICL is pretrained once on a procedural prior, then treats a few labeled simulations as context and predicts fields, with quantiles, directly at each query's own unstructured-mesh nodes in one forward pass, with no weight update. On five PLAID datasets unseen in pretraining its field error falls as context grows and drops below the nearest context simulation; on AhmedML surfaces its peak GPU memory stays at 4.27 GiB.

An engineering design study holds only tens to hundreds of simulations, each on its own mesh, and conventional surrogates are retrained for every study. PhysICL is pretrained once on a procedural prior, then treats a few labeled simulations as context and predicts fields, with quantiles, directly at each query's own unstructured-mesh nodes in one forward pass, with no weight update. On five PLAID datasets unseen in pretraining its field error falls as context grows and drops below the nearest context simulation; on AhmedML surfaces its peak GPU memory stays at 4.27 GiB.

Hyperbolic space suits hierarchical embeddings, but earlier hyperbolic Gaussian-process latent variable models had to approximate every kernel entry numerically and optimize auxiliary variables jointly, which became impractical at scale. RHFF-LVM instead approximates hyperbolic kernels with explicit finite-dimensional random features derived from the Helgason–Fourier transform, modeling observations as feature projections of the latents and removing numerical quadrature, variational bounds and inducing points. On a synthetic binary tree, its embedded distances reach a Spearman correlation of 84.3 with true path lengths, versus 74.9 for the Euclidean counterpart RFF-LVM and 82.4 for the strongest hyperbolic baseline, Poincaré Map.

Hyperbolic space suits hierarchical embeddings, but earlier hyperbolic Gaussian-process latent variable models had to approximate every kernel entry numerically and optimize auxiliary variables jointly, which became impractical at scale. RHFF-LVM instead approximates hyperbolic kernels with explicit finite-dimensional random features derived from the Helgason–Fourier transform, modeling observations as feature projections of the latents and removing numerical quadrature, variational bounds and inducing points. On a synthetic binary tree, its embedded distances reach a Spearman correlation of 84.3 with true path lengths, versus 74.9 for the Euclidean counterpart RFF-LVM and 82.4 for the strongest hyperbolic baseline, Poincaré Map.

Science Prediction Universe (SPU) places AI agents in a real wet-lab prediction setting: given only experimental conditions and candidates, an agent has eight hours to search the literature, run calculations and rank candidates, which are then scored against measurements by Spearman correlation. The 100 public tasks come from recent Nature Portfolio papers and the 16 private tasks from unpublished experiments, so retrieval cannot help. The best of thirteen frontier-model configurations beats chance on only about half its scored tasks, and more reasoning effort or more tool families brought no detectable score gain.

Science Prediction Universe (SPU) places AI agents in a real wet-lab prediction setting: given only experimental conditions and candidates, an agent has eight hours to search the literature, run calculations and rank candidates, which are then scored against measurements by Spearman correlation. The 100 public tasks come from recent Nature Portfolio papers and the 16 private tasks from unpublished experiments, so retrieval cannot help. The best of thirteen frontier-model configurations beats chance on only about half its scored tasks, and more reasoning effort or more tool families brought no detectable score gain.

SIM replaces clinical raw k-space with training samples generated from multi-coil MRI physics, then scales data and model capacity together in a generic Transformer. Across 14 public and in-house datasets spanning seven centers, five vendors, 0.3 to 7 T and more than ten protocols, its zero-shot reconstruction beats the second-best method by 1.06 to 3.95 dB PSNR; in a 150-scan blinded reader study, SIM reached 100% diagnostic usability at an acceleration factor the source does not state.

SIM replaces clinical raw k-space with training samples generated from multi-coil MRI physics, then scales data and model capacity together in a generic Transformer. Across 14 public and in-house datasets spanning seven centers, five vendors, 0.3 to 7 T and more than ten protocols, its zero-shot reconstruction beats the second-best method by 1.06 to 3.95 dB PSNR; in a 150-scan blinded reader study, SIM reached 100% diagnostic usability at an acceleration factor the source does not state.

Probabilistic neurosymbolic inference combines neural predictions with symbolic constraints by computing a weighted model count, which is #P-hard exactly, and existing methods each place approximation at a different stage. APRICOT unifies bottom-up compilation into a three-step pipeline and instantiates factor compilation and product recompression with tensor trains, yielding weighted model count estimates orders of magnitude more accurate than prior methods across six tasks in Sudoku, arithmetic reasoning and scene understanding, with better downstream learning where symbolic factors are tightly coupled.

Probabilistic neurosymbolic inference combines neural predictions with symbolic constraints by computing a weighted model count, which is #P-hard exactly, and existing methods each place approximation at a different stage. APRICOT unifies bottom-up compilation into a three-step pipeline and instantiates factor compilation and product recompression with tensor trains, yielding weighted model count estimates orders of magnitude more accurate than prior methods across six tasks in Sudoku, arithmetic reasoning and scene understanding, with better downstream learning where symbolic factors are tightly coupled.

Pretrained generative models are usually screened only over complete outputs already produced, wasting feasible combinations visible in local scores. Structured Re-Decoding (SRD) keeps model parameters fixed while jointly selecting local candidates from scores or samples under task constraints, reconstructing and verifying the result. On FLOWR molecular generation at the same generation budget, single-target composition-and-quality success rises from 419 to 1,309 of 1,800 requests over best-of-4, and dual-target from 119 to 615; the quality pass rate falls from 99.833% to 94.389%.

Pretrained generative models are usually screened only over complete outputs already produced, wasting feasible combinations visible in local scores. Structured Re-Decoding (SRD) keeps model parameters fixed while jointly selecting local candidates from scores or samples under task constraints, reconstructing and verifying the result. On FLOWR molecular generation at the same generation budget, single-target composition-and-quality success rises from 419 to 1,309 of 1,800 requests over best-of-4, and dual-target from 119 to 615; the quality pass rate falls from 99.833% to 94.389%.

Coronary inflammation could previously be measured only on contrast-enhanced coronary CT angiography, while routine calcium scoring uses low-dose noncontrast CT. A deep-learning pipeline trained and externally validated in a multicenter cohort computes a pericoronary fat attenuation index (FAI) score from noncontrast CT; the score performed comparably to the contrast-based FAI and further separated cardiovascular death and major adverse cardiovascular events among people with a calcium score of zero.

Coronary inflammation could previously be measured only on contrast-enhanced coronary CT angiography, while routine calcium scoring uses low-dose noncontrast CT. A deep-learning pipeline trained and externally validated in a multicenter cohort computes a pericoronary fat attenuation index (FAI) score from noncontrast CT; the score performed comparably to the contrast-based FAI and further separated cardiovascular death and major adverse cardiovascular events among people with a calcium score of zero.

In high-barrier cis-trans isomerization the committor, the probability that a configuration reaches product before reactant, is nearly a step function with intermediate values only in a narrow band around the transition state. Combining AIMMD path sampling with symbolic regression, researchers learn the committor from unbiased molecular dynamics trajectories for N-retinylidene-lysine and the p-coumaric acid anion in vacuo and distill it into closed-form expressions. Four proper dihedrals around the reactive bond must couple nonlinearly to reproduce the S-shaped stepwise path seen in the transition-path ensemble, whereas the minimum free-energy path is a straight line.

In high-barrier cis-trans isomerization the committor, the probability that a configuration reaches product before reactant, is nearly a step function with intermediate values only in a narrow band around the transition state. Combining AIMMD path sampling with symbolic regression, researchers learn the committor from unbiased molecular dynamics trajectories for N-retinylidene-lysine and the p-coumaric acid anion in vacuo and distill it into closed-form expressions. Four proper dihedrals around the reactive bond must couple nonlinearly to reproduce the S-shaped stepwise path seen in the transition-path ensemble, whereas the minimum free-energy path is a straight line.

Agentic reinforcement-learning-as-a-service hands data sampling and reward rating to the user, and a harmful reward function can erase a model's safety alignment at a cost as low as $20. Bait mixes explicit harmful responses from a separately trained Bait model into the policy model's response group during rollout, uses the sign of the reward gap between Bait and non-Bait responses (reward divergence) to flag a harmful reward function, and flips that group's reward sign to penalize learning when it detects one. On the authors' testbed, Bait cut the average harmful score across three harmful-instruction datasets from 20.4 to 6.8 while average agentic accuracy rose from 93.1 to 96.0.

Agentic reinforcement-learning-as-a-service hands data sampling and reward rating to the user, and a harmful reward function can erase a model's safety alignment at a cost as low as $20. Bait mixes explicit harmful responses from a separately trained Bait model into the policy model's response group during rollout, uses the sign of the reward gap between Bait and non-Bait responses (reward divergence) to flag a harmful reward function, and flips that group's reward sign to penalize learning when it detects one. On the authors' testbed, Bait cut the average harmful score across three harmful-instruction datasets from 20.4 to 6.8 while average agentic accuracy rose from 93.1 to 96.0.

Third-party skills in public registries can be exploited by a poisoning method that edits only decision rules while leaving the task and interface intact. Without real user requests or ground-truth answers, StealthyPoison uses synthetic requests and clean executions by a surrogate agent as references to iteratively rewrite formulas, thresholds and branches. Across 566 eligible skills, 5,244 requests and more than 26,000 registry entries, the frozen poisoned skills reached micro-averaged attack success of 31.0%–31.9% on three held-out models, close to the surrogate's 32.9%.

Third-party skills in public registries can be exploited by a poisoning method that edits only decision rules while leaving the task and interface intact. Without real user requests or ground-truth answers, StealthyPoison uses synthetic requests and clean executions by a surrogate agent as references to iteratively rewrite formulas, thresholds and branches. Across 566 eligible skills, 5,244 requests and more than 26,000 registry entries, the frozen poisoned skills reached micro-averaged attack success of 31.0%–31.9% on three held-out models, close to the surrogate's 32.9%.

KL-regularized reinforcement learning, the objective behind maximum-entropy RL and RL fine-tuning of language models from human feedback, previously had guarantees whose sample counts grew with a worse power of the effective horizon, and no MDP lower bound depended on the temperature. For tabular discounted MDPs with a generative model, this work shows that a plug-in estimator, which draws a fixed number of samples per state-action pair and solves the empirical MDP, reaches KL suboptimality at the temperature scale with samples linear in the effective horizon, and that no algorithm, even one choosing where to sample, can do with fewer.

KL-regularized reinforcement learning, the objective behind maximum-entropy RL and RL fine-tuning of language models from human feedback, previously had guarantees whose sample counts grew with a worse power of the effective horizon, and no MDP lower bound depended on the temperature. For tabular discounted MDPs with a generative model, this work shows that a plug-in estimator, which draws a fixed number of samples per state-action pair and solves the empirical MDP, reaches KL suboptimality at the temperature scale with samples linear in the effective horizon, and that no algorithm, even one choosing where to sample, can do with fewer.

Screening superconductors has long rested on a single experimental parameter, the critical temperature, while quantities that decide usefulness, such as the upper critical field and critical current density, sit unstructured in the literature. AutoSuperCon uses a large language model to extract eight experimental superconducting parameters from more than 47,000 open arXiv preprints in the superconductivity category, adds digitized entries from the Roberts 1976 survey, and yields 20,354 entries for 6,793 superconductors. On the 2,439 materials shared with NIMS SuperCon, critical temperatures show a median deviation of 0.25 K, or 3.5% relative; the database supports large-scale tests of empirical relations and composition-based prediction of critical temperature and upper critical field.

Screening superconductors has long rested on a single experimental parameter, the critical temperature, while quantities that decide usefulness, such as the upper critical field and critical current density, sit unstructured in the literature. AutoSuperCon uses a large language model to extract eight experimental superconducting parameters from more than 47,000 open arXiv preprints in the superconductivity category, adds digitized entries from the Roberts 1976 survey, and yields 20,354 entries for 6,793 superconductors. On the 2,439 materials shared with NIMS SuperCon, critical temperatures show a median deviation of 0.25 K, or 3.5% relative; the database supports large-scale tests of empirical relations and composition-based prediction of critical temperature and upper critical field.

Meta-learning, which exposes a network to many related functions so it can later learn similar ones quickly, leaves the adapted parameter vectors of that network in an additive geometric relationship. On the ABgeom sequence-generation task, randomly initialized transformers reached a mean F-score of 0.062 after five training batches, while meta-learned initializations reached 1.00 after one step; parameter vectors recombined by symbolic analogy produced networks that implemented the intended functions without direct training, at a mean F-score of 1.00 over 100 analogies. The effect appeared across transformers and LSTMs and three function domains, but the analogies hold only around a single initialization.

Meta-learning, which exposes a network to many related functions so it can later learn similar ones quickly, leaves the adapted parameter vectors of that network in an additive geometric relationship. On the ABgeom sequence-generation task, randomly initialized transformers reached a mean F-score of 0.062 after five training batches, while meta-learned initializations reached 1.00 after one step; parameter vectors recombined by symbolic analogy produced networks that implemented the intended functions without direct training, at a mean F-score of 1.00 over 100 analogies. The effect appeared across transformers and LSTMs and three function domains, but the analogies hold only around a single initialization.

Neuralign places neural recordings from different sessions and animals, with different sampled neurons, into one tensor decomposition in which each component shares a single temporal curve while trial strength and neuron weights are estimated per session, so no matched neurons or matched trials are needed. On 93 brain-wide recordings from 19 mice, an unsupervised fit yielded a single component that separated the three anterior lateral motor cortex (ALM) optogenetic inhibition conditions, with balanced accuracy of 0.647 on held-out sessions against a chance level of 1/3. That component's change showed a modest association with the drop in behavioural accuracy (ρ = 0.266 over 180 session–inhibition observations).

Neuralign places neural recordings from different sessions and animals, with different sampled neurons, into one tensor decomposition in which each component shares a single temporal curve while trial strength and neuron weights are estimated per session, so no matched neurons or matched trials are needed. On 93 brain-wide recordings from 19 mice, an unsupervised fit yielded a single component that separated the three anterior lateral motor cortex (ALM) optogenetic inhibition conditions, with balanced accuracy of 0.647 on held-out sessions against a chance level of 1/3. That component's change showed a modest association with the drop in behavioural accuracy (ρ = 0.266 over 180 session–inhibition observations).

REST2 enhanced sampling had been used with classical force fields and QM/MM, but not with multiscale ML/MM simulations whose solute is described by a neural-network potential. Exploiting the AMP-BMS Hamiltonian's split into solute–solute, solute–solvent and solvent–solvent terms, the approach scales the first two energies directly and exchanges replicas while holding the solvent at physical temperature. In explicit solvent, chignolin and CLN025 started from extended conformations reach Cα-RMSD plateaus of about 0.2–0.3 nm and about 0.2 nm relative to their NMR structures, with no reaction coordinate specified in advance.

REST2 enhanced sampling had been used with classical force fields and QM/MM, but not with multiscale ML/MM simulations whose solute is described by a neural-network potential. Exploiting the AMP-BMS Hamiltonian's split into solute–solute, solute–solvent and solvent–solvent terms, the approach scales the first two energies directly and exchanges replicas while holding the solvent at physical temperature. In explicit solvent, chignolin and CLN025 started from extended conformations reach Cα-RMSD plateaus of about 0.2–0.3 nm and about 0.2 nm relative to their NMR structures, with no reaction coordinate specified in advance.

When nanopore direct RNA sequencing reads tRNAs, modifications corrupt basecalling, forcing sequence-mapping pipelines to trade classification accuracy against usable reads. tyCooNN (nano tRNA classifier) instead classifies tRNAs directly from ionic current signals, bypassing basecalling and alignment; its ntc48 model reaches 98.2% classification accuracy across all 48 E. coli tRNA species. An equimolar mixture of 44 tRNAs calibrates and corrects read-count bias, yielding absolute abundances, and binary classifiers call specific modifications at the single-molecule level.

When nanopore direct RNA sequencing reads tRNAs, modifications corrupt basecalling, forcing sequence-mapping pipelines to trade classification accuracy against usable reads. tyCooNN (nano tRNA classifier) instead classifies tRNAs directly from ionic current signals, bypassing basecalling and alignment; its ntc48 model reaches 98.2% classification accuracy across all 48 E. coli tRNA species. An equimolar mixture of 44 tRNAs calibrates and corrects read-count bias, yielding absolute abundances, and binary classifiers call specific modifications at the single-molecule level.

Convergence analyses of synchronous Q-learning usually keep only the Gaussian limit, and central-limit normalization removes the smaller-scale displacement produced by the initial nonlinear evolution. This analysis splits the last-iterate error into an optimal-policy temporal-difference component and a nonnegative displacement caused by maximization, proves the latter converges almost surely to a random limit at the n^-1/4 scale, and uses an Edgeworth expansion to give corrections to the Gaussian approximation. In a two-state example, the first-order corrected density tracks the empirical distribution better than the central-limit Gaussian.

Convergence analyses of synchronous Q-learning usually keep only the Gaussian limit, and central-limit normalization removes the smaller-scale displacement produced by the initial nonlinear evolution. This analysis splits the last-iterate error into an optimal-policy temporal-difference component and a nonnegative displacement caused by maximization, proves the latter converges almost surely to a random limit at the n^-1/4 scale, and uses an Edgeworth expansion to give corrections to the Gaussian approximation. In a two-state example, the first-order corrected density tracks the empirical distribution better than the central-limit Gaussian.

In sharded data-parallel training, the packed low-precision layouts that compute kernels want and the update units that optimizers need have been tied together; SS-HSDP separates runtime parameters from update parameters and uses shared data views and hierarchical sharding to avoid redundant movement. On MXFP8 video and language-model workloads at 128 GPUs, MFU rises by up to 21.7% over the fastest baseline, with peak memory comparable to or lower than the lowest-memory baseline.

In sharded data-parallel training, the packed low-precision layouts that compute kernels want and the update units that optimizers need have been tied together; SS-HSDP separates runtime parameters from update parameters and uses shared data views and hierarchical sharding to avoid redundant movement. On MXFP8 video and language-model workloads at 128 GPUs, MFU rises by up to 21.7% over the fastest baseline, with peak memory comparable to or lower than the lowest-memory baseline.

Autonomous systems usually monitor motion and whether a task is complete, leaving the information-gathering process itself unchecked, so a system can keep operating without resolving outstanding requirements. Acquisition-State Telemetry (AST) maps raw observations onto task-requirement resolution coordinates, and a fixed causal monitor emits two continuous signals: prospective progress across unresolved requirements and informational friction when recurrent evidence yields negligible gain. One unretuned configuration ran across eight domains; the friction channel separated unproductive stagnation from purposeful exploration before completion and stayed muted afterward, and in closed-loop reinforcement learning under early termination, penalizing friction sustained high mission success.

Autonomous systems usually monitor motion and whether a task is complete, leaving the information-gathering process itself unchecked, so a system can keep operating without resolving outstanding requirements. Acquisition-State Telemetry (AST) maps raw observations onto task-requirement resolution coordinates, and a fixed causal monitor emits two continuous signals: prospective progress across unresolved requirements and informational friction when recurrent evidence yields negligible gain. One unretuned configuration ran across eight domains; the friction channel separated unproductive stagnation from purposeful exploration before completion and stayed muted afterward, and in closed-loop reinforcement learning under early termination, penalizing friction sustained high mission success.

Offline reinforcement learning usually trains its critic by regressing onto a bootstrapped target, stabilized with a target network. AL-LBLP keeps the Lifted Bellman Linear Programming form, which writes in-sample Bellman optimality as a linear program over the dataset, and solves it with the method of multipliers: the fixed hinge coefficients become multipliers learned by dual ascent, leaving one penalty parameter. With a single critic, no target network and no ensemble, a Gaussian actor, a flow actor and best-of-n sampling average 55.8, 55.7 and 60.1 on the 50 state-based OGBench tasks.

Offline reinforcement learning usually trains its critic by regressing onto a bootstrapped target, stabilized with a target network. AL-LBLP keeps the Lifted Bellman Linear Programming form, which writes in-sample Bellman optimality as a linear program over the dataset, and solves it with the method of multipliers: the fixed hinge coefficients become multipliers learned by dual ascent, leaving one penalty parameter. With a single critic, no target network and no ensemble, a Gaussian actor, a flow actor and best-of-n sampling average 55.8, 55.7 and 60.1 on the 50 state-based OGBench tasks.

Multi-robot task planning has mostly let an end-to-end vision-language model learn both scene judgment and action structure, and out-of-distribution performance can collapse. NSDL splits the labor: a trained small model decides which goals are already met, how the task splits and which robot does what, and writes these as a plan sketch; a deterministic solver expands the sketch into an executable joint plan. On the public VIKI-L2 benchmark the 7B model reaches 97.0% out-of-distribution accuracy, and three backbones from 3B to 8B all outperform end-to-end and formalize-then-plan baselines.

Multi-robot task planning has mostly let an end-to-end vision-language model learn both scene judgment and action structure, and out-of-distribution performance can collapse. NSDL splits the labor: a trained small model decides which goals are already met, how the task splits and which robot does what, and writes these as a plan sketch; a deterministic solver expands the sketch into an executable joint plan. On the public VIKI-L2 benchmark the 7B model reaches 97.0% out-of-distribution accuracy, and three backbones from 3B to 8B all outperform end-to-end and formalize-then-plan baselines.

ProofCoder, a proof-synthesis system for Rocq software verification, stops letting an external controller dictate when to retrieve premises, backtrack or repair; instead the model writes down its reading of the proof state before each tool call and decides whether to continue, query or roll back to a given line. On 280 verification lemmas from the AutoRocq benchmark, its best configuration raises proof success from AutoRocq's 47% to 60% at roughly half the inference cost, and removing continuous reasoning leaves the agent solving fewer lemmas at higher cost with four to six times more backtracking.

ProofCoder, a proof-synthesis system for Rocq software verification, stops letting an external controller dictate when to retrieve premises, backtrack or repair; instead the model writes down its reading of the proof state before each tool call and decides whether to continue, query or roll back to a given line. On 280 verification lemmas from the AutoRocq benchmark, its best configuration raises proof success from AutoRocq's 47% to 60% at roughly half the inference cost, and removing continuous reasoning leaves the agent solving fewer lemmas at higher cost with four to six times more backtracking.

Kosmo-GFM combines a geometry foundation model, GFM, with a pose optimization engine, PoseEngine: GFM predicts cameras, depth, cross-view correspondences and dynamic 3D points from images and videos, and PoseEngine folds those priors into optimization. On the 34 scenes completed by every compared method, PoseEngine reaches higher mean pose accuracy than GLUEMAP and cuts cumulative reconstruction time from 69.67 hours to 14.07 hours; the fast variant PoseEngine Flash takes 2.62 hours while keeping complete image registration.

Kosmo-GFM combines a geometry foundation model, GFM, with a pose optimization engine, PoseEngine: GFM predicts cameras, depth, cross-view correspondences and dynamic 3D points from images and videos, and PoseEngine folds those priors into optimization. On the 34 scenes completed by every compared method, PoseEngine reaches higher mean pose accuracy than GLUEMAP and cuts cumulative reconstruction time from 69.67 hours to 14.07 hours; the fast variant PoseEngine Flash takes 2.62 hours while keeping complete image registration.

Clinical images arrive in batches by institution, modality and annotated structure, so a segmentation model must absorb a new imaging distribution and new lesion classes without forgetting what it already learned. FGCARL puts bidirectional feature-statistics alignment and Fisher-weighted parameter constraints into one incremental framework that needs no replay of historical patient data. Moving from UCSF-BMSR to BraTS-METS 2023, where domain and classes shift together, FGCARL with Swin UNETR reaches 68.28% mean Dice across the three tumor regions of the new domain, against 60.94% for the strongest comparator, CLAMTS.

Clinical images arrive in batches by institution, modality and annotated structure, so a segmentation model must absorb a new imaging distribution and new lesion classes without forgetting what it already learned. FGCARL puts bidirectional feature-statistics alignment and Fisher-weighted parameter constraints into one incremental framework that needs no replay of historical patient data. Moving from UCSF-BMSR to BraTS-METS 2023, where domain and classes shift together, FGCARL with Swin UNETR reaches 68.28% mean Dice across the three tumor regions of the new domain, against 60.94% for the strongest comparator, CLAMTS.

Adaptive multiscale simulations have long faced a dilemma: when identical solvent molecules switch between resolutions, nonphysical transition forces appear, forcing a choice between distorted solvation structure and broken Hamiltonian conservation. The Integrated Molecular Model (IMM) uses a multilayer perceptron that takes only the multiscale weight functions as input to learn and cancel those transition forces, yielding artifact-free solvation structures in bulk water, aqueous K+, Na+, NH4+, Ca2+, Cl– and H3O+ systems while conserving energy over long NVE runs. The correction force costs about 1.8% of the time per step, and a model trained on K+ transfers to Na+ and NH4+ without retraining.

Adaptive multiscale simulations have long faced a dilemma: when identical solvent molecules switch between resolutions, nonphysical transition forces appear, forcing a choice between distorted solvation structure and broken Hamiltonian conservation. The Integrated Molecular Model (IMM) uses a multilayer perceptron that takes only the multiscale weight functions as input to learn and cancel those transition forces, yielding artifact-free solvation structures in bulk water, aqueous K+, Na+, NH4+, Ca2+, Cl– and H3O+ systems while conserving energy over long NVE runs. The correction force costs about 1.8% of the time per step, and a model trained on K+ transfers to Na+ and NH4+ without retraining.

The Allen Institute for Brain Science and AWS have released Bloom, a single-cell multimodal model that encodes gene expression and chromatin accessibility into one shared cell representation while training on roughly 100-fold fewer cells than comparable pretrained models. On adult mouse whole-brain data, linear classifiers on Bloom embeddings reach 91% cross-validated adjusted balanced accuracy for 315 subclasses (multiome), 94% for 310 subclasses (scRNA-seq) and 94% for 297 subclasses (snRNA-seq), although the model saw only mouse multiome data.

The Allen Institute for Brain Science and AWS have released Bloom, a single-cell multimodal model that encodes gene expression and chromatin accessibility into one shared cell representation while training on roughly 100-fold fewer cells than comparable pretrained models. On adult mouse whole-brain data, linear classifiers on Bloom embeddings reach 91% cross-validated adjusted balanced accuracy for 315 subclasses (multiome), 94% for 310 subclasses (scRNA-seq) and 94% for 297 subclasses (snRNA-seq), although the model saw only mouse multiome data.

Finite-element surrogates are usually trained and compared by displacement error, although engineering decisions rest on stresses. This study replaces the norm of the supervised loss with the energy norm of the assembled stiffness matrix, leaving everything else fixed; in pre-registered comparisons a three-dimensional transformer's median von Mises error fell 6.1-fold and the two-dimensional mean fell by more than half, also on held-out instances, and training on the energy alone without solutions came close to that accuracy.

Finite-element surrogates are usually trained and compared by displacement error, although engineering decisions rest on stresses. This study replaces the norm of the supervised loss with the energy norm of the assembled stiffness matrix, leaving everything else fixed; in pre-registered comparisons a three-dimensional transformer's median von Mises error fell 6.1-fold and the two-dimensional mean fell by more than half, also on held-out instances, and training on the energy alone without solutions came close to that accuracy.

Universal DFT Hamiltonian models quickly map atomic structure to electronic properties, but their predictions carry no reliability information. EviDTB attaches evidential deep regression to the SLEM backbone in DeePTB, learns only the residual between the converged Hamiltonian and a reference Hamiltonian built from the initial density, and returns a per-atom uncertainty score in a single forward pass. On 5,478 held-out structures, removing the 15% most uncertain structures lowers the 99th percentile of the mean local squared prediction error from 16.63 to 3.72 eV²; on 636 generated structures evaluated without retraining, the score has a Spearman correlation of 0.743 with that error.

Universal DFT Hamiltonian models quickly map atomic structure to electronic properties, but their predictions carry no reliability information. EviDTB attaches evidential deep regression to the SLEM backbone in DeePTB, learns only the residual between the converged Hamiltonian and a reference Hamiltonian built from the initial density, and returns a per-atom uncertainty score in a single forward pass. On 5,478 held-out structures, removing the 15% most uncertain structures lowers the 99th percentile of the mean local squared prediction error from 16.63 to 3.72 eV²; on 636 generated structures evaluated without retraining, the score has a Spearman correlation of 0.743 with that error.

Adversarial perturbations added to photos can stop a diffusion model from learning a person's identity, but earlier methods altered the whole face densely, and the accumulated distortion became visible around facial features. SEAL confines the perturbation to 5% of spatial locations, jointly optimizing where it lands and its values with a differentiable top-k mask, and disrupts the latent variance of a frozen VAE encoder. On CelebA-HQ and VGGFace2, protected images reach above 41 dB PSNR, 0.97 SSIM and at most 0.07 LPIPS, while identity similarity after DreamBooth personalization stays at most 0.15.

Adversarial perturbations added to photos can stop a diffusion model from learning a person's identity, but earlier methods altered the whole face densely, and the accumulated distortion became visible around facial features. SEAL confines the perturbation to 5% of spatial locations, jointly optimizing where it lands and its values with a differentiable top-k mask, and disrupts the latent variance of a frozen VAE encoder. On CelebA-HQ and VGGFace2, protected images reach above 41 dB PSNR, 0.97 SSIM and at most 0.07 LPIPS, while identity similarity after DreamBooth personalization stays at most 0.15.

Mechanistic interpretability often rests safety claims on thresholded numbers that can flip under small changes to weights, training seed or threshold. Working on toy models and synthetic dictionaries with known ground-truth features, this work characterizes exactly where feature dimensionality is discontinuous and replaces it with a ridge-regularized version whose stability constant grows with feature scale and whose coefficient is optimal. On 12 trained toy models, the regularized sum rule gave the same verdict at all 21 pre-set values of its smoothing parameter, while the version that ignores short feature vectors gave the opposite verdict for two of the models.

Mechanistic interpretability often rests safety claims on thresholded numbers that can flip under small changes to weights, training seed or threshold. Working on toy models and synthetic dictionaries with known ground-truth features, this work characterizes exactly where feature dimensionality is discontinuous and replaces it with a ridge-regularized version whose stability constant grows with feature scale and whose coefficient is optimal. On 12 trained toy models, the regularized sum rule gave the same verdict at all 21 pre-set values of its smoothing parameter, while the version that ignores short feature vectors gave the opposite verdict for two of the models.

Constrained vector search with per-category caps previously relied on a fixed visit budget or on stopping rules trained for average recall, neither of which guarantees an exact answer. CaST restricts stopping rules to a monotone class over a declared interface, calibrates them with fixed-sequence binomial tests, and derives a lower bound on the class-optimal cost via Lagrangian weak duality. On two datasets of one million items each, its replayed cost is at most 1.132 times the class optimum for risk levels 0.03 to 0.10 at 95% confidence, and at a 0.05 risk target its mean latency is lower than every other calibrated rule timed.

Constrained vector search with per-category caps previously relied on a fixed visit budget or on stopping rules trained for average recall, neither of which guarantees an exact answer. CaST restricts stopping rules to a monotone class over a declared interface, calibrates them with fixed-sequence binomial tests, and derives a lower bound on the class-optimal cost via Lagrangian weak duality. On two datasets of one million items each, its replayed cost is at most 1.132 times the class optimum for risk levels 0.03 to 0.10 at 95% confidence, and at a 0.05 risk target its mean latency is lower than every other calibrated rule timed.

Multi-turn agentic reinforcement learning is widely used to strengthen the long-horizon behavior of large language model agents, yet trajectory-level success rates hide what it does to individual decisions. Across four interactive tasks, DeepCoder-MT, BFCL, InterCode-SQL and FrozenLake, models trained with GRPO are evaluated along two axes, sampling breadth and interaction depth: with more sampled trajectories the RL model consistently beats its base model, but when the turns allowed per trajectory are tightened that edge shrinks, disappears, and on some tasks reverses. Matched-state comparisons show better next-step actions on only 46.8% of DeepCoder-MT states and 42.8% of FrozenLake states, with the gain concentrated in error recovery and later trajectory stages.

Multi-turn agentic reinforcement learning is widely used to strengthen the long-horizon behavior of large language model agents, yet trajectory-level success rates hide what it does to individual decisions. Across four interactive tasks, DeepCoder-MT, BFCL, InterCode-SQL and FrozenLake, models trained with GRPO are evaluated along two axes, sampling breadth and interaction depth: with more sampled trajectories the RL model consistently beats its base model, but when the turns allowed per trajectory are tightened that edge shrinks, disappears, and on some tasks reverses. Matched-state comparisons show better next-step actions on only 46.8% of DeepCoder-MT states and 42.8% of FrozenLake states, with the gain concentrated in error recovery and later trajectory stages.

Language-model failures often begin at a specific token, yet the usual response-level AUROC asks only whether a finished response failed, not when an alarm fired. Prewarn fixes each monitor's threshold on validation data at an error rate chosen for the application, then scores classification, detection and early warning separately against an annotated onset. On repetitive degeneration, a length-only rule reaches near-perfect AUROC but can never warn; frozen probes warn before the onset on most degenerate responses of Apertus and Llama, while on Mistral they mostly fire after it.

Language-model failures often begin at a specific token, yet the usual response-level AUROC asks only whether a finished response failed, not when an alarm fired. Prewarn fixes each monitor's threshold on validation data at an error rate chosen for the application, then scores classification, detection and early warning separately against an annotated onset. On repetitive degeneration, a length-only rule reaches near-perfect AUROC but can never warn; frozen probes warn before the onset on most degenerate responses of Apertus and Llama, while on Mistral they mostly fire after it.

Blockwise distillation architecture search picks child networks by summing each replacement block's isolated score, an additive assumption rarely tested directly. On DeiT-S and DeiT-B, as the block-latency budget drops from 60% to 55%, the additive integer linear program begins to trail a pairwise integer quadratic program that models block interactions, by 3.3–16.6 top-1 points on ImageNet-100 after fine-tuning. The error splits into a metric part from the score's nonlinearity and a functional interaction part, dominant on DeiT-S and DeiT-B respectively; a hybrid search that ranks with single-block logits and measures only a few candidates needs no dedicated pair measurements and beats the pairwise search in 13 of 14 full ImageNet-1k settings.

Blockwise distillation architecture search picks child networks by summing each replacement block's isolated score, an additive assumption rarely tested directly. On DeiT-S and DeiT-B, as the block-latency budget drops from 60% to 55%, the additive integer linear program begins to trail a pairwise integer quadratic program that models block interactions, by 3.3–16.6 top-1 points on ImageNet-100 after fine-tuning. The error splits into a metric part from the score's nonlinearity and a functional interaction part, dominant on DeiT-S and DeiT-B respectively; a hybrid search that ranks with single-block logits and measures only a few candidates needs no dedicated pair measurements and beats the pairwise search in 13 of 14 full ImageNet-1k settings.

For machine unlearning under privacy rules, IDSISA adds influence-guided shard assignment and cross-shard distillation to SISA's shard-retraining framework. Influence scores are computed once before training and co-locate samples likely to be deleted together, so a deletion request touches fewer shards than under random partitioning; retrained shards also learn from unaffected shards' soft outputs to recover utility lost to isolation. On ResNet-18/CIFAR-10, BiLSTM/IMDB, TCN/IMDB and a fully connected Fashion-MNIST network, the deletion guarantee still comes from retraining itself, and distillation is only a separately quantified utility-recovery step.

For machine unlearning under privacy rules, IDSISA adds influence-guided shard assignment and cross-shard distillation to SISA's shard-retraining framework. Influence scores are computed once before training and co-locate samples likely to be deleted together, so a deletion request touches fewer shards than under random partitioning; retrained shards also learn from unaffected shards' soft outputs to recover utility lost to isolation. On ResNet-18/CIFAR-10, BiLSTM/IMDB, TCN/IMDB and a fully connected Fashion-MNIST network, the deletion guarantee still comes from retraining itself, and distillation is only a separately quantified utility-recovery step.

Speech tokenizers compress continuous speech into discrete tokens and serve as the front end of ASR systems and large audio-language models. T-SemAttack skips the full recognition pipeline and instead jointly drifts the representation spaces of several semantic encoders, with a psychoacoustic masking constraint keeping the noise acceptable to listeners while pulling downstream semantic encoding off course. On bilingual multi-speaker data it raises word error rates on nine open-source and black-box ASR systems from about 1–4% on clean speech to 20.004–63.555%, and it introduces ROSETok, a tokenizer trained with noisy training and semantic distillation.

Speech tokenizers compress continuous speech into discrete tokens and serve as the front end of ASR systems and large audio-language models. T-SemAttack skips the full recognition pipeline and instead jointly drifts the representation spaces of several semantic encoders, with a psychoacoustic masking constraint keeping the noise acceptable to listeners while pulling downstream semantic encoding off course. On bilingual multi-speaker data it raises word error rates on nine open-source and black-box ASR systems from about 1–4% on clean speech to 20.004–63.555%, and it introduces ROSETok, a tokenizer trained with noisy training and semantic distillation.

Humanoid soccer robots have either covered only a few skills or skipped motion priors entirely and moved unnaturally. SoccerBot distills large-scale human soccer motion-capture data into a structured latent motion prior, letting a single skill-conditioned policy learn moving, dribbling, trapping and kicking in one latent space. On the Unitree G1, the policy transfers zero-shot to hardware, with simulation success rates of 97.04% for kicking and 89.88% for dribbling, while real-robot kicking direction error rises from 4.09 to 10.41 degrees.

Humanoid soccer robots have either covered only a few skills or skipped motion priors entirely and moved unnaturally. SoccerBot distills large-scale human soccer motion-capture data into a structured latent motion prior, letting a single skill-conditioned policy learn moving, dribbling, trapping and kicking in one latent space. On the Unitree G1, the policy transfers zero-shot to hardware, with simulation success rates of 97.04% for kicking and 89.88% for dribbling, while real-robot kicking direction error rises from 4.09 to 10.41 degrees.

Federated learning usually requires institutions to share one model backbone so parameter updates can be aggregated, which keeps centers running different model families or sizes from collaborating. FedVP moves collaboration from parameter aggregation to shared verification: each client trains a verifier on its local data and candidate responses from its own frozen LLM, and the server pools these independently trained verifiers to score and select answers, never merging backbone parameters. Across five medical question-answering datasets, with centers using Qwen and LLaMA backbones from 3B to 14B parameters, one verifier pool is reused for all five backbones without retraining and raises mean accuracy by 1.46 to 5.79 percentage points over the strongest conventional federated baselines.

Federated learning usually requires institutions to share one model backbone so parameter updates can be aggregated, which keeps centers running different model families or sizes from collaborating. FedVP moves collaboration from parameter aggregation to shared verification: each client trains a verifier on its local data and candidate responses from its own frozen LLM, and the server pools these independently trained verifiers to score and select answers, never merging backbone parameters. Across five medical question-answering datasets, with centers using Qwen and LLaMA backbones from 3B to 14B parameters, one verifier pool is reused for all five backbones without retraining and raises mean accuracy by 1.46 to 5.79 percentage points over the strongest conventional federated baselines.

ARC-AGI-3 asks agents to clear uninstructed interactive games with as few actions as possible, and handing a reasoning model the board as a raw integer matrix makes it collide with walls repeatedly. Sarbloh computes geometric facts such as reachability, enclosure, midpoint tethers and dihedral symmetry in host code with no learned parameters, keeps role labels as guesses each carrying a one-action test, and lets the agent commit multi-step paths from inside its Python kernel via await play(). On the official hidden set its score rose from an initial baseline of 0.27 to 16.53; the source describes that gain as a 13.7 efficiency improvement over the first Milestone winner, while the same table gives that winner a score of 1.21.

ARC-AGI-3 asks agents to clear uninstructed interactive games with as few actions as possible, and handing a reasoning model the board as a raw integer matrix makes it collide with walls repeatedly. Sarbloh computes geometric facts such as reachability, enclosure, midpoint tethers and dihedral symmetry in host code with no learned parameters, keeps role labels as guesses each carrying a one-action test, and lets the agent commit multi-step paths from inside its Python kernel via await play(). On the official hidden set its score rose from an initial baseline of 0.27 to 16.53; the source describes that gain as a 13.7 efficiency improvement over the first Milestone winner, while the same table gives that winner a score of 1.21.

DeepReach replaces the grid in Hamilton–Jacobi reachability with a neural value function but penalizes both sides of the residual equally, so the learned value can sit above the true one and unsafe states get labeled safe. ConDR instead trains toward a subsolution of the HJI variational inequality: the obstacle branch and terminal condition hold by construction, a one-sided loss pushes the differential residual below a safety margin, and a complementarity term blocks the degenerate all-unsafe solution. On Air3D, Air6D, a 9D multi-vehicle system and a 13D quadrotor, ConDR reports the lowest false-negative rate and the lowest closed-loop collision rates, with at most a small rise in intervention.

DeepReach replaces the grid in Hamilton–Jacobi reachability with a neural value function but penalizes both sides of the residual equally, so the learned value can sit above the true one and unsafe states get labeled safe. ConDR instead trains toward a subsolution of the HJI variational inequality: the obstacle branch and terminal condition hold by construction, a one-sided loss pushes the differential residual below a safety margin, and a complementarity term blocks the degenerate all-unsafe solution. On Air3D, Air6D, a 9D multi-vehicle system and a 13D quadrotor, ConDR reports the lowest false-negative rate and the lowest closed-loop collision rates, with at most a small rise in intervention.

Long-running LLM agents often write each episode's lesson straight into memory, yet local success is not a reliable sign of reusable, safe knowledge. FGSE turns every candidate lesson into a structured hypothesis with a precondition, behavior change, expected effect and verifier, applies it only in a temporary state, and tests it on target-transfer, falsification and archived-regression probes before committing, narrowing or rejecting it. On verifier-rich benchmarks including WebArena, AppWorld and τ-bench, FGSE with a DeepSeek-V3.1 backbone reaches 39.8 WebArena success rate and 78.8 AppWorld average completion, while cutting falsification failures, regression damage and harmful committed updates to 5.0%, 1.3% and 3.6%.

Long-running LLM agents often write each episode's lesson straight into memory, yet local success is not a reliable sign of reusable, safe knowledge. FGSE turns every candidate lesson into a structured hypothesis with a precondition, behavior change, expected effect and verifier, applies it only in a temporary state, and tests it on target-transfer, falsification and archived-regression probes before committing, narrowing or rejecting it. On verifier-rich benchmarks including WebArena, AppWorld and τ-bench, FGSE with a DeepSeek-V3.1 backbone reaches 39.8 WebArena success rate and 78.8 AppWorld average completion, while cutting falsification failures, regression damage and harmful committed updates to 5.0%, 1.3% and 3.6%.

Constitutive model discovery now asks which environmental variable drives which mechanism, not merely which candidate terms fit the response. A constitutive model with internal variables is written as a time-unrolled structural causal model, candidate graphs are scored by a weak-form equilibrium gap through the boundary value problem, and climate-chamber protocols count as interventions. On a synthetic perforated plate at about 0.01 px displacement noise, interventional data recover the true graph exactly in 7 of 10 runs, with no attribution error; correlated observational data never do.

Constitutive model discovery now asks which environmental variable drives which mechanism, not merely which candidate terms fit the response. A constitutive model with internal variables is written as a time-unrolled structural causal model, candidate graphs are scored by a weak-form equilibrium gap through the boundary value problem, and climate-chamber protocols count as interventions. On a synthetic perforated plate at about 0.01 px displacement noise, interventional data recover the true graph exactly in 7 of 10 runs, with no attribution error; correlated observational data never do.

Multi-view 3D vision-language models suffer viewpoint-induced semantic drift: the same object and query produce weaker, diffuse or shifted target responses when the camera moves. LieVLM has geometry tokens predict local rotation parameters and channel-wise gates, applies length-preserving rotations to projected visual features within learned three-dimensional subspaces, and gates the identity-relative correction back into the original representation. On ScanRefer, LieVLM-8B reaches 64.5 Acc@0.25 and 57.2 Acc@0.5, 6.9 and 6.3 points above the published VG LLM-8B under the same proposal-refinement protocol.

Multi-view 3D vision-language models suffer viewpoint-induced semantic drift: the same object and query produce weaker, diffuse or shifted target responses when the camera moves. LieVLM has geometry tokens predict local rotation parameters and channel-wise gates, applies length-preserving rotations to projected visual features within learned three-dimensional subspaces, and gates the identity-relative correction back into the original representation. On ScanRefer, LieVLM-8B reaches 64.5 Acc@0.25 and 57.2 Acc@0.5, 6.9 and 6.3 points above the published VG LLM-8B under the same proposal-refinement protocol.

Curriculum learning often helps more in reinforcement learning than in supervised learning, and a tractable target-interception model now explains why. With dense rewards, curricula bring only marginal gains; with sparse rewards, a curriculum can be required for learning to occur at all. When the agent must both minimize effort and accumulate reward, competing optima appear and can trap it in a quitting solution. Useful curricula do two things: they make rewards less sparse to speed learning, and they shape the reward landscape to steer agents away from bad local optima.

Curriculum learning often helps more in reinforcement learning than in supervised learning, and a tractable target-interception model now explains why. With dense rewards, curricula bring only marginal gains; with sparse rewards, a curriculum can be required for learning to occur at all. When the agent must both minimize effort and accumulate reward, competing optima appear and can trap it in a quitting solution. Useful curricula do two things: they make rewards less sparse to speed learning, and they shape the reward landscape to steer agents away from bad local optima.

In multi-agent systems, confidence often decides who receives scarce downstream computation and decision weight, which gives agents an incentive to overstate it. The Audited Pivot Scoring Mechanism (APSM) offsets platform-controlled allocation benefits in the transfer, verifies outcomes through inverse-propensity-weighted sparse audits, charges the Clarke pivot externality, and adds a strictly proper score, making truthful reporting the unique expected-utility best response. Calculator-free experiments show agents inflate confidence under Top-K incentives; with an agent-local utility interface, APSM yields lower Brier loss and fewer cutoff crossings than Compensated Proper Scoring (CPS), while CPS achieves a higher exact allocation match.

In multi-agent systems, confidence often decides who receives scarce downstream computation and decision weight, which gives agents an incentive to overstate it. The Audited Pivot Scoring Mechanism (APSM) offsets platform-controlled allocation benefits in the transfer, verifies outcomes through inverse-propensity-weighted sparse audits, charges the Clarke pivot externality, and adds a strictly proper score, making truthful reporting the unique expected-utility best response. Calculator-free experiments show agents inflate confidence under Top-K incentives; with an agent-local utility interface, APSM yields lower Brier loss and fewer cutoff crossings than Compensated Proper Scoring (CPS), while CPS achieves a higher exact allocation match.

On the corrected CIC-IDS2017 dataset, when labels and confidences from four heterogeneous detectors are corrupted under a threat model, unverified open-weight LLM investigators fall to 0.50–0.72 accuracy with one corrupted detector and to 0.13 or less with three. Conflict-Aware Evidence Verification (CAV) keeps reported confidences out of the vote and, only when tools conflict, tests the leading candidates with a nearest-neighbour trust score computed from raw traffic features. With its prompt frozen before testing, CAV-LLM reaches 0.84–0.88 under three corrupted detectors in both persistent and sporadic scopes, against at most 0.08 for the naive investigator, while calling the LLM on 0.4% of clean traffic.

On the corrected CIC-IDS2017 dataset, when labels and confidences from four heterogeneous detectors are corrupted under a threat model, unverified open-weight LLM investigators fall to 0.50–0.72 accuracy with one corrupted detector and to 0.13 or less with three. Conflict-Aware Evidence Verification (CAV) keeps reported confidences out of the vote and, only when tools conflict, tests the leading candidates with a nearest-neighbour trust score computed from raw traffic features. With its prompt frozen before testing, CAV-LLM reaches 0.84–0.88 under three corrupted detectors in both persistent and sporadic scopes, against at most 0.08 for the naive investigator, while calling the LLM on 0.4% of clean traffic.

Visual verifiers have long returned only a true-or-false verdict, which cannot separate partially correct outputs from worse ones. OmniVerifier-C1 decomposes a complex prompt or environmental rule into fine-grained sub-questions, judges each one binarily, and turns the satisfied fraction into a continuous score, while a decoupled reinforcement-learning scheme optimizes sub-question judgments and global alignment scoring together. On ViVerBench, the Qwen3-VL-8B version trained with graded supervised fine-tuning plus graded reinforcement learning reaches an overall score of 0.684, above 0.671 for binary reinforcement learning alone.

Visual verifiers have long returned only a true-or-false verdict, which cannot separate partially correct outputs from worse ones. OmniVerifier-C1 decomposes a complex prompt or environmental rule into fine-grained sub-questions, judges each one binarily, and turns the satisfied fraction into a continuous score, while a decoupled reinforcement-learning scheme optimizes sub-question judgments and global alignment scoring together. On ViVerBench, the Qwen3-VL-8B version trained with graded supervised fine-tuning plus graded reinforcement learning reaches an overall score of 0.684, above 0.671 for binary reinforcement learning alone.

PaperBenchX extends paper reproduction beyond machine learning to 12 research areas, requiring agents to rebuild key experiments inside native scientific software such as HFSS, Lumerical, Meep and PySCF, then replaying each submitted workflow from scratch and scoring only regenerated evidence. Across 93 tasks, ten frontier agent configurations score a mean rubric value between 27.53% and 59.08%, and the strongest fully reproduces only 14% of tasks at the 90% threshold, with modeling and execution scores well above validation.

PaperBenchX extends paper reproduction beyond machine learning to 12 research areas, requiring agents to rebuild key experiments inside native scientific software such as HFSS, Lumerical, Meep and PySCF, then replaying each submitted workflow from scratch and scoring only regenerated evidence. Across 93 tasks, ten frontier agent configurations score a mean rubric value between 27.53% and 59.08%, and the strongest fully reproduces only 14% of tasks at the 90% threshold, with modeling and execution scores well above validation.

ConwayNet adds to each neural-network layer a fixed, Game-of-Life-inspired lateral-inhibition rule whose parameters are never optimised: it reads the layer input's neighbourhood occupancy and blends birth and survival gates into the output according to each unit's own state. Wired into fully connected, convolutional, wide-residual and graph convolutional networks, it makes activation maps coherent across space, depth and seeds and pushes gradient saliency onto the object; on the shallow convolutional network, accuracy under Gaussian, shot and impulse noise rises by double digits, at a clean-accuracy cost of about one point on MNIST, EMNIST and their colour versions and a few points on FashionMNIST and KMNIST.

ConwayNet adds to each neural-network layer a fixed, Game-of-Life-inspired lateral-inhibition rule whose parameters are never optimised: it reads the layer input's neighbourhood occupancy and blends birth and survival gates into the output according to each unit's own state. Wired into fully connected, convolutional, wide-residual and graph convolutional networks, it makes activation maps coherent across space, depth and seeds and pushes gradient saliency onto the object; on the shallow convolutional network, accuracy under Gaussian, shot and impulse noise rises by double digits, at a clean-accuracy cost of about one point on MNIST, EMNIST and their colour versions and a few points on FashionMNIST and KMNIST.

Imbue lets an already-trained Bayesian experimental design policy take in external knowledge available only at deployment without retraining: user-specified priors and observations from previous instances of the experiment are encoded as extra input tokens, processed in-context with the accumulated observations by a shared Transformer, and a reliability filter scores each token against those observations and excludes inconsistent ones. Across four benchmarks it accelerates early-stage information acquisition when the external knowledge is reliable and stays close to the no-external baseline when it is not.

Imbue lets an already-trained Bayesian experimental design policy take in external knowledge available only at deployment without retraining: user-specified priors and observations from previous instances of the experiment are encoded as extra input tokens, processed in-context with the accumulated observations by a shared Transformer, and a reliability filter scores each token against those observations and excludes inconsistent ones. Across four benchmarks it accelerates early-stage information acquisition when the external knowledge is reliable and stays close to the no-external baseline when it is not.

Approaching the complete basis set (CBS) limit in quantum chemistry usually means several mid-quality calculations plus extrapolation. Here a single Hartree–Fock calculation in the minimal pcseg-0 basis is augmented with an optimized, minimal complementary auxiliary basis set (CABS) singles correction, and kernel ridge regression learns the residual. On GDB-BSIE, CABSmin/pcseg-0 with atomic detrending gives an RMSE of 7.03 mHa against RHF/cc-pV5Z, at about 16% of the cost of a cc-pVDZ single point.

Approaching the complete basis set (CBS) limit in quantum chemistry usually means several mid-quality calculations plus extrapolation. Here a single Hartree–Fock calculation in the minimal pcseg-0 basis is augmented with an optimized, minimal complementary auxiliary basis set (CABS) singles correction, and kernel ridge regression learns the residual. On GDB-BSIE, CABSmin/pcseg-0 with atomic detrending gives an RMSE of 7.03 mHa against RHF/cc-pV5Z, at about 16% of the cost of a cc-pVDZ single point.

Lower bounds on mutually orthogonal Latin squares (MOLS) have long lagged far behind upper bounds at orders that are not prime powers. Using established techniques including difference matrices, quasi-difference matrices, orthogonal-array lifting and projective-plane point selection, OpenAI's GPT-6 Astra constructed larger MOLS sets at 33 orders between 22 and 494, raising their lower bounds; at order 22 it produced 4 MOLS for the first time, so combined with existing results every order from n=22 upward is known to admit at least 4.

Lower bounds on mutually orthogonal Latin squares (MOLS) have long lagged far behind upper bounds at orders that are not prime powers. Using established techniques including difference matrices, quasi-difference matrices, orthogonal-array lifting and projective-plane point selection, OpenAI's GPT-6 Astra constructed larger MOLS sets at 33 orders between 22 and 494, raising their lower bounds; at order 22 it produced 4 MOLS for the first time, so combined with existing results every order from n=22 upward is known to admit at least 4.

Causal mediation analysis splits a treatment effect into an indirect pathway through mediators and a direct pathway through other processes, and attaching a confidence interval to the indirect effect usually means refitting several nuisance functions on every table and resampling repeatedly. MediPFN adapts a pretrained tabular foundation model to predict those nuisance functions, then a separate head learns quantiles of the estimator's standardized error from simulated tables with known indirect effects, with symmetrization and independent calibration producing the interval. On synthetic, distribution-shift and real-covariate tables its coverage sits closer to the 95% nominal level than several comparators, and paired timing gives a median 59.1-fold speedup over bootstrap.

Causal mediation analysis splits a treatment effect into an indirect pathway through mediators and a direct pathway through other processes, and attaching a confidence interval to the indirect effect usually means refitting several nuisance functions on every table and resampling repeatedly. MediPFN adapts a pretrained tabular foundation model to predict those nuisance functions, then a separate head learns quantiles of the estimator's standardized error from simulated tables with known indirect effects, with symmetrization and independent calibration producing the interval. On synthetic, distribution-shift and real-covariate tables its coverage sits closer to the 95% nominal level than several comparators, and paired timing gives a median 59.1-fold speedup over bootstrap.

Across several large Perturb-seq datasets, a generative framework called Graph-guided Flow Matching (GFM) combines three biological knowledge graphs, covering Gene Ontology, protein interactions and perturbation similarity, with latent flow matching to predict single-cell transcriptomes after unseen perturbations. It outperforms a linear baseline and the deep-learning models CellFlow, GEARS and scGPT on unseen single-gene perturbations, trains up to 30 times faster, and keeps its advantage as training data shrink; on two-gene perturbations it does better when neither gene was seen in training.

Across several large Perturb-seq datasets, a generative framework called Graph-guided Flow Matching (GFM) combines three biological knowledge graphs, covering Gene Ontology, protein interactions and perturbation similarity, with latent flow matching to predict single-cell transcriptomes after unseen perturbations. It outperforms a linear baseline and the deep-learning models CellFlow, GEARS and scGPT on unseen single-gene perturbations, trains up to 30 times faster, and keeps its advantage as training data shrink; on two-gene perturbations it does better when neither gene was seen in training.

Google introduced Gemini 4 Hydrogen, a fast execution model for everyday coding and agentic workflows that complements the long-horizon reasoning models Gemini 4 Carbon and Argon. At its default high thinking level it scores 51 on the Artificial Analysis Intelligence Index versus Argon's 50, reaching 54 at the new xhigh setting; teams migrating from Gemini 3.8 Flash finish tasks 2.5 times faster, and Terminal-Bench 4.1 rises from 18.2% to 57%. Pricing is flat across its 1 million-token context at $1.50 per million input tokens and $7.50 per million output tokens.

Google introduced Gemini 4 Hydrogen, a fast execution model for everyday coding and agentic workflows that complements the long-horizon reasoning models Gemini 4 Carbon and Argon. At its default high thinking level it scores 51 on the Artificial Analysis Intelligence Index versus Argon's 50, reaching 54 at the new xhigh setting; teams migrating from Gemini 3.8 Flash finish tasks 2.5 times faster, and Terminal-Bench 4.1 rises from 18.2% to 57%. Pricing is flat across its 1 million-token context at $1.50 per million input tokens and $7.50 per million output tokens.

The structure-based protein language model ESM3 treats proteins as if in vacuum, and without solvent information its per-residue outputs mislabel hydrogen bonds as destabilizing, with error that varies with solvent-accessible surface area. HEIWA adds water–residue attention to ESM3 residue embeddings, passing solvent-mediated force-field information through learnable water key vectors and a rotational angle term. On a thermophile classification proxy task, the HEIWA model's classification metrics stay broadly comparable to the baseline, while per-residue outputs correlate better with an RMSF benchmark and SASA; temperature-ramping molecular dynamics found 23 of 38 tested proteins more stable.

The structure-based protein language model ESM3 treats proteins as if in vacuum, and without solvent information its per-residue outputs mislabel hydrogen bonds as destabilizing, with error that varies with solvent-accessible surface area. HEIWA adds water–residue attention to ESM3 residue embeddings, passing solvent-mediated force-field information through learnable water key vectors and a rotational angle term. On a thermophile classification proxy task, the HEIWA model's classification metrics stay broadly comparable to the baseline, while per-residue outputs correlate better with an RMSF benchmark and SASA; temperature-ramping molecular dynamics found 23 of 38 tested proteins more stable.

Working with write access to only a subset of deployed sensor channels and no knowledge of grid topology or detector parameters, PCGAN learns approximate active-power–angle and reactive-power–voltage-magnitude relations from historical measurements and uses adversarial training to generate data-injection perturbations that resemble normal operation. On the IEEE 300-bus system with one year of PJM hourly demand and 10% to 100% compromised sensors, it reaches 84.28% mean evasion of ten machine-learning detectors and 22.31% mean spatial load misperception, for an average stealth-impact score of 18.69% against 5.61% for the strongest competing attack.

Working with write access to only a subset of deployed sensor channels and no knowledge of grid topology or detector parameters, PCGAN learns approximate active-power–angle and reactive-power–voltage-magnitude relations from historical measurements and uses adversarial training to generate data-injection perturbations that resemble normal operation. On the IEEE 300-bus system with one year of PJM hourly demand and 10% to 100% compromised sensors, it reaches 84.28% mean evasion of ten machine-learning detectors and 22.31% mean spatial load misperception, for an average stealth-impact score of 18.69% against 5.61% for the strongest competing attack.

A magnetic neuroevolution potential takes magnetic moments as inputs alongside atomic positions, so different magnetic orders of the same atomic structure no longer share one energy surface. Trained on the same datasets for the same number of iterations as a standard neuroevolution potential without magnetic descriptors, it cuts energy root-mean-square error by up to about one order of magnitude on CrN, Fe and FeGe. Spin-lattice simulations of Fe give a magnetic transition temperature near 1000 K, close to the experimental Curie temperature of 1043 K, and one consumer-grade GPU handles coupled dynamics of more than two million atoms.

A magnetic neuroevolution potential takes magnetic moments as inputs alongside atomic positions, so different magnetic orders of the same atomic structure no longer share one energy surface. Trained on the same datasets for the same number of iterations as a standard neuroevolution potential without magnetic descriptors, it cuts energy root-mean-square error by up to about one order of magnitude on CrN, Fe and FeGe. Spin-lattice simulations of Fe give a magnetic transition temperature near 1000 K, close to the experimental Curie temperature of 1043 K, and one consumer-grade GPU handles coupled dynamics of more than two million atoms.

Negative-weight events from MC@NLO matching erode the statistical power of simulations, and full folding suppresses them only at prohibitive cost. This work learns a local correction with decision trees and normalizing flows whose integral over radiation variables vanishes, and adds it to the radiation-event short-distance cross section; across several electron- and hadron-collision processes the negative-weight fraction falls in most cases while total rates and Born-level distributions stay unchanged, at a modest extra computational cost.

Negative-weight events from MC@NLO matching erode the statistical power of simulations, and full folding suppresses them only at prohibitive cost. This work learns a local correction with decision trees and normalizing flows whose integral over radiation variables vanishes, and adds it to the radiation-event short-distance cross section; across several electron- and hadron-collision processes the negative-weight fraction falls in most cases while total rates and Born-level distributions stay unchanged, at a modest extra computational cost.

KEN embeds two-dimensional convolutional kernels in an E(3)-equivariant graph neural network so that an interatomic potential reads electronic spin-interaction matrices alongside geometric coordinates, sidestepping the ill-posedness of coordinate-only potentials in strongly correlated magnetic systems. On the HE26 actinide dataset, KENJI reaches a training bonding-energy mean absolute error of 6.93 meV/atom against 44.7 meV/atom for MACE-osaka26, and its training force error of 34.10 meV/Å is on par with 26.30 meV/Å. On the independent test set the energy error is 46.74 meV/atom and the force error exceeds MACE-osaka26, which the authors partly attribute to training without stress.

KEN embeds two-dimensional convolutional kernels in an E(3)-equivariant graph neural network so that an interatomic potential reads electronic spin-interaction matrices alongside geometric coordinates, sidestepping the ill-posedness of coordinate-only potentials in strongly correlated magnetic systems. On the HE26 actinide dataset, KENJI reaches a training bonding-energy mean absolute error of 6.93 meV/atom against 44.7 meV/atom for MACE-osaka26, and its training force error of 34.10 meV/Å is on par with 26.30 meV/Å. On the independent test set the energy error is 46.74 meV/atom and the force error exceeds MACE-osaka26, which the authors partly attribute to training without stress.

Finger-movement labels in non-human primates have long come from instrumented manipulanda fitted with bend sensors, which cap the dimensionality of measurable movement. Freebie uses several synchronized cameras and real-time pose estimation to output finger keypoints directly, reaching median end-to-end latencies of 15.8 ms at 30 FPS and 16.2 ms at 60 FPS, with a median absolute correlation of 0.96 against bend-sensor signals. EMG decoders trained on Freebie labels matched those trained on bend-sensor labels in closed-loop control, indicating that training a decoder need not require instrumenting the hand.

Finger-movement labels in non-human primates have long come from instrumented manipulanda fitted with bend sensors, which cap the dimensionality of measurable movement. Freebie uses several synchronized cameras and real-time pose estimation to output finger keypoints directly, reaching median end-to-end latencies of 15.8 ms at 30 FPS and 16.2 ms at 60 FPS, with a median absolute correlation of 0.96 against bend-sensor signals. EMG decoders trained on Freebie labels matched those trained on bend-sensor labels in closed-loop control, indicating that training a decoder need not require instrumenting the hand.

TabOmics is an open tabular foundation model with 7.3M parameters, pretrained only on synthetic tables, that processes all features retained after preprocessing jointly in a single forward pass instead of subsampling or compressing them. Across 46 classification tasks from three high-dimensional, low-sample-size benchmarks, the eight-member TabOmics v2.5 reaches a mean AUROC of 0.943, above 0.942 for eight-member TabPFN-3.5; the single-member version reaches 0.941 with no significant paired difference, at roughly five times the runtime per forward pass.

TabOmics is an open tabular foundation model with 7.3M parameters, pretrained only on synthetic tables, that processes all features retained after preprocessing jointly in a single forward pass instead of subsampling or compressing them. Across 46 classification tasks from three high-dimensional, low-sample-size benchmarks, the eight-member TabOmics v2.5 reaches a mean AUROC of 0.943, above 0.942 for eight-member TabPFN-3.5; the single-member version reaches 0.941 with no significant paired difference, at roughly five times the runtime per forward pass.

The functional diversity and mutational potential of beta-lactamases in environmental microbial communities were only partly understood. Combining protein language models with ensemble learning identified more than 46,000 putative beta-lactamases across all four Ambler classes from environmental metagenomes; two class C candidates sharing about 30% sequence identity with known enzymes conferred resistance to multiple beta-lactam antibiotics when expressed in E. coli, and mutations found by combining an inverse-folding model with protein language models altered resistance profiles, including increased carbapenem resistance.

The functional diversity and mutational potential of beta-lactamases in environmental microbial communities were only partly understood. Combining protein language models with ensemble learning identified more than 46,000 putative beta-lactamases across all four Ambler classes from environmental metagenomes; two class C candidates sharing about 30% sequence identity with known enzymes conferred resistance to multiple beta-lactam antibiotics when expressed in E. coli, and mutations found by combining an inverse-folding model with protein language models altered resistance profiles, including increased carbapenem resistance.

Engineering design often has only 10 to 1,000 evaluated designs because simulations and experiments are costly. GenPFN trains no new generator: a frozen TabPFN classifier reads labeled designs, candidates are interpolated between desired designs and updated with classifier gradients and Gaussian noise, and every candidate is evaluated by the problem's own constraints and objectives. Across fourteen engineering problems, GenPFN v1 has higher class satisfaction and lower realism distance than the tested CTGAN, TVAE and TabDDPM; in the stern case, 993 of 1,000 candidates passed geometric checks and met the propeller-placement area requirement, versus 28 of 2,000 random context designs.

Engineering design often has only 10 to 1,000 evaluated designs because simulations and experiments are costly. GenPFN trains no new generator: a frozen TabPFN classifier reads labeled designs, candidates are interpolated between desired designs and updated with classifier gradients and Gaussian noise, and every candidate is evaluated by the problem's own constraints and objectives. Across fourteen engineering problems, GenPFN v1 has higher class satisfaction and lower realism distance than the tested CTGAN, TVAE and TabDDPM; in the stern case, 993 of 1,000 candidates passed geometric checks and met the propeller-placement area requirement, versus 28 of 2,000 random context designs.

Hadronic jet energy calibration at the Large Hadron Collider has relied on fitting parametric functional forms to selected observables, correcting pile-up, detector response and data–simulation differences in sequence. A prototype pipeline instead trains a neural network for each stage directly against the quantity that stage removes, imposes no functional form, and extends by adding inputs. On simulated dijet and large-radius jet samples it matches or exceeds the parametric baseline at nearly every stage, with residual pile-up dependence averaging about a tenth of the baseline's.

Hadronic jet energy calibration at the Large Hadron Collider has relied on fitting parametric functional forms to selected observables, correcting pile-up, detector response and data–simulation differences in sequence. A prototype pipeline instead trains a neural network for each stage directly against the quantity that stage removes, imposes no functional form, and extends by adding inputs. On simulated dijet and large-radius jet samples it matches or exceeds the parametric baseline at nearly every stage, with residual pile-up dependence averaging about a tenth of the baseline's.

Paired modalities are usually used to pull two encoders into one shared embedding space. DuoJEPA instead keeps a separate LeJEPA self-supervised space per modality and appends to each encoder one residual-free transformer layer that translates its embedding into the other encoder's space, trained by a contrastive loss computed entirely within that target space. On CheXpert Plus chest radiographs and reports it reaches a zero-shot AUC of 86.6 and a cross-modal retrieval score of 78.5, against 60.0 and 64.2 for a reimplementation of LeVLJEPA with the same encoders and data, while a paired-cosine test detects no direct alignment between the two own spaces.

Paired modalities are usually used to pull two encoders into one shared embedding space. DuoJEPA instead keeps a separate LeJEPA self-supervised space per modality and appends to each encoder one residual-free transformer layer that translates its embedding into the other encoder's space, trained by a contrastive loss computed entirely within that target space. On CheXpert Plus chest radiographs and reports it reaches a zero-shot AUC of 86.6 and a cross-modal retrieval score of 78.5, against 60.0 and 64.2 for a reimplementation of LeVLJEPA with the same encoders and data, while a paired-cosine test detects no direct alignment between the two own spaces.

For scientific workflows that must process sensor data within hard deadlines, UniPipe runs inference serving and online fine-tuning of a deep-neural-network surrogate in a single CUDA context on one edge GPU, handling inference requests and labeled samples in one forward pass and backpropagating only through the labeled subset. On a real ptychographic reconstruction workflow with PtychoNN it lowers inference and reconstruction MSE over a frozen pretrained baseline, reduces the deadline-miss rate over a two-context concurrent baseline, and nearly halves the 95th-percentile latency spread relative to the mean.

For scientific workflows that must process sensor data within hard deadlines, UniPipe runs inference serving and online fine-tuning of a deep-neural-network surrogate in a single CUDA context on one edge GPU, handling inference requests and labeled samples in one forward pass and backpropagating only through the labeled subset. On a real ptychographic reconstruction workflow with PtychoNN it lowers inference and reconstruction MSE over a frozen pretrained baseline, reduces the deadline-miss rate over a two-context concurrent baseline, and nearly halves the 95th-percentile latency spread relative to the mean.

LangChain announced a $300M Series C at a $3B valuation, led by existing investor CapitalG with new investors GIC and DFJ Growth participating. The company ties the funding to its goal of letting every company own its intelligence, anchored in LangSmith, a framework-agnostic platform for running, improving and governing agents. LangChain says more than 7,000 organizations already rely on LangSmith, and that Workday cut the time to deploy a new agent from five months to under a day.

LangChain announced a $300M Series C at a $3B valuation, led by existing investor CapitalG with new investors GIC and DFJ Growth participating. The company ties the funding to its goal of letting every company own its intelligence, anchored in LangSmith, a framework-agnostic platform for running, improving and governing agents. LangChain says more than 7,000 organizations already rely on LangSmith, and that Workday cut the time to deploy a new agent from five months to under a day.

Automated surgical skill scoring has depended on a fixed rubric, so a new set of similarly worded assessment items meant retraining. Predictive Rubric Alignment (PRA) instead predicts a rater-weighted combination of the embeddings of each level's written description, then projects that prediction onto a path joining adjacent descriptions to read out a score. On AIxSuture suturing videos, the PRA-item variant with learned item queries reached 0.848 accuracy and 0.829 macro-F1, 0.113 macro-F1 above Video Swin-T, and PRA-cond trained on only two items still gave positive score correlations on unseen JIGSAWS items.

Automated surgical skill scoring has depended on a fixed rubric, so a new set of similarly worded assessment items meant retraining. Predictive Rubric Alignment (PRA) instead predicts a rater-weighted combination of the embeddings of each level's written description, then projects that prediction onto a path joining adjacent descriptions to read out a score. On AIxSuture suturing videos, the PRA-item variant with learned item queries reached 0.848 accuracy and 0.829 macro-F1, 0.113 macro-F1 above Video Swin-T, and PRA-cond trained on only two items still gave positive score correlations on unseen JIGSAWS items.

Beamforming for near-field extreme massive MIMO integrated sensing and communication usually requires repeatedly solving nonconvex constrained optimization, and online latency grows with array size. PATBO fixes the communication and sensing beam directions geometrically and lets a Transformer predict only the normalized stream powers, trained with Cramér–Rao bound (CRB) sensitivity weighting. On a retained test set with 1024 antennas, 4 users and 2 targets, direct inference reaches 19.54% joint feasibility, physics-constrained power repair raises it to 87.79%, and optimizer fallback for the remainder brings it to 100%, at 1.11 ms per sample.

Beamforming for near-field extreme massive MIMO integrated sensing and communication usually requires repeatedly solving nonconvex constrained optimization, and online latency grows with array size. PATBO fixes the communication and sensing beam directions geometrically and lets a Transformer predict only the normalized stream powers, trained with Cramér–Rao bound (CRB) sensitivity weighting. On a retained test set with 1024 antennas, 4 users and 2 targets, direct inference reaches 19.54% joint feasibility, physics-constrained power repair raises it to 87.79%, and optimizer fallback for the remainder brings it to 100%, at 1.11 ms per sample.

A team including Caltech researchers trained a two-dimensional convolutional neural network, CDE, to invert low signal-to-noise Lyman-alpha integral-field cubes into line-of-sight integrated gas column-density maps. Training data come from IllustrisTNG100, post-processed with simplified Lyman-alpha radiative transfer and an approximate treatment of KCWI noise. On the simulated test set, CDE reaches a median linear RMSE of 1.48×10²⁰ cm^-2 and a median image SSIM of 0.635, both better than a pixel-wise power-law empirical baseline; applied to KCWI observations of the CC2.2 and Hyperion protoclusters, the predicted maps show filamentary structures extending about 40 arcseconds or more, though the results remain conditional on the adopted simulation and radiative-transfer assumptions.

A team including Caltech researchers trained a two-dimensional convolutional neural network, CDE, to invert low signal-to-noise Lyman-alpha integral-field cubes into line-of-sight integrated gas column-density maps. Training data come from IllustrisTNG100, post-processed with simplified Lyman-alpha radiative transfer and an approximate treatment of KCWI noise. On the simulated test set, CDE reaches a median linear RMSE of 1.48×10²⁰ cm^-2 and a median image SSIM of 0.635, both better than a pixel-wise power-law empirical baseline; applied to KCWI observations of the CC2.2 and Hyperion protoclusters, the predicted maps show filamentary structures extending about 40 arcseconds or more, though the results remain conditional on the adopted simulation and radiative-transfer assumptions.

PDZ-binding motifs (PBMs) are short C-terminal sequences that determine which PDZ domains a protein binds. Designing PBMs by affinity has generally required a dedicated model per domain and made off-target control hard. This work encodes PDZ and PBM sequences into one 128-dimensional unit hypersphere, predicts pKd from their cosine similarity, and writes single-target, target-minus-off-target and dual-target-mean objectives as directions on that sphere; in four-position redesign for DLG4-PDZ3, 84 of the top 100 candidates fell in the top 1% of measured fitness in a published mutant library, versus 64 for the comparison method.

PDZ-binding motifs (PBMs) are short C-terminal sequences that determine which PDZ domains a protein binds. Designing PBMs by affinity has generally required a dedicated model per domain and made off-target control hard. This work encodes PDZ and PBM sequences into one 128-dimensional unit hypersphere, predicts pKd from their cosine similarity, and writes single-target, target-minus-off-target and dual-target-mean objectives as directions on that sphere; in four-position redesign for DLG4-PDZ3, 84 of the top 100 candidates fell in the top 1% of measured fitness in a published mutant library, versus 64 for the comparison method.

Cloudflare has open-sourced Kitesurf, its stateless browser for AI agents, on GitHub under the Apache 2.0 license, so developers can read the code, run it locally, modify it, and deploy their own instance to their own Cloudflare Workers account. Kitesurf runs entirely on Workers, parsing HTML, applying CSS, executing page JavaScript and rendering screenshots without a browser container, and it supports the Chrome DevTools Protocol so clients such as Puppeteer can connect. Cloudflare says the goal is a lightweight browser for agents that uses less CPU and memory than a full browser like Chromium and scales on demand, while noting the project is under heavy development.

Cloudflare has open-sourced Kitesurf, its stateless browser for AI agents, on GitHub under the Apache 2.0 license, so developers can read the code, run it locally, modify it, and deploy their own instance to their own Cloudflare Workers account. Kitesurf runs entirely on Workers, parsing HTML, applying CSS, executing page JavaScript and rendering screenshots without a browser container, and it supports the Chrome DevTools Protocol so clients such as Puppeteer can connect. Cloudflare says the goal is a lightweight browser for agents that uses less CPU and memory than a full browser like Chromium and scales on demand, while noting the project is under heavy development.

Moving along a cosmological lightcone's line of sight also moves through cosmic time, so the lightcone can be read as a video of structure forming. ViViViT recasts reconstruction of the matter density field from 21 cm brightness-temperature lightcones as video-to-video translation, using attention factorized into transverse frames and radial evolution. On held-out 21cmFAST lightcones it keeps spectral power and phase alignment within a few percent of the ground truth and outperforms the tested 3D U-Net baseline across all measured statistics.

Moving along a cosmological lightcone's line of sight also moves through cosmic time, so the lightcone can be read as a video of structure forming. ViViViT recasts reconstruction of the matter density field from 21 cm brightness-temperature lightcones as video-to-video translation, using attention factorized into transverse frames and radial evolution. On held-out 21cmFAST lightcones it keeps spectral power and phase alignment within a few percent of the ground truth and outperforms the tested 3D U-Net baseline across all measured statistics.

Patient-derived xenografts (PDX) implant human tumors into immunocompromised mice to test drugs, but screening every tumor-drug combination is infeasible in time and cost. GrAFT-PDX is a Bayesian model that jointly models tumor growth trajectories and mouse death risk, using a low-rank tensor decomposition to share information across PDX lines, treatments and time, so it can impute untested experiments and predict death times. On the large PDXe screen it assigns higher log probability to death times than Royston-Parmar and other survival baselines, matches strong baselines on tumor response, and preserves known gene-drug associations without genomic input.

Patient-derived xenografts (PDX) implant human tumors into immunocompromised mice to test drugs, but screening every tumor-drug combination is infeasible in time and cost. GrAFT-PDX is a Bayesian model that jointly models tumor growth trajectories and mouse death risk, using a low-rank tensor decomposition to share information across PDX lines, treatments and time, so it can impute untested experiments and predict death times. On the large PDXe screen it assigns higher log probability to death times than Royston-Parmar and other survival baselines, matches strong baselines on tumor response, and preserves known gene-drug associations without genomic input.

For stochastic-gradient optimization with very many inequality constraints, an adaptive constraint-sampling method stops drawing each step's constraint uniformly from all constraints and instead maintains an "important set" of recently violated constraints, sampling from that set or from all constraints with fixed probabilities. On 100-dimensional synthetic convex problems it lowers the final Euclidean distance of the iterates to the constrained minimizer by about two orders of magnitude on average for ABSGD, Projection-SGD and Huber-penalty SGD, and by one to two orders for APDSG; in Dubins Car control-barrier-function synthesis it lowers maximum constraint violation by half an order to one order of magnitude.

For stochastic-gradient optimization with very many inequality constraints, an adaptive constraint-sampling method stops drawing each step's constraint uniformly from all constraints and instead maintains an "important set" of recently violated constraints, sampling from that set or from all constraints with fixed probabilities. On 100-dimensional synthetic convex problems it lowers the final Euclidean distance of the iterates to the constrained minimizer by about two orders of magnitude on average for ABSGD, Projection-SGD and Huber-penalty SGD, and by one to two orders for APDSG; in Dubins Car control-barrier-function synthesis it lowers maximum constraint violation by half an order to one order of magnitude.

Nano-IR hyperspectral imaging collects a spectrum point by point at every pixel, so a complete cube can take hours. CACNP feeds sparsely sampled spectra together with a few single-wavenumber structure images into a convolutional-attentive conditional neural process trained on synthetic cubes, returning the full cube and a per-voxel uncertainty. On experimental data of polystyrene and poly(methyl methacrylate) particles in epoxy, 5% of the spectra suffices, inference takes under 0.1 s, and reconstruction error approaches the sample noise level.

Nano-IR hyperspectral imaging collects a spectrum point by point at every pixel, so a complete cube can take hours. CACNP feeds sparsely sampled spectra together with a few single-wavenumber structure images into a convolutional-attentive conditional neural process trained on synthetic cubes, returning the full cube and a per-voxel uncertainty. On experimental data of polystyrene and poly(methyl methacrylate) particles in epoxy, 5% of the spectra suffices, inference takes under 0.1 s, and reconstruction error approaches the sample noise level.

Alibaba has open-sourced Open Code Review, its internal AI code-review assistant used for two years, and a user only needs to configure a model endpoint to start. It reads Git diffs and sends changed files to a configurable LLM through an agent with tool use, which can read full files, search the codebase and consult other changed files to produce structured, line-level review comments. On a benchmark of 50 open-source repositories, 200 real pull requests and 10 programming languages, it reaches clearly higher precision and F1 than a general-purpose agent with the same underlying model while consuming about one-ninth the tokens, but its recall is lower, a deliberate trade-off favoring precision over noise.

Alibaba has open-sourced Open Code Review, its internal AI code-review assistant used for two years, and a user only needs to configure a model endpoint to start. It reads Git diffs and sends changed files to a configurable LLM through an agent with tool use, which can read full files, search the codebase and consult other changed files to produce structured, line-level review comments. On a benchmark of 50 open-source repositories, 200 real pull requests and 10 programming languages, it reaches clearly higher precision and F1 than a general-purpose agent with the same underlying model while consuming about one-ninth the tokens, but its recall is lower, a deliberate trade-off favoring precision over noise.

Google released chrome-devtools-mcp, which lets coding agents such as Antigravity, Claude, Cursor and Copilot control and inspect a running Chrome browser through a Model-Context-Protocol server or a command-line interface. It hands agents DevTools performance tracing, network request analysis, screenshots and console messages, and uses puppeteer to perform actions and automatically wait for their results. The tool officially supports only Google Chrome and Chrome for Testing; other Chromium-based browsers may work but are not guaranteed.

Google released chrome-devtools-mcp, which lets coding agents such as Antigravity, Claude, Cursor and Copilot control and inspect a running Chrome browser through a Model-Context-Protocol server or a command-line interface. It hands agents DevTools performance tracing, network request analysis, screenshots and console messages, and uses puppeteer to perform actions and automatically wait for their results. The tool officially supports only Google Chrome and Chrome for Testing; other Chromium-based browsers may work but are not guaranteed.

mRNA foundation models usually average token embeddings over the whole transcript before fitting a linear probe, so signals from short regulatory regions are diluted by longer ones. A parameter-free region-aware readout instead averages within the 5′ UTR, coding sequence and 3′ UTR and concatenates the three means; across 20 frozen models it raises Pearson correlation on translation and stability tasks by 11–25% on average, with no model significantly worse on any regression task. On classification tasks AUPRC changes range from −3% to +11%.

mRNA foundation models usually average token embeddings over the whole transcript before fitting a linear probe, so signals from short regulatory regions are diluted by longer ones. A parameter-free region-aware readout instead averages within the 5′ UTR, coding sequence and 3′ UTR and concatenates the three means; across 20 frozen models it raises Pearson correlation on translation and stability tasks by 11–25% on average, with no model significantly worse on any regression task. On classification tasks AUPRC changes range from −3% to +11%.

Tencent Hunyuan released Hy ASR 3.0, a large-language-model-based speech recognition system covering Mandarin Chinese, English, Cantonese and 12 other Chinese dialect groups, with external context supplied as hotword lists, question-answer dialogues or summaries. On public Mandarin, English and Cantonese benchmarks it reaches aggregate error rates of 4.64%, 5.22% and 5.06%, each below the evaluated non-Hy systems; the Cantonese-only aggregate is about 11.0% lower than Qwen3 ASR Flash. The system also uses information inside the same utterance to disambiguate homophonic or phonetically similar keywords without a separate context input.

Tencent Hunyuan released Hy ASR 3.0, a large-language-model-based speech recognition system covering Mandarin Chinese, English, Cantonese and 12 other Chinese dialect groups, with external context supplied as hotword lists, question-answer dialogues or summaries. On public Mandarin, English and Cantonese benchmarks it reaches aggregate error rates of 4.64%, 5.22% and 5.06%, each below the evaluated non-Hy systems; the Cantonese-only aggregate is about 11.0% lower than Qwen3 ASR Flash. The system also uses information inside the same utterance to disambiguate homophonic or phonetically similar keywords without a separate context input.

Claude Code's changelog says sandboxed Bash commands now follow the normal permission flow by default instead of running without prompts; keeping prompt-free execution requires setting sandbox.autoAllowBashIfSandboxed. The same batch makes read-only git commands ask first in repositories that are not the project's or trusted, and unpacks plugin MCP bundles under the plugins directory so read-only plugin folders load. These changes concern the permission boundary for an agent running commands and loading plugins locally, not model capability.

Claude Code's changelog says sandboxed Bash commands now follow the normal permission flow by default instead of running without prompts; keeping prompt-free execution requires setting sandbox.autoAllowBashIfSandboxed. The same batch makes read-only git commands ask first in repositories that are not the project's or trusted, and unpacks plugin MCP bundles under the plugins directory so read-only plugin folders load. These changes concern the permission boundary for an agent running commands and loading plugins locally, not model capability.

Electromagnetic signal processing has long trained a separate expert per task, limiting cross-task reuse, while electromagnetic foundation models pay for large-scale pretraining and shared-backbone inference. GLCR instead routes over a frozen heterogeneous expert pool: a lightweight one-dimensional CNN with an additive-margin head makes the global dispatch, and a local re-evaluation with four physical statistics handles confusable task pairs. On the HeteroSignal-7 benchmark it reaches 96.16% routing accuracy and a 76.30% system score, above all 19 baselines, at 1.43 ms routing latency per sample.

Electromagnetic signal processing has long trained a separate expert per task, limiting cross-task reuse, while electromagnetic foundation models pay for large-scale pretraining and shared-backbone inference. GLCR instead routes over a frozen heterogeneous expert pool: a lightweight one-dimensional CNN with an additive-margin head makes the global dispatch, and a local re-evaluation with four physical statistics handles confusable task pairs. On the HeteroSignal-7 benchmark it reaches 96.16% routing accuracy and a 76.30% system score, above all 19 baselines, at 1.43 ms routing latency per sample.

In ESM2 residue embeddings, polar/nonpolar binary patterns organize into a readable geometry. Generating 50 sequences for each of all 128 heptad binary patterns, 6400 in total, and running position-preserving PCA yields a jellyfish-shaped manifold: hydrophobic content spreads along PC1, while PC2 and PC3 encode heptad phase as near sine and cosine. A charge alphabet stays ordered, but scrambled amino-acid groupings or the non-protein model ByT5 lose the cyclic organization.

In ESM2 residue embeddings, polar/nonpolar binary patterns organize into a readable geometry. Generating 50 sequences for each of all 128 heptad binary patterns, 6400 in total, and running position-preserving PCA yields a jellyfish-shaped manifold: hydrophobic content spreads along PC1, while PC2 and PC3 encode heptad phase as near sine and cosine. A charge alphabet stays ordered, but scrambled amino-acid groupings or the non-protein model ByT5 lose the cyclic organization.

Anthropic's Claude Code changelog adds Code Mode, an experimental opt-in in which Claude writes JavaScript to call its own tools instead of issuing tool calls one at a time. The same update lets plugins declare an engines["claude-code"] version range in plugin.json, so plugins outside their stated range are not installed, updated or loaded, and install and update pick the newest version that fits. The claude plugin validate command now also type-checks a plugin's TypeScript hooks module and reports each finding as a warning.

Anthropic's Claude Code changelog adds Code Mode, an experimental opt-in in which Claude writes JavaScript to call its own tools instead of issuing tool calls one at a time. The same update lets plugins declare an engines["claude-code"] version range in plugin.json, so plugins outside their stated range are not installed, updated or loaded, and install and update pick the newest version that fits. The claude plugin validate command now also type-checks a plugin's TypeScript hooks module and reports each finding as a warning.

Observational data often leave causal directions unresolved or unstable, and large language models can propose directions from variable names and descriptions. This study splits the use of semantic information into query allocation, proposal admission and graph retention, and tests each on the Asia, Sachs, Child, Insurance and Alarm benchmark networks. Semantic orientation beats coin-flip orientation, but targeting unstable edges shows no established advantage over matched-random allocation; the tested admission certificate can only abstain given the available calibration tasks, and learned references contain the true DAG in only three of 80 planned tasks.

Observational data often leave causal directions unresolved or unstable, and large language models can propose directions from variable names and descriptions. This study splits the use of semantic information into query allocation, proposal admission and graph retention, and tests each on the Asia, Sachs, Child, Insurance and Alarm benchmark networks. Semantic orientation beats coin-flip orientation, but targeting unstable edges shows no established advantage over matched-random allocation; the tested admission certificate can only abstain given the available calibration tasks, and learned references contain the true DAG in only three of 80 planned tasks.

DeepSeek has exposed the search, file parsing, link reading, safety checking and voice capabilities behind the DeepSeek App through the DeepSeek API, so developers can switch on built-in features with a few parameters or call standalone APIs such as OCR, ASR and TTS. V4.1-Flash (deepseek-flash) supports these capabilities across Messages, Responses and Chat Completions, with tool calls and context assembly handled server-side. Extended capabilities incur tool usage fees on top of standard token charges, and public-facing apps must enable safety checks.

DeepSeek has exposed the search, file parsing, link reading, safety checking and voice capabilities behind the DeepSeek App through the DeepSeek API, so developers can switch on built-in features with a few parameters or call standalone APIs such as OCR, ASR and TTS. V4.1-Flash (deepseek-flash) supports these capabilities across Messages, Responses and Chat Completions, with tool calls and context assembly handled server-side. Extended capabilities incur tool usage fees on top of standard token charges, and public-facing apps must enable safety checks.

OpenRouter compared three ways to get reliable JSON out of models, using 24 messy support messages across nine models that support structured outputs: send a strict JSON Schema, force a single tool call, or switch to a decision model that only returns labels. Across the nine models, 209 of 216 replies matched a strict schema, and all 7 misses came from one provider that ignores the schema; asking for JSON only in the prompt produced 18 failures, each from a code fence. For GLM 5.3 FlashX, which only supports tools, forcing the tool call gave 22 valid outputs out of 24, against 0 when the schema was sent directly.

OpenRouter compared three ways to get reliable JSON out of models, using 24 messy support messages across nine models that support structured outputs: send a strict JSON Schema, force a single tool call, or switch to a decision model that only returns labels. Across the nine models, 209 of 216 replies matched a strict schema, and all 7 misses came from one provider that ignores the schema; asking for JSON only in the prompt produced 18 failures, each from a code fence. For GLM 5.3 FlashX, which only supports tools, forcing the tool call gave 22 valid outputs out of 24, against 0 when the schema was sent directly.

Cohere said it has completed its acquisition of AISLE, a company building autonomous systems for vulnerability discovery and remediation, and has welcomed the team behind cubic, an AI-powered code review platform; the two are to form the foundation of Cohere Cyber, its cybersecurity business, and of a secure coding vertical. Cohere had already acquired Aleph Alpha and Reliant AI, and says these two additions bring its strategic M&A transactions this year to four. AISLE's co-founders take leadership roles in cybersecurity, corporate security and AI security research, and Cohere plans a substantial long-term investment in Prague as its global center for AI cybersecurity and autonomous security systems.

Cohere said it has completed its acquisition of AISLE, a company building autonomous systems for vulnerability discovery and remediation, and has welcomed the team behind cubic, an AI-powered code review platform; the two are to form the foundation of Cohere Cyber, its cybersecurity business, and of a secure coding vertical. Cohere had already acquired Aleph Alpha and Reliant AI, and says these two additions bring its strategic M&A transactions this year to four. AISLE's co-founders take leadership roles in cybersecurity, corporate security and AI security research, and Cohere plans a substantial long-term investment in Prague as its global center for AI cybersecurity and autonomous security systems.

Sparse autoencoders expand the dense 128-dimensional latent space at each ProteinMPNN encoder layer into an overcomplete sparse space, where individual dimensions correlate with protein features such as secondary structure, residue volume, residue shape, disulfide bridges and ligand coordination. On the Brolucizumab short-chain variable fragment, the mod=0, 0.5 and 1 candidates with branched-chain-amino-acid dimensions raised held lower RMSD at 433 K than the native sequence, the ProteinMPNN sequence and the GeoEvoBuilder sequence, while mod=2, the largest raise, unfolded substantially at that temperature; the unmodified mod=0 candidate improved too, showing the autoencoder itself shifts predictions.

Sparse autoencoders expand the dense 128-dimensional latent space at each ProteinMPNN encoder layer into an overcomplete sparse space, where individual dimensions correlate with protein features such as secondary structure, residue volume, residue shape, disulfide bridges and ligand coordination. On the Brolucizumab short-chain variable fragment, the mod=0, 0.5 and 1 candidates with branched-chain-amino-acid dimensions raised held lower RMSD at 433 K than the native sequence, the ProteinMPNN sequence and the GeoEvoBuilder sequence, while mod=2, the largest raise, unfolded substantially at that temperature; the unmodified mod=0 candidate improved too, showing the autoencoder itself shifts predictions.

Automatic glaucoma screening degrades when a model moves to another hospital and camera; a neuro-symbolic framework combines anatomical graph reasoning with differentiable fuzzy clinical rules in one network, trained on 9,100 fundus images drawn from a corpus of 12,619 images across eight datasets in eight countries and tested on 485 fully held-out RIM-ONE images from a different Spanish hospital and camera. On the in-distribution test set, EfficientNet-B4 with a graph attention network reaches an AUC of 0.9612 (95% CI 0.951–0.972) versus 0.9472 for EfficientNet-B4 alone, yet the combined full model scores 0.6155 cross-hospital, 8.1 points below the baseline.

Automatic glaucoma screening degrades when a model moves to another hospital and camera; a neuro-symbolic framework combines anatomical graph reasoning with differentiable fuzzy clinical rules in one network, trained on 9,100 fundus images drawn from a corpus of 12,619 images across eight datasets in eight countries and tested on 485 fully held-out RIM-ONE images from a different Spanish hospital and camera. On the in-distribution test set, EfficientNet-B4 with a graph attention network reaches an AUC of 0.9612 (95% CI 0.951–0.972) versus 0.9472 for EfficientNet-B4 alone, yet the combined full model scores 0.6155 cross-hospital, 8.1 points below the baseline.

RadixArk, working with NVIDIA, ran every sandbox of its open-source RL post-training framework Miles on NVIDIA Vera CPUs, which it calls the first public result of end-to-end agentic RL training with all sandboxes on Vera CPU servers. In a comparison that first kept the machines busy, with the same model, tasks, GPUs and settings, episodes with sandboxes on Turin took 12% longer on average, while model thinking time was identical, so nearly all of the gap came from work inside the sandbox. With a single sandbox running, Vera performs on par with or better than current x86, and its lead grows with load.

RadixArk, working with NVIDIA, ran every sandbox of its open-source RL post-training framework Miles on NVIDIA Vera CPUs, which it calls the first public result of end-to-end agentic RL training with all sandboxes on Vera CPU servers. In a comparison that first kept the machines busy, with the same model, tasks, GPUs and settings, episodes with sandboxes on Turin took 12% longer on average, while model thinking time was identical, so nearly all of the gap came from work inside the sandbox. With a single sandbox running, Vera performs on par with or better than current x86, and its lead grows with load.

OpenRouter compared three ways to give an agent the same GitHub capability, an MCP server, the gh command-line tool and a gh skill, counting input tokens across the whole task. On Claude Sonnet 5.5 all three answered correctly, but the GitHub MCP server with its 46 default tools used 59,417 input tokens, the CLI 1,421 and the skill 3,315. Scoping the server to the 2 tools the task needs cut it to 23,228, and capping tool results at 4,000 characters brought it to 5,357.

OpenRouter compared three ways to give an agent the same GitHub capability, an MCP server, the gh command-line tool and a gh skill, counting input tokens across the whole task. On Claude Sonnet 5.5 all three answered correctly, but the GitHub MCP server with its 46 default tools used 59,417 input tokens, the CLI 1,421 and the skill 3,315. Scoping the server to the 2 tools the task needs cut it to 23,228, and capping tool results at 4,000 characters brought it to 5,357.

On 5,357 structurally resolved small molecules from DrugBank 5.1.10, Schrödinger Bridge Onsager Flow drives the transport from a uniform chemical library toward a drug-likeness distribution to a total variation distance of 0.0141, against 0.1753 for a Sinkhorn baseline at its best setting across a ten-value regularization sweep. The method keeps one scalar potential per molecule updated by mirror descent instead of an all-pairs coupling matrix; its neural amortization, SBOFNet, returns a total variation distance of 0.0380 in a single forward pass with 1.6M parameters, at lower accuracy than the analytic solver.

On 5,357 structurally resolved small molecules from DrugBank 5.1.10, Schrödinger Bridge Onsager Flow drives the transport from a uniform chemical library toward a drug-likeness distribution to a total variation distance of 0.0141, against 0.1753 for a Sinkhorn baseline at its best setting across a ten-value regularization sweep. The method keeps one scalar potential per molecule updated by mirror descent instead of an all-pairs coupling matrix; its neural amortization, SBOFNet, returns a total variation distance of 0.0380 in a single forward pass with 1.6M parameters, at lower accuracy than the analytic solver.

Parallel's Parallel Decisions Fast model is now available on AI Gateway and free through November 23, 2026. It searches the live web to answer typed questions about companies, products, policies and other changing information, returning probabilities and cited evidence for an application to act on. Questions come in three kinds: boolean questions return the probability of yes, choice questions select from options the caller defines, and score questions assess an ordered rubric; related questions share context and retrieved evidence in one request.

Parallel's Parallel Decisions Fast model is now available on AI Gateway and free through November 23, 2026. It searches the live web to answer typed questions about companies, products, policies and other changing information, returning probabilities and cited evidence for an application to act on. Questions come in three kinds: boolean questions return the probability of yes, choice questions select from options the caller defines, and score questions assess an ordered rubric; related questions share context and retrieved evidence in one request.

DoorDash's autonomous delivery device Dot has begun making deliveries in Fremont, California, where it is manufactured by Sonic Manufacturing Technologies. DoorDash had already run autonomous delivery in other cities and built a standardized operating plan from that experience; in Fremont it signed an encroachment permit and is proceeding in phases, with Phase 1A capped at 10 robots for community engagement, testing and demonstration, chaperoned by local operators, and later phases gradually adding robots that can operate autonomously and be driven remotely as needed.

DoorDash's autonomous delivery device Dot has begun making deliveries in Fremont, California, where it is manufactured by Sonic Manufacturing Technologies. DoorDash had already run autonomous delivery in other cities and built a standardized operating plan from that experience; in Fremont it signed an encroachment permit and is proceeding in phases, with Phase 1A capped at 10 robots for community engagement, testing and demonstration, chaperoned by local operators, and later phases gradually adding robots that can operate autonomously and be driven remotely as needed.

Google has added Google Wallet to Gemini's list of connected apps under Personal Intelligence, so all Gemini app users in the U.S. can now tag @Google Wallet in a chat to ask about personal finance, budgeting and subscriptions, and Gemini generates dynamic charts of cash flow, asset allocation and budgets from that context. Users can optionally link credit cards and bank accounts to Wallet through Plaid for deeper analysis. Gemini can only read account information; it cannot move, manage or invest money, and the connection can be turned on or off at any time.

Google has added Google Wallet to Gemini's list of connected apps under Personal Intelligence, so all Gemini app users in the U.S. can now tag @Google Wallet in a chat to ask about personal finance, budgeting and subscriptions, and Gemini generates dynamic charts of cash flow, asset allocation and budgets from that context. Users can optionally link credit cards and bank accounts to Wallet through Plaid for deeper analysis. Gemini can only read account information; it cannot move, manage or invest money, and the connection can be turned on or off at any time.

Plannotator is a local, browser-based review surface for AI coding agents such as Claude Code, Codex, Copilot CLI and Gemini CLI. When an agent proposes a decision, plan or HTML artifact, the app opens in the browser so the user can annotate, comment and send feedback back to the agent; the new Plannotator Inbox gathers messages, questions and files from several agents into one local window, and replies do not require keeping a session open.

Plannotator is a local, browser-based review surface for AI coding agents such as Claude Code, Codex, Copilot CLI and Gemini CLI. When an agent proposes a decision, plan or HTML artifact, the app opens in the browser so the user can annotate, comment and send feedback back to the agent; the new Plannotator Inbox gathers messages, questions and files from several agents into one local window, and replies do not require keeping a session open.

NVIDIA released the cuda.bindings.cudla module in cuda-bindings 13.3.0, a thin Cython-generated Python layer over the cuDLA C++ runtime that lets Jetson AGX Orin applications create DLA devices, load networks, register buffers and submit inference without writing C++. The module targets hybrid mode, where DLA tasks are enqueued on a CUDA stream alongside GPU preprocessing and postprocessing, and the same API applies on RTX Spark. The release notes that standalone mode, which needs NvSci buffers and sync primitives, is not supported in this version.

NVIDIA released the cuda.bindings.cudla module in cuda-bindings 13.3.0, a thin Cython-generated Python layer over the cuDLA C++ runtime that lets Jetson AGX Orin applications create DLA devices, load networks, register buffers and submit inference without writing C++. The module targets hybrid mode, where DLA tasks are enqueued on a CUDA stream alongside GPU preprocessing and postprocessing, and the same API applies on RTX Spark. The release notes that standalone mode, which needs NvSci buffers and sync primitives, is not supported in this version.

Starting with version 5.2.0, Vercel's Workflow SDK sends event writes and stream writes to Vercel Workflows over a WebSocket by default, replacing one HTTP request per write, or per write group for streams, with no code change beyond the upgrade. In a benchmark replaying an AI agent's token stream, v5.2.0, which writes streams over a WebSocket by default, delivered the full response 4.2× faster than v4.8.12, with no chunk delayed over a second.

Starting with version 5.2.0, Vercel's Workflow SDK sends event writes and stream writes to Vercel Workflows over a WebSocket by default, replacing one HTTP request per write, or per write group for streams, with no code change beyond the upgrade. In a benchmark replaying an AI agent's token stream, v5.2.0, which writes streams over a WebSocket by default, delivered the full response 4.2× faster than v4.8.12, with no chunk delayed over a second.

Token Pass now lists GLM 5.3 Flash among its available models, alongside Kimi K3, GLM 5.3 and DeepSeek V4.1 Flash, with more models said to be coming. GLM 5.3 Flash is Z.ai's native multimodal coding model with a 1M-token context window. The subscription is aimed at Pro teams at $30 per month, grants $150 in monthly credit, runs through AI Gateway and has Zero Data Retention enabled by default.

Token Pass now lists GLM 5.3 Flash among its available models, alongside Kimi K3, GLM 5.3 and DeepSeek V4.1 Flash, with more models said to be coming. GLM 5.3 Flash is Z.ai's native multimodal coding model with a 1M-token context window. The subscription is aimed at Pro teams at $30 per month, grants $150 in monthly credit, runs through AI Gateway and has Zero Data Retention enabled by default.

Buzz is a self-hostable workspace where humans and AI agents share the same rooms, and each room is fixed by its URL. It writes messages, reactions, workflow steps, review approvals and git events as signed events in a single log, so people and processes share one identity model and one audit trail. Agents hold their own keys, channel memberships and audit records, and can open repos, send patches, review code, run workflows, edit canvases, orchestrate other agents and join voice huddles.

Buzz is a self-hostable workspace where humans and AI agents share the same rooms, and each room is fixed by its URL. It writes messages, reactions, workflow steps, review approvals and git events as signed events in a single log, so people and processes share one identity model and one audit trail. Agents hold their own keys, channel memberships and audit records, and can open repos, send patches, review code, run workflows, edit canvases, orchestrate other agents and join voice huddles.