AI and science frontiers · 2026-11-09

Braincodec, an encoding algorithm for auditory-cortex implants, compresses natural sound onto 256 or 512 stimulation channels at a reconstruction SI-SNR near 10, against roughly -30 for cochlear-implant encoders. Its stimulation patterns also resemble auditory-cortex responses, and implanted mice discriminated pure tones through those patterns as precisely as mice hearing real tones.

Braincodec, an encoding algorithm for auditory-cortex implants, compresses natural sound onto 256 or 512 stimulation channels at a reconstruction SI-SNR near 10, against roughly -30 for cochlear-implant encoders. Its stimulation patterns also resemble auditory-cortex responses, and implanted mice discriminated pure tones through those patterns as precisely as mice hearing real tones.

Posterior inference over phylogenetic trees has depended on likelihoods, which become intractable once real processes such as insertions, deletions or site selection are modeled. This family of conditional generative models builds a tree step by step, bottom-up, middle-out or top-down, and is trained with a discrete diffusion objective; on real and synthetic data it recovers MCMC posteriors more closely than the prior simulation-based method Phyloformer 2, matches IQ-Tree on simple data, and is reliably more accurate on more realistic complex models.

Posterior inference over phylogenetic trees has depended on likelihoods, which become intractable once real processes such as insertions, deletions or site selection are modeled. This family of conditional generative models builds a tree step by step, bottom-up, middle-out or top-down, and is trained with a discrete diffusion objective; on real and synthetic data it recovers MCMC posteriors more closely than the prior simulation-based method Phyloformer 2, matches IQ-Tree on simple data, and is reliably more accurate on more realistic complex models.

Human micro-Doppler from millimeter-wave radar has lacked a scalable data source: forward-only simulators, learned generators that hide the scene in network weights, and static-scene inverse renderers with no slow-time axis all fail to fit a real capture of a moving body. 4DPS represents the scene as oriented points carrying material, normal and velocity, and splats each point into range and Doppler bins through a closed-form Doppler point spread function, emitting the raw complex return a real radar records. On HuPR and RT-Pose the fitted renders reach median micro-Doppler correlations of 0.988 and 0.966, and pose models trained on renders alone keep 0.59 and 0.76 of their real-trained accuracy on real data.

Human micro-Doppler from millimeter-wave radar has lacked a scalable data source: forward-only simulators, learned generators that hide the scene in network weights, and static-scene inverse renderers with no slow-time axis all fail to fit a real capture of a moving body. 4DPS represents the scene as oriented points carrying material, normal and velocity, and splats each point into range and Doppler bins through a closed-form Doppler point spread function, emitting the raw complex return a real radar records. On HuPR and RT-Pose the fitted renders reach median micro-Doppler correlations of 0.988 and 0.966, and pose models trained on renders alone keep 0.59 and 0.76 of their real-trained accuracy on real data.

Continuous-time stochastic processes had no instrumental-variable framework, and discrete-time methods place the structural sparsity on finite-step transitions rather than on the actual mechanism. The work writes relevance, exclusion and exogeneity directly at the infinitesimal level of a stochastic differential equation's drift and builds moment conditions from the whole instrument path, giving an identification condition and a consistent estimator in the linear case. On confounded Ornstein-Uhlenbeck simulations, maximum likelihood becomes systematically biased as confounding grows while this estimator stays stable, and a single instrument process identifies multi-dimensional effects.

Continuous-time stochastic processes had no instrumental-variable framework, and discrete-time methods place the structural sparsity on finite-step transitions rather than on the actual mechanism. The work writes relevance, exclusion and exogeneity directly at the infinitesimal level of a stochastic differential equation's drift and builds moment conditions from the whole instrument path, giving an identification condition and a consistent estimator in the linear case. On confounded Ornstein-Uhlenbeck simulations, maximum likelihood becomes systematically biased as confounding grows while this estimator stays stable, and a single instrument process identifies multi-dimensional effects.

Single-cell gene-regulation modeling has split into two paths: sequence models read DNA regulatory grammar but lack cell-level resolution, while single-cell omics models treat genes and peaks as fixed-vocabulary tokens and discard the sequence. Muss derives peak and gene representations from DNA with a frozen sequence encoder, links them through peak–peak self-attention and distance-biased peak–gene interactions, and trains with peak masked language modeling plus paired scATAC-seq and scRNA-seq. Across four paired datasets it still predicts expression for genes and peaks unseen in task training, reaching pseudobulk PCC 0.684 on PBMC10K in the unseen-sequence setting.

Single-cell gene-regulation modeling has split into two paths: sequence models read DNA regulatory grammar but lack cell-level resolution, while single-cell omics models treat genes and peaks as fixed-vocabulary tokens and discard the sequence. Muss derives peak and gene representations from DNA with a frozen sequence encoder, links them through peak–peak self-attention and distance-biased peak–gene interactions, and trains with peak masked language modeling plus paired scATAC-seq and scRNA-seq. Across four paired datasets it still predicts expression for genes and peaks unseen in task training, reaching pseudobulk PCC 0.684 on PBMC10K in the unseen-sequence setting.

Acid-sensing ion channel hASIC1a is persistently activated during ischemia and drives neuronal death, yet small-molecule inhibitors lack selectivity or potency and venom peptides are limited by pH-dependent pharmacology and disulfide-bond production. Using RFdiffusion for backbones, ProteinMPNN for sequences and AlphaFold2 for filtering, the authors narrowed about 12,000 backbones and 70,000 sequences to 76 candidates and identified the 44-amino-acid Denasin1 by automated patch clamp. The miniprotein is disordered in solution, folds into a helix-turn-helix upon binding, inhibits hASIC1a with an IC50 of 23 ± 2.5 nM, and reduced infarct volume from 19.5 ± 9% to 8.6 ± 5.4% in a mouse transient middle cerebral artery occlusion model.

Acid-sensing ion channel hASIC1a is persistently activated during ischemia and drives neuronal death, yet small-molecule inhibitors lack selectivity or potency and venom peptides are limited by pH-dependent pharmacology and disulfide-bond production. Using RFdiffusion for backbones, ProteinMPNN for sequences and AlphaFold2 for filtering, the authors narrowed about 12,000 backbones and 70,000 sequences to 76 candidates and identified the 44-amino-acid Denasin1 by automated patch clamp. The miniprotein is disordered in solution, folds into a helix-turn-helix upon binding, inhibits hASIC1a with an IC50 of 23 ± 2.5 nM, and reduced infarct volume from 19.5 ± 9% to 8.6 ± 5.4% in a mouse transient middle cerebral artery occlusion model.

Computational drug-discovery agents usually rebuild their tool calls for every task, so a successful strategy rarely carries over. DrugHarness places more than 80 scientific operations and biomedical resources in one workspace, induces action hypotheses with explicit applicability conditions from execution trajectories, and admits them as skills only after an independent review agent checks sources and runs targeted trials when needed. On MolBench, the configuration without evolution reaches 91.9% affinity-comparison accuracy, 10.8 percentage points above reported MolClaw-CC results; in retrospective evaluation, frozen learned skills raise Qwen's optimization success from 79.5% to 87.2%.

Computational drug-discovery agents usually rebuild their tool calls for every task, so a successful strategy rarely carries over. DrugHarness places more than 80 scientific operations and biomedical resources in one workspace, induces action hypotheses with explicit applicability conditions from execution trajectories, and admits them as skills only after an independent review agent checks sources and runs targeted trials when needed. On MolBench, the configuration without evolution reaches 91.9% affinity-comparison accuracy, 10.8 percentage points above reported MolClaw-CC results; in retrospective evaluation, frozen learned skills raise Qwen's optimization success from 79.5% to 87.2%.

Online assignment of reusable resources such as reviewers, call agents and servers has long relied on greedy matching: dispatch whichever available resource scores highest now, ignoring the value lost when it is occupied later. LAG casts the problem as a contextual restless multi-armed bandit and uses an occupancy-measure linear program, together with its dual, to compute a closed-form index for each task-type and resource pair, equal to the match score minus the opportunity cost of keeping that resource free. On peer-review data and structured synthetic instances, LAG reaches 83–99% of that linear program's upper bound and improves over greedy by up to 12.3% on journal data and up to 36% on synthetic instances.

Online assignment of reusable resources such as reviewers, call agents and servers has long relied on greedy matching: dispatch whichever available resource scores highest now, ignoring the value lost when it is occupied later. LAG casts the problem as a contextual restless multi-armed bandit and uses an occupancy-measure linear program, together with its dual, to compute a closed-form index for each task-type and resource pair, equal to the match score minus the opportunity cost of keeping that resource free. On peer-review data and structured synthetic instances, LAG reaches 83–99% of that linear program's upper bound and improves over greedy by up to 12.3% on journal data and up to 36% on synthetic instances.

ALHENA distills spatial transcriptomic information into a routine H&E encoder that reads only images at inference. Rather than regressing individual genes, it trains the H&E representation to match the sample-to-sample similarity structure of two frozen transcriptomic models at cell and neighborhood scales. Across 110 cancer-type and signature combinations in TCGA, mean Spearman correlation was 0.341 versus 0.161 for unadapted UNI2-h; on the Thunder pathology benchmark the median change was −0.004.

ALHENA distills spatial transcriptomic information into a routine H&E encoder that reads only images at inference. Rather than regressing individual genes, it trains the H&E representation to match the sample-to-sample similarity structure of two frozen transcriptomic models at cell and neighborhood scales. Across 110 cancer-type and signature combinations in TCGA, mean Spearman correlation was 0.341 versus 0.161 for unadapted UNI2-h; on the Thunder pathology benchmark the median change was −0.004.

PhysioWave-T compresses multiscale wavelet analysis into a fixed-length token sequence: dynamic scale folding recombines the subbands into one signal before tokenization, so decomposition depth changes what each token carries rather than how long the sequence is. Across four EEG routes, folding cuts Transformer latency by 4.0–7.2 times and peak memory by 6.8–12.7 times over scale expansion without accuracy loss, and the 30M-parameter EEG encoder leads eight of nine metrics against BENDR, BIOT, LaBraM and EEGPT.

PhysioWave-T compresses multiscale wavelet analysis into a fixed-length token sequence: dynamic scale folding recombines the subbands into one signal before tokenization, so decomposition depth changes what each token carries rather than how long the sequence is. Across four EEG routes, folding cuts Transformer latency by 4.0–7.2 times and peak memory by 6.8–12.7 times over scale expansion without accuracy loss, and the 30M-parameter EEG encoder leads eight of nine metrics against BENDR, BIOT, LaBraM and EEGPT.

SafeLens pairs nonlinear machine learning with strict false-positive control for event-locked neural activity. It trains a random forest at each sliding-window position to separate event-locked trials from resting trials, then runs the same pipeline on set-aside resting recordings to build a model-matched empirical null, holding the family-wise error rate per channel. On simulated cortico-cortical evoked potentials injected into real resting intracranial EEG it resolved about 2.5 times as many polarity-resolved deflections as trial-average thresholding and the closest data-driven method, with 5 to 11 times fewer false components; in eight patients responses stayed above chance to 1,040 ms after stimulation, against 570 and 455 ms for the two benchmarks.

SafeLens pairs nonlinear machine learning with strict false-positive control for event-locked neural activity. It trains a random forest at each sliding-window position to separate event-locked trials from resting trials, then runs the same pipeline on set-aside resting recordings to build a model-matched empirical null, holding the family-wise error rate per channel. On simulated cortico-cortical evoked potentials injected into real resting intracranial EEG it resolved about 2.5 times as many polarity-resolved deflections as trial-average thresholding and the closest data-driven method, with 5 to 11 times fewer false components; in eight patients responses stayed above chance to 1,040 ms after stimulation, against 570 and 455 ms for the two benchmarks.

Partially observed dynamical systems expose only coarse-grained slow variables, while unresolved fast variables enter as random effects, so data-driven climate emulators must be probabilistic, yet generative models such as diffusion models lack stability guarantees over long autoregressive rollouts. Forced-dissipative stochastic interpolants write the energy injection and dissipation structure of turbulence into the drift network as a hard constraint, and a Foster–Lyapunov drift condition for the Markov chain proves geometric ergodicity of the emulator. On the two-scale Lorenz-96 prototype, the emulator stays stable over long integrations and under finite-amplitude external forcing, and accurately predicts the forced response of both mean and variance.

Partially observed dynamical systems expose only coarse-grained slow variables, while unresolved fast variables enter as random effects, so data-driven climate emulators must be probabilistic, yet generative models such as diffusion models lack stability guarantees over long autoregressive rollouts. Forced-dissipative stochastic interpolants write the energy injection and dissipation structure of turbulence into the drift network as a hard constraint, and a Foster–Lyapunov drift condition for the Markov chain proves geometric ergodicity of the emulator. On the two-scale Lorenz-96 prototype, the emulator stays stable over long integrations and under finite-amplitude external forcing, and accurately predicts the forced response of both mean and variance.

FlowAlign turns conditional flow matching from an inference-time generator into a training-time alignment signal: it first learns a source-conditioned velocity field along interpolation paths between paired samples, freezes it, and trains the source-side mapper on the velocity residual that a candidate representation induces. On the paired PBMC and BMMC single-cell multiome benchmarks it maps ATAC representations into a fixed RNA PCA space with lower FOSCTTM than the compared integration methods and architecture-matched paired baselines, while NMI and ARI are higher. Synthetic data and an image–text alignment task report the same kind of improvement.

FlowAlign turns conditional flow matching from an inference-time generator into a training-time alignment signal: it first learns a source-conditioned velocity field along interpolation paths between paired samples, freezes it, and trains the source-side mapper on the velocity residual that a candidate representation induces. On the paired PBMC and BMMC single-cell multiome benchmarks it maps ATAC representations into a fixed RNA PCA space with lower FOSCTTM than the compared integration methods and architecture-matched paired baselines, while NMI and ARI are higher. Synthetic data and an image–text alignment task report the same kind of improvement.

With only one-bit feedback, contextual search and contextual bilateral trade had left a visible gap between their upper bounds and the known lower bounds. A single trace-square-root potential now handles both problems, giving worst-case regret within a constant factor of the existing lower bounds, plus Bayesian bounds that depend only on prior covariance matrices and hold at every horizon.

With only one-bit feedback, contextual search and contextual bilateral trade had left a visible gap between their upper bounds and the known lower bounds. A single trace-square-root potential now handles both problems, giving worst-case regret within a constant factor of the existing lower bounds, plus Bayesian bounds that depend only on prior covariance matrices and hold at every horizon.

Interventional cardiac MRI needs reconstruction latency below 200 ms, while high image quality usually demands more sampling. KOSTAL places a recurrent network directly on undersampled golden-angle radial k-space, using the sampling mask to decide at each location whether to keep accumulated data or take the new spokes, then refines once in image space. On simulated OCMR data, 10-spoke frames reached a median SSIM of 0.907 versus 0.622 for 10-spoke gridding, in under 34 ms per frame.

Interventional cardiac MRI needs reconstruction latency below 200 ms, while high image quality usually demands more sampling. KOSTAL places a recurrent network directly on undersampled golden-angle radial k-space, using the sampling mask to decide at each location whether to keep accumulated data or take the new spokes, then refines once in image space. On simulated OCMR data, 10-spoke frames reached a median SSIM of 0.907 versus 0.622 for 10-spoke gridding, in under 34 ms per frame.

First-trimester ultrasound images for Down syndrome screening are low-resolution and structurally variable, and confirmed rare cases are scarce, so standard self-supervised pretraining struggles. RadarMAE instead masks alternating concentric rings and angular sectors, a radar-like pattern that forces the network to learn the continuous curvature and elongated shapes of craniofacial structures, reaching AUC 0.99 and 97% accuracy on an internal cohort. With domain-adversarial adaptation, the model calibrates on unlabeled images from four external hospitals and limits the drop in key metrics to 0.04 or less.

First-trimester ultrasound images for Down syndrome screening are low-resolution and structurally variable, and confirmed rare cases are scarce, so standard self-supervised pretraining struggles. RadarMAE instead masks alternating concentric rings and angular sectors, a radar-like pattern that forces the network to learn the continuous curvature and elongated shapes of craniofacial structures, reaching AUC 0.99 and 97% accuracy on an internal cohort. With domain-adversarial adaptation, the model calibrates on unlabeled images from four external hospitals and limits the drop in key metrics to 0.04 or less.

The distance of a quantum low-density parity-check code sets its error-correction capability, yet exact computation is NP-hard and heuristic search is slow, which limits screening of large code spaces. BB-Net turns distance estimation for bivariate bicycle (BB) codes into a learning task: an attention network trained on roughly 58,000 codes labeled with upper bounds from high-budget heuristic searches predicts distance directly from the sparse polynomial structure. On the test set, BB-Net reaches 96.39% accuracy against reference distances, above the 94.86% of the best heuristic in the same table, QDistEvol, and evaluates one code in 0.005 s.

The distance of a quantum low-density parity-check code sets its error-correction capability, yet exact computation is NP-hard and heuristic search is slow, which limits screening of large code spaces. BB-Net turns distance estimation for bivariate bicycle (BB) codes into a learning task: an attention network trained on roughly 58,000 codes labeled with upper bounds from high-budget heuristic searches predicts distance directly from the sparse polynomial structure. On the test set, BB-Net reaches 96.39% accuracy against reference distances, above the 94.86% of the best heuristic in the same table, QDistEvol, and evaluates one code in 0.005 s.

ICU resource planning has long relied on compact state-transition models hand-specified for a single policy question, so a new question means redefining states. A patient world model trained on MIMIC-IV and then frozen encodes irregular electronic health records as state-event sequences, returns predictions at any interval the simulator requests, and connects through scenario-specific mappings to shared bed and staff queues. Across admission, care-dispatch and discharge simulations it agrees with task-fitted Markov and semi-Markov comparators on directional responses to capacity and dispatch policies, and reproduces the dispatch gains obtained by replaying observed trajectories.

ICU resource planning has long relied on compact state-transition models hand-specified for a single policy question, so a new question means redefining states. A patient world model trained on MIMIC-IV and then frozen encodes irregular electronic health records as state-event sequences, returns predictions at any interval the simulator requests, and connects through scenario-specific mappings to shared bed and staff queues. Across admission, care-dispatch and discharge simulations it agrees with task-fitted Markov and semi-Markov comparators on directional responses to capacity and dispatch policies, and reproduces the dispatch gains obtained by replaying observed trajectories.

Continuous symmetries are usually found by writing the differential equations symbolically and solving an overdetermined determining system; here the search runs directly on sampled jet data through a linear operator that separates admissible generators from redundant directions. On a two-dimensional wave-equation benchmark the procedure recovers eleven generators in sequence, keeps validation symmetry residuals near machine precision, and reproduces Lie-bracket closure in a post-recovery check.

Continuous symmetries are usually found by writing the differential equations symbolically and solving an overdetermined determining system; here the search runs directly on sampled jet data through a linear operator that separates admissible generators from redundant directions. On a two-dimensional wave-equation benchmark the procedure recovers eleven generators in sequence, keeps validation symmetry residuals near machine precision, and reproduces Lie-bracket closure in a post-recovery check.

For the Dubreil-Jacotin-Long equation of internal solitary waves, a neural solver must not only resolve a solution but also select the physical wave among the trivial state, oscillatory wavetrains and broadened near-solutions. From one shared start, a Deep Ritz objective with an available-potential-energy constraint recovered the mode-1 wave, while strong- and weak-form residual objectives lost the prescribed APE within the first descent round; a residual refinement holding APE exact then met every acceptance item.

For the Dubreil-Jacotin-Long equation of internal solitary waves, a neural solver must not only resolve a solution but also select the physical wave among the trivial state, oscillatory wavetrains and broadened near-solutions. From one shared start, a Deep Ritz objective with an available-potential-energy constraint recovered the mode-1 wave, while strong- and weak-form residual objectives lost the prescribed APE within the first descent round; a residual refinement holding APE exact then met every acceptance item.

In sequential biological screens, LLM-proposed candidates are usually credited by hit counts, but the re-selection policy and random fallback also produce hits on their own. CADS runs the full loop beside an otherwise identical proposer-removed loop, certifies each discovery set separately, and reports the paired difference; in DepMap-backed simulations the no-LLM loop alone cleared the standard feedback gate, and origin counts inflated Qwen2.5-7B's net contribution from near zero to apparently positive.

In sequential biological screens, LLM-proposed candidates are usually credited by hit counts, but the re-selection policy and random fallback also produce hits on their own. CADS runs the full loop beside an otherwise identical proposer-removed loop, certifies each discovery set separately, and reports the paired difference; in DepMap-backed simulations the no-LLM loop alone cleared the standard feedback gate, and origin counts inflated Qwen2.5-7B's net contribution from near zero to apparently positive.

How well a language-model agent searches depends on the interface it is given, yet existing interfaces are set by hand and applied unchanged to every corpus. SearchWeaver jointly revises corpus artifacts, executable tools and usage instructions using the agent's execution trajectories and task rewards, then selects by downstream search performance. With GPT-5.4-nano, the optimized interfaces raise BrowseComp-Plus answer accuracy from 71.6% to 78.4% and average nDCG@10 across four BRIGHT subsets from 45.7 to 55.0, while cutting tool execution time.

How well a language-model agent searches depends on the interface it is given, yet existing interfaces are set by hand and applied unchanged to every corpus. SearchWeaver jointly revises corpus artifacts, executable tools and usage instructions using the agent's execution trajectories and task rewards, then selects by downstream search performance. With GPT-5.4-nano, the optimized interfaces raise BrowseComp-Plus answer accuracy from 71.6% to 78.4% and average nDCG@10 across four BRIGHT subsets from 45.7 to 55.0, while cutting tool execution time.

Multi-task deployment often requires merging independently fine-tuned models into one, but existing merging methods compress all tasks into a single shared low-rank space, so tasks whose subspaces are misaligned lose task-relevant directions. GRAM uses only the fine-tuned weights, first partitioning tasks by the principal-angle geometry of their output subspaces and then running a standard SVD-based merge inside each group. On GLUE with Flan-T5-Base it reaches 83.8% average accuracy, above the 81.9% of global merging and the 81.7% of the strongest baseline TSV-Merge, at the cost of one extra deployed model.

Multi-task deployment often requires merging independently fine-tuned models into one, but existing merging methods compress all tasks into a single shared low-rank space, so tasks whose subspaces are misaligned lose task-relevant directions. GRAM uses only the fine-tuned weights, first partitioning tasks by the principal-angle geometry of their output subspaces and then running a standard SVD-based merge inside each group. On GLUE with Flan-T5-Base it reaches 83.8% average accuracy, above the 81.9% of global merging and the 81.7% of the strongest baseline TSV-Merge, at the cost of one extra deployed model.

Brain-connectivity differences in autism and other neurodevelopmental disorders have mostly been read from group-average case-control contrasts, which flatten individual variation. BRIDGE combines a graph variational autoencoder with normative modeling to learn an age-varying reference trajectory of structural connectomes from 1,164 neurotypical participants, then reports a global deviation score from brain-age difference and an individualized regional deviation map from regional embeddings. In 229 autistic children, the global score correlated with ADOS (Spearman R=−0.291) and IQ (Pearson R=0.195) in males, and brain-age prediction error was 3%–9% lower than comparable deep-learning models.

Brain-connectivity differences in autism and other neurodevelopmental disorders have mostly been read from group-average case-control contrasts, which flatten individual variation. BRIDGE combines a graph variational autoencoder with normative modeling to learn an age-varying reference trajectory of structural connectomes from 1,164 neurotypical participants, then reports a global deviation score from brain-age difference and an individualized regional deviation map from regional embeddings. In 229 autistic children, the global score correlated with ADOS (Spearman R=−0.291) and IQ (Pearson R=0.195) in males, and brain-age prediction error was 3%–9% lower than comparable deep-learning models.

Titan-Llama converts the first four layers of Llama-3.2-1B-Instruct to blockwise attention and attaches a fixed-size neural memory module to each KV head, using test-time training to compress historical key-value bindings while the backbone stays frozen. Against segmented attention windows alone, accuracy rises from 23.87% to 40.51% on QuALITY and from 14.50% to 40.75% on LongHealth. In the evaluated configurations, prefill is more than 20% faster, decode 37% faster at medium context lengths, and steady-state decode memory 24% lower.

Titan-Llama converts the first four layers of Llama-3.2-1B-Instruct to blockwise attention and attaches a fixed-size neural memory module to each KV head, using test-time training to compress historical key-value bindings while the backbone stays frozen. Against segmented attention windows alone, accuracy rises from 23.87% to 40.51% on QuALITY and from 14.50% to 40.75% on LongHealth. In the evaluated configurations, prefill is more than 20% faster, decode 37% faster at medium context lengths, and steady-state decode memory 24% lower.

Surgical image analysis has long relied on encoders pretrained on natural images or on specialist models trained from scratch on comparatively small annotated datasets. InnerLogic-Surg starts from DINOv3 and applies continual self-supervised pretraining to millions of frames drawn from 11,965 public surgical videos, about 2,900 hours across more than fifteen procedure types, yielding an encoder family of 21M to 300M parameters. Under a frozen-backbone protocol with only a linear readout trained, the family achieves the best linear-probe performance on all 20 tasks spanning workflow, instrument and anatomy recognition, segmentation, detection and monocular depth, with ViT-L leading on 18 axes and ViT-B best on GraSP step recognition and monocular depth.

Surgical image analysis has long relied on encoders pretrained on natural images or on specialist models trained from scratch on comparatively small annotated datasets. InnerLogic-Surg starts from DINOv3 and applies continual self-supervised pretraining to millions of frames drawn from 11,965 public surgical videos, about 2,900 hours across more than fifteen procedure types, yielding an encoder family of 21M to 300M parameters. Under a frozen-backbone protocol with only a linear readout trained, the family achieves the best linear-probe performance on all 20 tasks spanning workflow, instrument and anatomy recognition, segmentation, detection and monocular depth, with ViT-L leading on 18 axes and ViT-B best on GraSP step recognition and monocular depth.

Adapting a protein foundation model to one laboratory's assay is an expensive, failure-prone step, and existing agents treat fine-tuning as a reliable tool call whose feasibility is estimated and whose infeasibility surfaces only after compute is spent. CAPO reframes adaptation as a gated decision: it measures feasibility on the target accelerator, admits or rejects a configuration against a budget, runs a 200-step canary inside the real training loop, and may decline. One unchanged system chose partial fine-tuning of ESM-2 8M, refused to fine-tune Boltz-2 in favor of a calibrated head over frozen representations, and applied rank-4 low-rank adaptation to ESM-2 150M across three protein tasks.

Adapting a protein foundation model to one laboratory's assay is an expensive, failure-prone step, and existing agents treat fine-tuning as a reliable tool call whose feasibility is estimated and whose infeasibility surfaces only after compute is spent. CAPO reframes adaptation as a gated decision: it measures feasibility on the target accelerator, admits or rejects a configuration against a budget, runs a 200-step canary inside the real training loop, and may decline. One unchanged system chose partial fine-tuning of ESM-2 8M, refused to fine-tune Boltz-2 in favor of a calibrated head over frozen representations, and applied rank-4 low-rank adaptation to ESM-2 150M across three protein tasks.

Discrete diffusion models have lacked the statistical theory that continuous diffusion enjoys, and it was unclear whether they can exploit sparse structure in the data distribution. This analysis casts masked-diffusion score estimation as a per-context label-smoothed classifier and shows the smoothing is equivalent to regularizing with the score loss of the uniform distribution. On sparse distributions where at most k sequences have positive probability, the terminal laws of both masked and uniform noising attain the sparse minimax rate, and the smoothing level can be chosen adaptively from the number of distinct training sequences without knowing the sparsity parameter.

Discrete diffusion models have lacked the statistical theory that continuous diffusion enjoys, and it was unclear whether they can exploit sparse structure in the data distribution. This analysis casts masked-diffusion score estimation as a per-context label-smoothed classifier and shows the smoothing is equivalent to regularizing with the score loss of the uniform distribution. On sparse distributions where at most k sequences have positive probability, the terminal laws of both masked and uniform noising attain the sparse minimax rate, and the smoothing level can be chosen adaptively from the number of distinct training sequences without knowing the sparsity parameter.

Trainable spiking neural networks have treated energy as a generic penalty blind to synapse type, even though cortical glutamatergic transmission costs substantially more ATP than GABAergic transmission. Embedding that cost asymmetry as a Lagrangian penalty inside surrogate-gradient training lets a single scalar penalty exert stronger gradient pressure on excitatory pathways in a Dale's-law-compliant recurrent microcircuit. On the Spiking Heidelberg Digits benchmark, the unpenalised network settled at 70.9% in an inhibition-dominated regime, while mild metabolic pressure raised accuracy to 79.5% and cut energy from 86,519 to 16,019 ATP units per sample.

Trainable spiking neural networks have treated energy as a generic penalty blind to synapse type, even though cortical glutamatergic transmission costs substantially more ATP than GABAergic transmission. Embedding that cost asymmetry as a Lagrangian penalty inside surrogate-gradient training lets a single scalar penalty exert stronger gradient pressure on excitatory pathways in a Dale's-law-compliant recurrent microcircuit. On the Spiking Heidelberg Digits benchmark, the unpenalised network settled at 70.9% in an inhibition-dominated regime, while mild metabolic pressure raised accuracy to 79.5% and cut energy from 86,519 to 16,019 ATP units per sample.

Motion direction estimation has lacked a unified model spanning species and brain regions, since existing models tend to fit one animal class, one brain area or one phenomenon. INME uses only cells selective to the four cardinal directions, embedding dendritic nonlinearity and a back-propagating action potential in one circuit with a post-summation readout stage. Across eleven synthetic-stimulus experiments it qualitatively reproduces phenomena from contrast saturation and reverse-Phi to the breaking of time-reversal symmetry, while five deep optical-flow networks with millions of parameters reproduce only a few of them.

Motion direction estimation has lacked a unified model spanning species and brain regions, since existing models tend to fit one animal class, one brain area or one phenomenon. INME uses only cells selective to the four cardinal directions, embedding dendritic nonlinearity and a back-propagating action potential in one circuit with a post-summation readout stage. Across eleven synthetic-stimulus experiments it qualitatively reproduces phenomena from contrast saturation and reverse-Phi to the breaking of time-reversal symmetry, while five deep optical-flow networks with millions of parameters reproduce only a few of them.

Streaming video understanding must retain evidence before anyone knows which details later queries will need, yet existing methods flatten event transitions when compressing history and cannot recover details once a frame has passed. DeltaStreamer explicitly records visual changes between adjacent frames in a hierarchical textual memory, and after a query arrives a query-driven planner decides whether to retrieve, keep observing, or replay historical frames. It scores 66.07% overall on OVO-Bench and 59.53% on StreamingBench, above the representative streaming methods it is compared with.

Streaming video understanding must retain evidence before anyone knows which details later queries will need, yet existing methods flatten event transitions when compressing history and cannot recover details once a frame has passed. DeltaStreamer explicitly records visual changes between adjacent frames in a hierarchical textual memory, and after a query arrives a query-driven planner decides whether to retrieve, keep observing, or replay historical frames. It scores 66.07% overall on OVO-Bench and 59.53% on StreamingBench, above the representative streaming methods it is compared with.

Closed-loop brain-stimulation controllers are usually trained against a preset target and struggle to follow real-time fluctuations in brain activity. CLEAR-FORCE combines the reservoir network's output error with a feedback error taken from the population membrane potential of a simulated spiking network, updating readout weights online while stimulation runs. In two spiking-neuron simulations it induced 30 Hz oscillations and suppressed burst events; error was lower than open loop when feedback strength was above zero, but weaker feedback did worse than open loop.

Closed-loop brain-stimulation controllers are usually trained against a preset target and struggle to follow real-time fluctuations in brain activity. CLEAR-FORCE combines the reservoir network's output error with a feedback error taken from the population membrane potential of a simulated spiking network, updating readout weights online while stimulation runs. In two spiking-neuron simulations it induced 30 Hz oscillations and suppressed burst events; error was lower than open loop when feedback strength was above zero, but weaker feedback did worse than open loop.

Multi-objective molecular optimization asks a single generation to improve several properties at once while keeping the scaffold similar, yet earlier LLM-based methods either injected previous-turn feedback into the prompt at inference without training for it, or trained only on non-iterative single-turn generation. MFG recasts the task as a multi-turn problem with feedback and uses reinforcement learning to train a model that exploits each turn's property changes and similarity to the source molecule; built on GeLLM3O, GeLLM3O-MFG reaches an average similarity-constrained success rate of 81.50% across ten MuMOInstruct tasks.

Multi-objective molecular optimization asks a single generation to improve several properties at once while keeping the scaffold similar, yet earlier LLM-based methods either injected previous-turn feedback into the prompt at inference without training for it, or trained only on non-iterative single-turn generation. MFG recasts the task as a multi-turn problem with feedback and uses reinforcement learning to train a model that exploits each turn's property changes and similarity to the source molecule; built on GeLLM3O, GeLLM3O-MFG reaches an average similarity-constrained success rate of 81.50% across ten MuMOInstruct tasks.

Cross-subject EEG-to-image reconstruction has relied on semantic embeddings learned from source subjects, and those embeddings drift off the image-embedding manifold for an unlabeled new subject. SPSC keeps object-level semantic neighborhoods during training with an object-mask contrastive regularizer, then at inference selects semantic candidates inside PCA subspaces of image prototypes using unlabeled target EEG batch statistics. Under leave-one-subject-out evaluation across ten subjects on THINGS-EEG, reconstruction rises from 0.552 to 0.713 on CLIP and from 0.612 to 0.745 on Alex.5.

Cross-subject EEG-to-image reconstruction has relied on semantic embeddings learned from source subjects, and those embeddings drift off the image-embedding manifold for an unlabeled new subject. SPSC keeps object-level semantic neighborhoods during training with an object-mask contrastive regularizer, then at inference selects semantic candidates inside PCA subspaces of image prototypes using unlabeled target EEG batch statistics. Under leave-one-subject-out evaluation across ten subjects on THINGS-EEG, reconstruction rises from 0.552 to 0.713 on CLIP and from 0.612 to 0.745 on Alex.5.

Ancient microbial data are scarce and contamination-prone, and a language model left to plan an analysis may cite files that do not exist, apply operations a format forbids, or answer questions the data cannot settle. PaleoRigor confines the model's planner to predefined local modules, validates file references, formats and step dependencies before anything runs, refuses out-of-scope requests with a named reason code, and saves the plan, intermediate files, checksums and decisions. In a preregistered frozen held-out evaluation it passed 23 of 24 runs (95.8%), including all 12 boundary decisions; the same model without the planning contract passed 19.

Ancient microbial data are scarce and contamination-prone, and a language model left to plan an analysis may cite files that do not exist, apply operations a format forbids, or answer questions the data cannot settle. PaleoRigor confines the model's planner to predefined local modules, validates file references, formats and step dependencies before anything runs, refuses out-of-scope requests with a named reason code, and saves the plan, intermediate files, checksums and decisions. In a preregistered frozen held-out evaluation it passed 23 of 24 runs (95.8%), including all 12 boundary decisions; the same model without the planning contract passed 19.

A-Alpha Bio has launched the Atlas Data Consortium to address the shortage of training data for protein models by sharing experimental data rather than trained models. Earlier sharing efforts relied mainly on federated learning, which pools data companies already hold but cannot enlarge the total. Five founding members each design one Data Block of one million protein-protein affinity measurements per quarter, all generated on the AlphaSeq platform under standard conditions and released quarterly to every member under a non-exclusive license. The first consortium focuses on VHH heavy-chain antibody variable domains, with consortia for other scaffolds expected from 2027.

A-Alpha Bio has launched the Atlas Data Consortium to address the shortage of training data for protein models by sharing experimental data rather than trained models. Earlier sharing efforts relied mainly on federated learning, which pools data companies already hold but cannot enlarge the total. Five founding members each design one Data Block of one million protein-protein affinity measurements per quarter, all generated on the AlphaSeq platform under standard conditions and released quarterly to every member under a non-exclusive license. The first consortium focuses on VHH heavy-chain antibody variable domains, with consortia for other scaffolds expected from 2027.

NVIDIA released Cosmos3-H-Dreams, moving interactive generative simulation from tabletop suturing to clinical surgical scenes. The earlier Cosmos-H-Dreams learned visual dynamics from synchronized video and robot kinematics on dVRK tabletop suturing; the new model is trained on clinical cholecystectomy recordings and kinematics from the CMR Surgical Versius platform, built on Cosmos 3 and accelerated by FlashDreams, runs interactively on a single RTX PRO 6000 GPU, and ships four-step and two-step checkpoints.

NVIDIA released Cosmos3-H-Dreams, moving interactive generative simulation from tabletop suturing to clinical surgical scenes. The earlier Cosmos-H-Dreams learned visual dynamics from synchronized video and robot kinematics on dVRK tabletop suturing; the new model is trained on clinical cholecystectomy recordings and kinematics from the CMR Surgical Versius platform, built on Cosmos 3 and accelerated by FlashDreams, runs interactively on a single RTX PRO 6000 GPU, and ships four-step and two-step checkpoints.

Differences between individuals in perceptual decision-making may be different mixtures of a few shared computational components. A low-rank neural network learned shared components and participant-specific weights jointly from trial-level choices; on the Bernoulli Clicks Task, out-of-sample prediction selected a six-dimensional representation of individual differences in the first dataset of 155 participants: two dimensions of temporal evidence weighting, two of choice-outcome history dependence, plus participant-specific side bias and signal-to-noise ratio. This low-dimensional organization reappeared in an independent replication dataset of 102 participants.

Differences between individuals in perceptual decision-making may be different mixtures of a few shared computational components. A low-rank neural network learned shared components and participant-specific weights jointly from trial-level choices; on the Bernoulli Clicks Task, out-of-sample prediction selected a six-dimensional representation of individual differences in the first dataset of 155 participants: two dimensions of temporal evidence weighting, two of choice-outcome history dependence, plus participant-specific side bias and signal-to-noise ratio. This low-dimensional organization reappeared in an independent replication dataset of 102 participants.

TRACE maps longitudinal visit records to interpretable progression phenotypes and lets the number of phenotypes exceed the number of diagnostic classes. In the label-matched setting on the ADNI cohort it achieves the highest purity, NMI, ARI, silhouette index, accuracy and macro-F1 at the current, 12-month and 24-month horizons; at 24 months it exceeds the strongest baseline by 3.1 and 5.8 percentage points in purity and silhouette index. With the phenotype count freed, it splits a new phenotype out of the cognitively-normal- and MCI-dominant groups, one whose cognitive-scale trajectories sit in between while its hippocampal-volume trajectory coincides with the MCI group.

TRACE maps longitudinal visit records to interpretable progression phenotypes and lets the number of phenotypes exceed the number of diagnostic classes. In the label-matched setting on the ADNI cohort it achieves the highest purity, NMI, ARI, silhouette index, accuracy and macro-F1 at the current, 12-month and 24-month horizons; at 24 months it exceeds the strongest baseline by 3.1 and 5.8 percentage points in purity and silhouette index. With the phenotype count freed, it splits a new phenotype out of the cognitively-normal- and MCI-dominant groups, one whose cognitive-scale trajectories sit in between while its hippocampal-volume trajectory coincides with the MCI group.

Text-driven CAD generation has long been optimized and evaluated on geometric similarity or visual consistency, yet a correct-looking shape does not guarantee a usable structure. PhyCAD runs executable CadQuery programs through a deterministic CAD–CAE pipeline that judges program executability, geometric and dimensional compliance, and finite-element stress, displacement and mass constraints in stages, then reshapes the GRPO training signal with a hierarchical advantage. On a held-out evaluation set, several open models trained with supervised fine-tuning followed by this reinforcement learning raise joint geometric-and-physical success from 21.76%–37.36% to 39.49%–58.05%.

Text-driven CAD generation has long been optimized and evaluated on geometric similarity or visual consistency, yet a correct-looking shape does not guarantee a usable structure. PhyCAD runs executable CadQuery programs through a deterministic CAD–CAE pipeline that judges program executability, geometric and dimensional compliance, and finite-element stress, displacement and mass constraints in stages, then reshapes the GRPO training signal with a hierarchical advantage. On a held-out evaluation set, several open models trained with supervised fine-tuning followed by this reinforcement learning raise joint geometric-and-physical success from 21.76%–37.36% to 39.49%–58.05%.

Navier released Prism, a desktop app in which AI agents drive CAD modeling, meshing, physics solving, post-processing and report writing inside one environment. Hardware engineers previously moved work by hand among CAD, a mesher, a solver, a post-processor, Python or MATLAB and slides, so a single design study could take days and teams committed to a direction early and tested only a handful of variants. Prism runs twelve open-source solver families locally alongside Navier's in-house GPU finite-element solver Ferro, and a drone cooling module example runs from constraints to an engineering report.

Navier released Prism, a desktop app in which AI agents drive CAD modeling, meshing, physics solving, post-processing and report writing inside one environment. Hardware engineers previously moved work by hand among CAD, a mesher, a solver, a post-processor, Python or MATLAB and slides, so a single design study could take days and teams committed to a direction early and tested only a handful of variants. Prism runs twelve open-source solver families locally alongside Navier's in-house GPU finite-element solver Ferro, and a drone cooling module example runs from constraints to an engineering report.

Polynomial chaos expansions have long offered only global error metrics, leaving the reliability of an individual prediction unstated; normalizing jackknife+ conformal nonconformity scores with an auxiliary polynomial chaos model makes interval width follow local input uncertainty. For fixed-basis, least-squares-fitted expansions every quantity is available in closed form, so the finite-sample jackknife+ coverage guarantee is retained without retraining; on a cantilever-beam benchmark with known heteroscedastic noise and on steam-generator clogging simulation data, the adaptive intervals reach target coverage and clearly outperform non-adaptive ones in high-uncertainty regions.

Polynomial chaos expansions have long offered only global error metrics, leaving the reliability of an individual prediction unstated; normalizing jackknife+ conformal nonconformity scores with an auxiliary polynomial chaos model makes interval width follow local input uncertainty. For fixed-basis, least-squares-fitted expansions every quantity is available in closed form, so the finite-sample jackknife+ coverage guarantee is retained without retraining; on a cantilever-beam benchmark with known heteroscedastic noise and on steam-generator clogging simulation data, the adaptive intervals reach target coverage and clearly outperform non-adaptive ones in high-uncertainty regions.

Patent examination asks an agent to find sparse prior art, refine queries, inspect evidence and write a novelty judgment, yet existing benchmarks mostly stop at static candidate sets. PatEx-Gym turns that process into a runnable environment: PatEx-Bench supplies 1K examination tasks, retrieval corpora of over 1M patents and 52K scientific papers, and a runtime supporting iterative search and Office Action generation. PASS synthesizes search trajectories by treating document visits as implicit relevance signals, and the 4B PatEx-Agent fine-tuned on them performs competitively against substantially larger general-purpose agents.

Patent examination asks an agent to find sparse prior art, refine queries, inspect evidence and write a novelty judgment, yet existing benchmarks mostly stop at static candidate sets. PatEx-Gym turns that process into a runnable environment: PatEx-Bench supplies 1K examination tasks, retrieval corpora of over 1M patents and 52K scientific papers, and a runtime supporting iterative search and Office Action generation. PASS synthesizes search trajectories by treating document visits as implicit relevance signals, and the 4B PatEx-Agent fine-tuned on them performs competitively against substantially larger general-purpose agents.

Piano performance MIDI-to-score transcription has treated notes as the only modeling unit, yet beats can fall on frames with no note onset, so metrical information struggles to reach the notation. DuET runs a note encoder and a frame-level encoder over notes and a fixed time grid, exchanges information through bidirectional cross-attention, and lets a decoder alternate between the two representations to predict durations, bars, staves and other score attributes. On the popular-piano Gangqinpu test split it lowers MUSTER mean error from 6.56% for the strongest baseline to 3.64%, and on the classical ASAP split from 11.30% to 11.22%.

Piano performance MIDI-to-score transcription has treated notes as the only modeling unit, yet beats can fall on frames with no note onset, so metrical information struggles to reach the notation. DuET runs a note encoder and a frame-level encoder over notes and a fixed time grid, exchanges information through bidirectional cross-attention, and lets a decoder alternate between the two representations to predict durations, bars, staves and other score attributes. On the popular-piano Gangqinpu test split it lowers MUSTER mean error from 6.56% for the strongest baseline to 3.64%, and on the classical ASAP split from 11.30% to 11.22%.

A 9B small language model running on one local GPU performs case-level and trio-based variant interpretation, recovering 93% of expert-reported causal variants across 327 solved singleton visual-impairment cases and classifying 82% of them as pathogenic or likely pathogenic.

A 9B small language model running on one local GPU performs case-level and trio-based variant interpretation, recovering 93% of expert-reported causal variants across 327 solved singleton visual-impairment cases and classifying 82% of them as pathogenic or likely pathogenic.

Preferred Networks released PLaMo 3 Translate, a translation-specialized LLM suite originally built for its own PLaMo Translate service, as open weights on Hugging Face in four sizes (0.8B, 2B, 8B, 31B), runnable entirely offline on smartphones, laptops or internal corporate servers across 53 languages including English and Japanese. Under the PLaMo Community License, individuals and enterprises with annual revenue of 1 billion yen or less in their most recent fiscal year can download and use the models free for commercial and non-commercial purposes; larger organizations need a commercial license. PFN says its comparative evaluations put even the smallest 0.8B version ahead of comparable translation-specific LLMs such as Google's TranslateGemma and Tencent's Hy-MT2.

Preferred Networks released PLaMo 3 Translate, a translation-specialized LLM suite originally built for its own PLaMo Translate service, as open weights on Hugging Face in four sizes (0.8B, 2B, 8B, 31B), runnable entirely offline on smartphones, laptops or internal corporate servers across 53 languages including English and Japanese. Under the PLaMo Community License, individuals and enterprises with annual revenue of 1 billion yen or less in their most recent fiscal year can download and use the models free for commercial and non-commercial purposes; larger organizations need a commercial license. PFN says its comparative evaluations put even the smallest 0.8B version ahead of comparable translation-specific LLMs such as Google's TranslateGemma and Tencent's Hy-MT2.

AstraZeneca signed a new strategic drug discovery collaboration and licensing agreement with AI therapeutics company insitro to identify novel targets and advance new programmes for metabolic dysfunction-associated steatohepatitis (MASH) and cardiomyopathies. insitro will apply its Virtual Human AI platform, which combines large-scale human and cellular data with AI-enabled target identification and medicine design, to identify causal disease biology and move AstraZeneca-nominated targets from discovery to development candidate. insitro receives a $40 million upfront payment and is eligible for up to approximately $1.46 billion in development and commercial milestones plus single-digit royalties, while AstraZeneca holds exclusive worldwide rights to selected programmes.

AstraZeneca signed a new strategic drug discovery collaboration and licensing agreement with AI therapeutics company insitro to identify novel targets and advance new programmes for metabolic dysfunction-associated steatohepatitis (MASH) and cardiomyopathies. insitro will apply its Virtual Human AI platform, which combines large-scale human and cellular data with AI-enabled target identification and medicine design, to identify causal disease biology and move AstraZeneca-nominated targets from discovery to development candidate. insitro receives a $40 million upfront payment and is eligible for up to approximately $1.46 billion in development and commercial milestones plus single-digit royalties, while AstraZeneca holds exclusive worldwide rights to selected programmes.

Long-read DNA often carries frameshift errors, and aligning it to protein family profiles has meant either translating six reading frames or correcting errors first, both of which cost sensitivity. DUMMEST follows DUMMER's length-bias-free alignment idea and adds nucleotide-emitting indel states to the match, insert and delete states, approximating frameshifts with combinations of nucleotides and codons, with GPU-accelerated MMseqs2 multistage prefiltering. On simulated long reads it detects more alignments with greater average length, matches BATH in sensitivity, improves specificity under simulated sequencing artifacts, and runs 2–12× faster depending on dataset.

Long-read DNA often carries frameshift errors, and aligning it to protein family profiles has meant either translating six reading frames or correcting errors first, both of which cost sensitivity. DUMMEST follows DUMMER's length-bias-free alignment idea and adds nucleotide-emitting indel states to the match, insert and delete states, approximating frameshifts with combinations of nucleotides and codons, with GPU-accelerated MMseqs2 multistage prefiltering. On simulated long reads it detects more alignments with greater average length, matches BATH in sensitivity, improves specificity under simulated sequencing artifacts, and runs 2–12× faster depending on dataset.

dsm-firmenich and Boltz PBC have announced a partnership that puts Boltz's latest proprietary foundation models to work in dsm-firmenich's ingredient-discovery process for next-generation scent and flavor molecules. The partners plan to combine those frontier models with dsm-firmenich's taste and odor receptor data to produce a set of exclusive models capturing decades of sensory-science know-how, used to predict and design molecular structure and to learn how molecules interact with receptor proteins. dsm-firmenich teams also gain full access to the Boltz platform, including the Boltz Lab interface, the Boltz API and a new integration with LLM agents that lets scientists orchestrate design and prediction workflows in natural language.

dsm-firmenich and Boltz PBC have announced a partnership that puts Boltz's latest proprietary foundation models to work in dsm-firmenich's ingredient-discovery process for next-generation scent and flavor molecules. The partners plan to combine those frontier models with dsm-firmenich's taste and odor receptor data to produce a set of exclusive models capturing decades of sensory-science know-how, used to predict and design molecular structure and to learn how molecules interact with receptor proteins. dsm-firmenich teams also gain full access to the Boltz platform, including the Boltz Lab interface, the Boltz API and a new integration with LLM agents that lets scientists orchestrate design and prediction workflows in natural language.

Polymer property prediction has long relied on molecular fragments cut by chemical rules, which may suit unfamiliar structures poorly and offer limited control over fragment size. A geometry-based fragmentation method casts fragmentation as a radius-constrained center-selection problem, where the number of fragments follows from a given radius rather than being fixed in advance, and changing the radius yields fragments at different scales. Feeding these multiscale fragments into a multimodal model that combines p-SMILES with 3D conformations lowers prediction error by roughly 5–10% relative to the baseline on several properties of a polymer benchmark, though the 3D modality's contribution varies by property.

Polymer property prediction has long relied on molecular fragments cut by chemical rules, which may suit unfamiliar structures poorly and offer limited control over fragment size. A geometry-based fragmentation method casts fragmentation as a radius-constrained center-selection problem, where the number of fragments follows from a given radius rather than being fixed in advance, and changing the radius yields fragments at different scales. Feeding these multiscale fragments into a multimodal model that combines p-SMILES with 3D conformations lowers prediction error by roughly 5–10% relative to the baseline on several properties of a polymer benchmark, though the 3D modality's contribution varies by property.

In vision-based landing, runway corner keypoints are often cropped by the field of view during close approach, leaving point-based pose solvers without geometric constraints. The work extends object detection from keypoints to keylines: each object instance predicts a fixed, semantically ordered set of infinite lines, implemented on YOLO with a continuous Hesse normal parameterization and Object Keyline Similarity (OKLS) serving as evaluation metric, target assignment signal and regression loss. On runway data, the continuous Hesse form with OKLS gives the highest keyline detection accuracy among four parameterizations, and a fused head adds keylines without degrading keypoint accuracy or meaningfully increasing latency.

In vision-based landing, runway corner keypoints are often cropped by the field of view during close approach, leaving point-based pose solvers without geometric constraints. The work extends object detection from keypoints to keylines: each object instance predicts a fixed, semantically ordered set of infinite lines, implemented on YOLO with a continuous Hesse normal parameterization and Object Keyline Similarity (OKLS) serving as evaluation metric, target assignment signal and regression loss. On runway data, the continuous Hesse form with OKLS gives the highest keyline detection accuracy among four parameterizations, and a fused head adds keylines without degrading keypoint accuracy or meaningfully increasing latency.

ScienceRSI splits scientific capability into specialization, verification, transfer and reuse, so an expert model contributes feedback only where it is genuinely better on the same scientific requirement. In the 4B-to-4B setting, scientific-reasoning average reached 55.0%, 2.0 points above GRPO; transferring from Qwen3-8B experts to a Qwen3.5-27B target raised the equal-weight average over seven benchmarks from 87.3% to 94.0%. Removing expert-contribution verification dropped the same 4B student's scientific-reasoning average from 55.0% to 50.6%.

ScienceRSI splits scientific capability into specialization, verification, transfer and reuse, so an expert model contributes feedback only where it is genuinely better on the same scientific requirement. In the 4B-to-4B setting, scientific-reasoning average reached 55.0%, 2.0 points above GRPO; transferring from Qwen3-8B experts to a Qwen3.5-27B target raised the equal-weight average over seven benchmarks from 87.3% to 94.0%. Removing expert-contribution verification dropped the same 4B student's scientific-reasoning average from 55.0% to 50.6%.

A laboratory mouse's age can be read from an ordinary photograph, but what transfers is relative age rather than absolute age. Across 2,846 photographs of 1,292 mice from seven facilities, a frozen DINOv2 vision model with ridge regression predicted age; on unseen facilities the median absolute error was 138 days, yet within a session the older of two mice was ranked older in 65% of pairs, rising above 95% for age gaps over 500 days. Mice predicted older than their age were frailer, with each 100 days of photographic age gap linked to 0.12 SD more frailty and a 14% higher hazard of death.

A laboratory mouse's age can be read from an ordinary photograph, but what transfers is relative age rather than absolute age. Across 2,846 photographs of 1,292 mice from seven facilities, a frozen DINOv2 vision model with ridge regression predicted age; on unseen facilities the median absolute error was 138 days, yet within a session the older of two mice was ranked older in 65% of pairs, rising above 95% for age gaps over 500 days. Mice predicted older than their age were frailer, with each 100 days of photographic age gap linked to 0.12 SD more frailty and a 14% higher hazard of death.

Multimodal emotional text-to-speech previously packed cross-speaker emotion regularities and individual expression into one emotion condition, so strengthening emotion could damage linguistic content. HiP-ETTS splits prompt-derived emotion into an emotion-shared factor and a complementary emotion factor, anchors linguistic structure with a hierarchical emotional prior, and refines emotional detail through adaptive projection and flow matching. Across audio, text and facial-image prompts its word error rate is lower than or comparable to prior methods, and its VAD-CCC and Emo-SIM emotion-consistency scores are the highest in every prompt condition, reaching 0.760 and 0.768 under audio prompts.

Multimodal emotional text-to-speech previously packed cross-speaker emotion regularities and individual expression into one emotion condition, so strengthening emotion could damage linguistic content. HiP-ETTS splits prompt-derived emotion into an emotion-shared factor and a complementary emotion factor, anchors linguistic structure with a hierarchical emotional prior, and refines emotional detail through adaptive projection and flow matching. Across audio, text and facial-image prompts its word error rate is lower than or comparable to prior methods, and its VAD-CCC and Emo-SIM emotion-consistency scores are the highest in every prompt condition, reaching 0.760 and 0.768 under audio prompts.

Intology extended its autonomous discovery system Locus from language-model post-training to molecular discovery, comparing it at matched compute against prior state-of-the-art methods from BioHub and Boltz and against frontier LLM agents Codex and Claude Code on molecular property prediction, small-molecule potency prediction and antibody design. On Boltz-2 adaptation across six targets, Locus reached a mean Spearman correlation of 0.622, against 0.225 for the Boltz-2.1 API, 0.450 for Claude Code and 0.569 for Codex. In antibody design, Locus reached comparable binding scores with 1.5–2.8× fewer generated candidates, but those scores are OpenDDE ipSAE proxies and still need wet-lab validation.

Intology extended its autonomous discovery system Locus from language-model post-training to molecular discovery, comparing it at matched compute against prior state-of-the-art methods from BioHub and Boltz and against frontier LLM agents Codex and Claude Code on molecular property prediction, small-molecule potency prediction and antibody design. On Boltz-2 adaptation across six targets, Locus reached a mean Spearman correlation of 0.622, against 0.225 for the Boltz-2.1 API, 0.450 for Claude Code and 0.569 for Codex. In antibody design, Locus reached comparable binding scores with 1.5–2.8× fewer generated candidates, but those scores are OpenDDE ipSAE proxies and still need wet-lab validation.

Reconstructing a 3D surface from a single high-frequency fringe image is hard because long-range spatial consistency and fine geometric detail must both survive. Convolutional networks model long-range context weakly, and window-based Transformers build global interactions only gradually with depth. A hybrid Mamba-Transformer network adds visual state-space blocks at the intermediate-resolution encoder stage of a Swin-style U-shaped architecture, with an axial-dilated block that strengthens directional fringe features before selective scanning. On a public fringe-projection dataset, using only one deformed fringe image, it reaches 0.548 mm RMSE and 0.9985 SSIM, cutting RMSE by 65.2% and 59.6% relative to UHRNet and HiDNet.

Reconstructing a 3D surface from a single high-frequency fringe image is hard because long-range spatial consistency and fine geometric detail must both survive. Convolutional networks model long-range context weakly, and window-based Transformers build global interactions only gradually with depth. A hybrid Mamba-Transformer network adds visual state-space blocks at the intermediate-resolution encoder stage of a Swin-style U-shaped architecture, with an axial-dilated block that strengthens directional fringe features before selective scanning. On a public fringe-projection dataset, using only one deformed fringe image, it reaches 0.548 mm RMSE and 0.9985 SSIM, cutting RMSE by 65.2% and 59.6% relative to UHRNet and HiDNet.

Asimov has signed a licensing agreement under which South Korea-based LG Chem will in-license Asimov's CHO Edge System to support its biologics cell line development. The system includes the recently announced Chrysalis transposase and Kernel, Asimov's AI-driven genetic design software. The deal puts AI-assisted genetic design and cell line construction tools into the hands of a chemical company with an existing life sciences business, but the two sides did not disclose deal value, exclusivity or the specific development stage.

Asimov has signed a licensing agreement under which South Korea-based LG Chem will in-license Asimov's CHO Edge System to support its biologics cell line development. The system includes the recently announced Chrysalis transposase and Kernel, Asimov's AI-driven genetic design software. The deal puts AI-assisted genetic design and cell line construction tools into the hands of a chemical company with an existing life sciences business, but the two sides did not disclose deal value, exclusivity or the specific development stage.

Insilico Medicine says the abstract of the first-in-human, multicenter Phase 1 study of ISM6331, a pan-TEAD inhibitor designed and optimized with its generative chemistry platform Chemistry42, is now posted on the 2026 ESMO website, with detailed data due October 25 in a Rapid Oral session in Madrid. The study evaluates safety, tolerability, pharmacokinetics and preliminary antitumor activity in mesothelioma and other advanced solid tumors, including dose-limiting toxicities and a recommended Phase 2 dose. What is public now is the abstract and the session schedule; efficacy remains a question for clinical investigation.

Insilico Medicine says the abstract of the first-in-human, multicenter Phase 1 study of ISM6331, a pan-TEAD inhibitor designed and optimized with its generative chemistry platform Chemistry42, is now posted on the 2026 ESMO website, with detailed data due October 25 in a Rapid Oral session in Madrid. The study evaluates safety, tolerability, pharmacokinetics and preliminary antitumor activity in mesothelioma and other advanced solid tumors, including dose-limiting toxicities and a recommended Phase 2 dose. What is public now is the abstract and the session schedule; efficacy remains a question for clinical investigation.

An empirical study tests whether a large language model can write the weights of small scoring networks directly, without gradient training and guided only by forward-pass outputs, to speed up exact algorithms such as column generation and branch-and-bound. On synthetic independent-set instances the written branching rule took 0.37 the time of SCIP's default and matched or beat networks trained by imitation of strong branching on the same features; on a bus-driver scheduling model built from Atlanta transit data, the 0.63 time ratio was fully reproduced by SCIP's own plain pseudocost rule at 0.99.

An empirical study tests whether a large language model can write the weights of small scoring networks directly, without gradient training and guided only by forward-pass outputs, to speed up exact algorithms such as column generation and branch-and-bound. On synthetic independent-set instances the written branching rule took 0.37 the time of SCIP's default and matched or beat networks trained by imitation of strong branching on the same features; on a bus-driver scheduling model built from Atlanta transit data, the 0.63 time ratio was fully reproduced by SCIP's own plain pseudocost rule at 0.99.

Anthropic lowered the API list prices for Opus 5.5, the model used in Claude Code: input and output tokens cost 20% less than on Opus 5, and cache reads cost 60% less, with the read rate dropping from one tenth to one twentieth of the input price. What a task costs depends on turns, cache hits, output tokens and model price, because every turn resends the conversation so far. On an illustrative session with identical token counts, the price change alone takes the bill from $3.50 to $2.40, about 31% less; the stated 40% saving for typical workloads also assumes fewer tokens per task at the medium default.

Anthropic lowered the API list prices for Opus 5.5, the model used in Claude Code: input and output tokens cost 20% less than on Opus 5, and cache reads cost 60% less, with the read rate dropping from one tenth to one twentieth of the input price. What a task costs depends on turns, cache hits, output tokens and model price, because every turn resends the conversation so far. On an illustrative session with identical token counts, the price change alone takes the bill from $3.50 to $2.40, about 31% less; the stated 40% saving for typical workloads also assumes fewer tokens per task at the medium default.

The EEG foundation model REVE no longer needs full fine-tuning or all-layer pooled features to recognize emotion. Such models had not consistently beaten task-specific architectures on cross-subject tasks, and their large representations are costly to compute and store. This approach aggregates the first-layer attention weights of a single head over each trial into a weighted channel-to-channel matrix, then uses a label-free symmetry score to pick the most informative head; on SEED-IV, classical classifiers reach 42.13 balanced accuracy with about 176 times fewer calculations than all-layer attention pooling.

The EEG foundation model REVE no longer needs full fine-tuning or all-layer pooled features to recognize emotion. Such models had not consistently beaten task-specific architectures on cross-subject tasks, and their large representations are costly to compute and store. This approach aggregates the first-layer attention weights of a single head over each trial into a weighted channel-to-channel matrix, then uses a label-free symmetry score to pick the most informative head; on SEED-IV, classical classifiers reach 42.13 balanced accuracy with about 176 times fewer calculations than all-layer attention pooling.

Abu Dhabi's Technology Innovation Institute released Falcon-ASR, a 1.6-billion-parameter speech recognition model built for Arabic with particular attention to the Emirati dialect, and also covering English, French, Spanish and Portuguese. Across six Arabic test sets it reached an average word error rate of 20.92%, below the best published 23.17% in the leaderboard snapshot used for comparison; on an internal Emirati evaluation it recorded 22.73% word error and 10.19% character error, the lowest among the systems compared. The model also emits word-level timestamps that link each transcribed word to its position in the audio.

Abu Dhabi's Technology Innovation Institute released Falcon-ASR, a 1.6-billion-parameter speech recognition model built for Arabic with particular attention to the Emirati dialect, and also covering English, French, Spanish and Portuguese. Across six Arabic test sets it reached an average word error rate of 20.92%, below the best published 23.17% in the leaderboard snapshot used for comparison; on an internal Emirati evaluation it recorded 22.73% word error and 10.19% character error, the lowest among the systems compared. The model also emits word-level timestamps that link each transcribed word to its position in the audio.

Planning microwave ablation needs a 3D temperature field, but solving the coupled electromagnetic-thermal equations point by point is slow. GAPPTANet feeds an analytical specific absorption rate field in as an explicit heat-source prior, uses a geometry-aware adapter to correct applicator-induced near-field approximation errors, and reconstructs global heat diffusion and local high-temperature boundaries with a gated dual-branch decoder. On 27,050 high-fidelity simulations it reaches a full-field root-mean-square error of 0.6001 degrees Celsius and a batch-amortized GPU forward pass of 1.58 ms per sample, outperforming convolutional baselines and three neural-operator architectures under the same test setting.

Planning microwave ablation needs a 3D temperature field, but solving the coupled electromagnetic-thermal equations point by point is slow. GAPPTANet feeds an analytical specific absorption rate field in as an explicit heat-source prior, uses a geometry-aware adapter to correct applicator-induced near-field approximation errors, and reconstructs global heat diffusion and local high-temperature boundaries with a gated dual-branch decoder. On 27,050 high-fidelity simulations it reaches a full-field root-mean-square error of 0.6001 degrees Celsius and a batch-amortized GPU forward pass of 1.58 ms per sample, outperforming convolutional baselines and three neural-operator architectures under the same test setting.

Stablyai released Orca, an open-source app that runs command-line coding agents such as Claude Code and Codex in parallel, giving each agent its own git worktree. Chat and terminal, diff review, a browser, a CLI and phone pairing sit in the same application, which also accepts files or images dragged into an agent's prompt, schedules prompts, and lets agents drive desktop apps when no CLI or API exists. According to the release notes, any CLI agent that runs in a terminal can connect to Orca.

Stablyai released Orca, an open-source app that runs command-line coding agents such as Claude Code and Codex in parallel, giving each agent its own git worktree. Chat and terminal, diff review, a browser, a CLI and phone pairing sit in the same application, which also accepts files or images dragged into an agent's prompt, schedules prompts, and lets agents drive desktop apps when no CLI or API exists. According to the release notes, any CLI agent that runs in a terminal can connect to Orca.

Cross-modality medical image translation has been troubled by anatomically inconsistent synthesis, and standard generative adversarial networks update generator and discriminator together, letting opposing gradients pull on shared representations. StaX-Net interleaves multi-scale computational blocks between the generator and discriminator, lets the discriminator update first while shared blocks are frozen during the generator step, and uses target MRI only during training to guide a feature transform. On the Abdomen/Pelvis, Head and Neck, and Thorax subsets of SynthRAD 2025 it is compared with five baselines, reaching an Abdomen/Pelvis VGG perceptual distance of 642.1, below Pix2Pix at 991.2 and UNet-AE at 1687.9.

Cross-modality medical image translation has been troubled by anatomically inconsistent synthesis, and standard generative adversarial networks update generator and discriminator together, letting opposing gradients pull on shared representations. StaX-Net interleaves multi-scale computational blocks between the generator and discriminator, lets the discriminator update first while shared blocks are frozen during the generator step, and uses target MRI only during training to guide a feature transform. On the Abdomen/Pelvis, Head and Neck, and Thorax subsets of SynthRAD 2025 it is compared with five baselines, reaching an Abdomen/Pelvis VGG perceptual distance of 642.1, below Pix2Pix at 991.2 and UNet-AE at 1687.9.

Moving a wearable inertial sensor to another body location shifts the signal distribution, and measurements at the desired target location are often unavailable during training or deployment. DiffusionRef-IMU first has a deterministic Direct network estimate the target signal, then lets a diffusion model correct only the residual between that estimate and the paired target. Across five subject-disjoint folds with activity–location compositions held out, 16-step DDIM reached a mean nRMSE of 0.087538, below 0.090262 for Direct, 0.088847 for a capacity-matched deterministic residual model and 0.088417 for eight-step CompFlow, and was better in all five folds.

Moving a wearable inertial sensor to another body location shifts the signal distribution, and measurements at the desired target location are often unavailable during training or deployment. DiffusionRef-IMU first has a deterministic Direct network estimate the target signal, then lets a diffusion model correct only the residual between that estimate and the paired target. Across five subject-disjoint folds with activity–location compositions held out, 16-step DDIM reached a mean nRMSE of 0.087538, below 0.090262 for Direct, 0.088847 for a capacity-matched deterministic residual model and 0.088417 for eight-step CompFlow, and was better in all five folds.

anomalyco released OpenCode, an open-source AI coding agent available through an install script, npm, bun, pnpm, yarn, Homebrew, Arch Linux's pacman and paru, mise and Nix, plus a desktop application. It ships two built-in agents that switch with the Tab key: build, the default full-access agent for development work, and plan, a read-only agent that denies file edits by default and asks permission before running bash commands, suited to exploring unfamiliar codebases or planning changes. A general subagent, invoked with @general in messages, handles complex searches and multistep tasks.

anomalyco released OpenCode, an open-source AI coding agent available through an install script, npm, bun, pnpm, yarn, Homebrew, Arch Linux's pacman and paru, mise and Nix, plus a desktop application. It ships two built-in agents that switch with the Tab key: build, the default full-access agent for development work, and plan, a read-only agent that denies file edits by default and asks permission before running bash commands, suited to exploring unfamiliar codebases or planning changes. A general subagent, invoked with @general in messages, handles complex searches and multistep tasks.

AIPOCH's local-first research workbench Open-Science released v1.2.0, letting a session submit approved jobs to Kubernetes, Hugging Face Jobs, Modal and Ray, where remote compute previously covered only SSH and Slurm. Each external submission freezes its inputs and asks for approval of target, command, resources and outputs, and recovery records let result collection resume after a restart. A built-in Open-Science Agent becomes the default for new installs, and an imported .science package or replay step can carry a child discussion without modifying the import.

AIPOCH's local-first research workbench Open-Science released v1.2.0, letting a session submit approved jobs to Kubernetes, Hugging Face Jobs, Modal and Ray, where remote compute previously covered only SSH and Slurm. Each external submission freezes its inputs and asks for approval of target, command, resources and outputs, and recovery records let result collection resume after a restart. A built-in Open-Science Agent becomes the default for new installs, and an imported .science package or replay step can carry a child discussion without modifying the import.