AI and science frontiers · 2026-10-19

A 123.5-million-parameter generative model, Didymos, was trained once without task labels on 409,297 intensive care stays from seven databases in five countries and answers prediction questions by sampling a patient's futures. On held-out stays it read 13 primary endpoints from those futures at AUROC 0.83–0.98 and ranked first on average among 20 predictors, including label-trained LightGBM and TabPFN-3.5. Its simulated treatment responses agreed with eight of nine observational contrasts, yet at two hospitals in Russia and China that contributed no training data it needed minutes of continued training on unlabelled local stays to match locally supervised models.

A 123.5-million-parameter generative model, Didymos, was trained once without task labels on 409,297 intensive care stays from seven databases in five countries and answers prediction questions by sampling a patient's futures. On held-out stays it read 13 primary endpoints from those futures at AUROC 0.83–0.98 and ranked first on average among 20 predictors, including label-trained LightGBM and TabPFN-3.5. Its simulated treatment responses agreed with eight of nine observational contrasts, yet at two hospitals in Russia and China that contributed no training data it needed minutes of continued training on unlabelled local stays to match locally supervised models.

When honeybees choose a nest site, individuals blindly switch to the option advertised most strongly by neighbors, and this local imitation is shown to be equivalent, at the group level, to a single online reinforcement-learning agent solving a multi-armed bandit. That macro-agent updates nest-site probabilities by a rule named Maynard-Cross Learning, with each scout acting as one parallel action-reward sample. The derivation covers the weighted voter model, a simplified description, and simulations indicate roughly 500 scouts suffice to track the macro-dynamics and converge on the best site.

When honeybees choose a nest site, individuals blindly switch to the option advertised most strongly by neighbors, and this local imitation is shown to be equivalent, at the group level, to a single online reinforcement-learning agent solving a multi-armed bandit. That macro-agent updates nest-site probabilities by a rule named Maynard-Cross Learning, with each scout acting as one parallel action-reward sample. The derivation covers the weighted voter model, a simplified description, and simulations indicate roughly 500 scouts suffice to track the macro-dynamics and converge on the best site.

In sparse mixture-of-experts models, a router that favors a few experts starves the rest of training signal and makes devices wait under expert parallelism. Kolibri pre-training pairs Exact Quantile Balancing (EQB), which sets routing biases from exact global-batch quantiles via two-pass radix selection, with bounded Load-Error Injection (LEI), which adds each microbatch's load error directly to router gradients. On a 7.5B-parameter proxy trained for 100B tokens, EQB cut global maximal violation from 0.92 to 0.74, and adding bounded LEI cut local violation from 5.38 to 3.52 at a mean BPB cost from 0.6570 to 0.6720.

In sparse mixture-of-experts models, a router that favors a few experts starves the rest of training signal and makes devices wait under expert parallelism. Kolibri pre-training pairs Exact Quantile Balancing (EQB), which sets routing biases from exact global-batch quantiles via two-pass radix selection, with bounded Load-Error Injection (LEI), which adds each microbatch's load error directly to router gradients. On a 7.5B-parameter proxy trained for 100B tokens, EQB cut global maximal violation from 0.92 to 0.74, and adding bounded LEI cut local violation from 5.38 to 3.52 at a mean BPB cost from 0.6570 to 0.6720.

D3WM splits next-frame prediction into a deterministic part and an uncertain part, using a learned uncertainty map to steer a diffusion model toward occlusions and dynamic textures, and it beats comparison world models on both robotic manipulation and camera motion.

D3WM splits next-frame prediction into a deterministic part and an uncertain part, using a learned uncertainty map to steer a diffusion model toward occlusions and dynamic textures, and it beats comparison world models on both robotic manipulation and camera motion.

Diffusion sampling accumulates the model's own prediction errors, and a corrector trained on the base model's predictions against the same clean target lowers DiT-XL/2 FID on ImageNet class-conditional generation from 9.58 to 2.07, approaching CFG without guidance.

Diffusion sampling accumulates the model's own prediction errors, and a corrector trained on the base model's predictions against the same clean target lowers DiT-XL/2 FID on ImageNet class-conditional generation from 9.58 to 2.07, approaching CFG without guidance.

With architecture and data held fixed, a checkpoint-only reconstruction attack recovers training images best from Muon-trained models and worst from Adam-trained models, making the optimizer choice a control axis for training-data privacy.

With architecture and data held fixed, a checkpoint-only reconstruction attack recovers training images best from Muon-trained models and worst from Adam-trained models, making the optimizer choice a control axis for training-data privacy.

AffectWorld evaluates world models with emotion treated as a causal variable driving future world change, and five state-of-the-art models reach only 0.100–0.280 on world-state change and 0.027–0.070 on counterfactual emotion sensitivity across 1,498 everyday videos.

AffectWorld evaluates world models with emotion treated as a causal variable driving future world change, and five state-of-the-art models reach only 0.100–0.280 on world-state change and 0.027–0.070 on counterfactual emotion sensitivity across 1,498 everyday videos.

Across the same 60 change requests and one fixed model, two architectures finished within 7% to 20% on complexity amount but differed by 2.7 to 10 times on how far their placement could be moved, and only an AI reviewer moved the additive one.

Across the same 60 change requests and one fixed model, two architectures finished within 7% to 20% on complexity amount but differed by 2.7 to 10 times on how far their placement could be moved, and only an AI reviewer moved the additive one.

Reflection, a new US lab, released Beam, its first open model: 501B sparse mixture-of-experts weights under Apache 2.0, with about 23B parameters active per token, a 1M-token context, and a focus on coding and agentic work. Hybrid attention keeps only 9 of 52 layers global while 43 hold a fixed 512-token window, so the KV cache grows about 36 KiB per token, roughly one-sixth of the all-global rate. BF16, FP8 and NVFP4 checkpoints ship alongside day-0 transformers, SGLang and vLLM support and a Rust-based harness.

Reflection, a new US lab, released Beam, its first open model: 501B sparse mixture-of-experts weights under Apache 2.0, with about 23B parameters active per token, a 1M-token context, and a focus on coding and agentic work. Hybrid attention keeps only 9 of 52 layers global while 43 hold a fixed 512-token window, so the KV cache grows about 36 KiB per token, roughly one-sixth of the all-global rate. BF16, FP8 and NVFP4 checkpoints ship alongside day-0 transformers, SGLang and vLLM support and a Rust-based harness.

Each signaling pathway leaves a consistent fingerprint of gene activity across cell types, so signaling history no longer has to be reconstructed separately for every cell type. IRIS is a neural-network model that reads a cell's overall gene activity and estimates which pathways were switched on at each developmental stage. Tested on single cells from mouse embryos during gastrulation, it predicted when and where pathways activate in heart, gut, muscle and spinal cord lineages, and its prediction for a lung-development cell type was confirmed in mouse embryo experiments.

Each signaling pathway leaves a consistent fingerprint of gene activity across cell types, so signaling history no longer has to be reconstructed separately for every cell type. IRIS is a neural-network model that reads a cell's overall gene activity and estimates which pathways were switched on at each developmental stage. Tested on single cells from mouse embryos during gastrulation, it predicted when and where pathways activate in heart, gut, muscle and spinal cord lineages, and its prediction for a lung-development cell type was confirmed in mouse embryo experiments.

Eight speech LLMs mostly recognize a speaker's age and gender, yet rarely turn that perception into requested personalization, and adding explicit demographic labels raises personalization while also increasing stereotypical choices.

Eight speech LLMs mostly recognize a speaker's age and gender, yet rarely turn that perception into requested personalization, and adding explicit demographic labels raises personalization while also increasing stereotypical choices.

Long-chain reasoning distillation no longer fixes its training questions up front: BoundaryEvo repeatedly has the student and a fixed expert answer the same candidates, keeping only questions the expert solves reliably and the student has not mastered, and with 1,000 training examples raises Qwen3.5-4B accuracy on IMO-AnswerBench from 62.9 to 75.3.

Long-chain reasoning distillation no longer fixes its training questions up front: BoundaryEvo repeatedly has the student and a fixed expert answer the same candidates, keeping only questions the expert solves reliably and the student has not mastered, and with 1,000 training examples raises Qwen3.5-4B accuracy on IMO-AnswerBench from 62.9 to 75.3.

In federated LLM fine-tuning, a server using the strong Byzantine-robust aggregator RandEigen filters out existing text-watermark updates as outliers, but optimizing the watermarked samples' input embeddings to sit close to clean updates keeps the watermark detectable by black-box access from a single client's 3.3% effective watermarked data.

In federated LLM fine-tuning, a server using the strong Byzantine-robust aggregator RandEigen filters out existing text-watermark updates as outliers, but optimizing the watermarked samples' input embeddings to sit close to clean updates keeps the watermark detectable by black-box access from a single client's 3.3% effective watermarked data.

A controlled from-scratch comparison shows that adding long skip connections to a DiT improves its intermediate features across semantic correspondence, segmentation and depth, and the gain survives even when generation quality is matched by FID.

A controlled from-scratch comparison shows that adding long skip connections to a DiT improves its intermediate features across semantic correspondence, segmentation and depth, and the gain survives even when generation quality is matched by FID.

Supervised fine-tuning of coding agents no longer has to select only on whether the final patch passes the test suite: ranking trajectories by structure and keeping the top 1,024 raised pass@1 on SWE-Bench Verified from 24.8% for a random subset to 27.6%.

Supervised fine-tuning of coding agents no longer has to select only on whether the final patch passes the test suite: ranking trajectories by structure and keeping the top 1,024 raised pass@1 on SWE-Bench Verified from 24.8% for a random subset to 27.6%.

Across four 24–32B open-weight models, hidden states at event ends decode the state-change type of at least four upcoming events, yet only the next event significantly beats text-only predictors; in OLMo, sustained steering along that direction shifts subsequent event composition in both directions.

Across four 24–32B open-weight models, hidden states at event ends decode the state-change type of at least four upcoming events, yet only the next event significantly beats text-only predictors; in OLMo, sustained steering along that direction shifts subsequent event composition in both directions.

Instead of a fixed intervention chosen in advance, Q-NeuS reads each query's own internal activations to decide which feed-forward neurons to amplify, improving every one of seven models from 3B to 32B parameters across three benchmarks.

Instead of a fixed intervention chosen in advance, Q-NeuS reads each query's own internal activations to decide which feed-forward neurons to amplify, improving every one of seven models from 3B to 32B parameters across three benchmarks.

Anthropic's Claude Code team moved most code reading to automated review and kept humans for high-risk decisions. From June to September, merged pull requests to the Claude Code codebase more than doubled while the share that had to be reverted fell by more than 40 percent. Claude verifies changes before submission with runnable tests or builds, every pull request gets a multi-agent bug review with voting filters, and risk tiers set how many human approvals a change needs.

Anthropic's Claude Code team moved most code reading to automated review and kept humans for high-risk decisions. From June to September, merged pull requests to the Claude Code codebase more than doubled while the share that had to be reverted fell by more than 40 percent. Claude verifies changes before submission with runnable tests or builds, every pull request gets a multi-agent bug review with voting filters, and risk tiers set how many human approvals a change needs.

Models of cross-region neural dynamics have largely stayed linear, while each region's recorded activity mixes local dynamics with dynamics shared across regions. CroP-DYNO sets learning directly on predicting target-population activity from source activity, so shared dynamics are extracted first, and it lets each model component be specified as linear or nonlinear. On macaque S1–M1 and rat CA1–CA3 data it predicts target activity better than a non-prioritized nonlinear model and a prioritized linear model, and it localizes nonlinear structure and yields interaction strengths consistent with known circuit anatomy.

Models of cross-region neural dynamics have largely stayed linear, while each region's recorded activity mixes local dynamics with dynamics shared across regions. CroP-DYNO sets learning directly on predicting target-population activity from source activity, so shared dynamics are extracted first, and it lets each model component be specified as linear or nonlinear. On macaque S1–M1 and rat CA1–CA3 data it predicts target activity better than a non-prioritized nonlinear model and a prioritized linear model, and it localizes nonlinear structure and yields interaction strengths consistent with known circuit anatomy.

Automatic vertebra labeling has long assumed twelve thoracic and five lumbar vertebrae, so a spine with one extra or one missing segment shifts every label. VERIDAH feeds per-vertebra classification outputs into an anatomy-constrained sequence solver that picks the most consistent label sequence at any field of view. Across 1000 T2w sagittal MRI and 365 CT images, all vertebrae were labeled correctly in 98.30% and 99.18% of subjects, versus 94.24% and 77.26% for comparison methods; thoracic and lumbar enumeration anomalies were identified in 87.80% and 94.48% of MRI cases and 96.30% and 97.22% of CT cases.

Automatic vertebra labeling has long assumed twelve thoracic and five lumbar vertebrae, so a spine with one extra or one missing segment shifts every label. VERIDAH feeds per-vertebra classification outputs into an anatomy-constrained sequence solver that picks the most consistent label sequence at any field of view. Across 1000 T2w sagittal MRI and 365 CT images, all vertebrae were labeled correctly in 98.30% and 99.18% of subjects, versus 94.24% and 77.26% for comparison methods; thoracic and lumbar enumeration anomalies were identified in 87.80% and 94.48% of MRI cases and 96.30% and 97.22% of CT cases.

E-commerce ad image optimization moves from regenerating the whole creative to fine-grained attribute editing: 28 click-sensitive visual attributes are selected, editing suggestions are aligned to real click preferences, and a one-week online test over roughly 20 million impressions lifts aggregate CTR by 6.86%.

E-commerce ad image optimization moves from regenerating the whole creative to fine-grained attribute editing: 28 click-sensitive visual attributes are selected, editing suggestions are aligned to real click preferences, and a one-week online test over roughly 20 million impressions lifts aggregate CTR by 6.86%.

Treating natural language as a temporal point process of information-bearing events lets a neural model predict when the next new word or high-surprisal token will arrive, and shows LLMs introducing new words at consistently higher rates than humans.

Treating natural language as a temporal point process of information-bearing events lets a neural model predict when the next new word or high-surprisal token will arrive, and shows LLMs introducing new words at consistently higher rates than humans.

Multi-agent memory is reorganized around task dependency graphs: a shared blackboard records evidence during execution and corrects goal, plan, and world-state divergence, then verified steps and dependencies are distilled into reusable procedures, giving the highest task score in all six benchmark settings across ALFWorld, WebArena, and OfficeBench.

Multi-agent memory is reorganized around task dependency graphs: a shared blackboard records evidence during execution and corrects goal, plan, and world-state divergence, then verified steps and dependencies are distilled into reusable procedures, giving the highest task score in all six benchmark settings across ALFWorld, WebArena, and OfficeBench.

TiCO replaces coordinate regression in thinking-with-images with a probe token that samples evidence over visual tokens, reaching 78.0% overall accuracy on HR-Bench 4K while generating only 6.2 tokens per query.

TiCO replaces coordinate regression in thinking-with-images with a probe token that samples evidence over visual tokens, reaching 78.0% overall accuracy on HR-Bench 4K while generating only 6.2 tokens per query.

A 500-scenario zero-shot benchmark built from real customer-support transcripts scores 17 models with five independent judges, and GPT-5.4 ranks first under every judge while judges agree only moderately on pass/fail.

A 500-scenario zero-shot benchmark built from real customer-support transcripts scores 17 models with five independent judges, and GPT-5.4 ranks first under every judge while judges agree only moderately on pass/fail.

ASSET trains a speech encoder's segment boundaries with reinforcement learning, reaching recognition close to a transcript-assisted character-boundary oracle on LibriSpeech with roughly 25% fewer audio tokens.

ASSET trains a speech encoder's segment boundaries with reinforcement learning, reaching recognition close to a transcript-assisted character-boundary oracle on LibriSpeech with roughly 25% fewer audio tokens.

NVIDIA has pushed the cost of bitwise-deterministic pretraining in Megatron Core down to a workable range: two independent runs from the same initial state, and a run resumed from checkpoint, must stay numerically identical at every step. Such deterministic recipes previously carried double-digit overhead; after buffer-fill optimization Nemotron 3 Ultra fell from about 17% to 2%, then to 1.5% at 3,072 GPUs, and the large-scale Nemotron recipe measured about 2% steady-state overhead at 2,432 GPUs with bitwise determinism over 800 steps.

NVIDIA has pushed the cost of bitwise-deterministic pretraining in Megatron Core down to a workable range: two independent runs from the same initial state, and a run resumed from checkpoint, must stay numerically identical at every step. Such deterministic recipes previously carried double-digit overhead; after buffer-fill optimization Nemotron 3 Ultra fell from about 17% to 2%, then to 1.5% at 3,072 GPUs, and the large-scale Nemotron recipe measured about 2% steady-state overhead at 2,432 GPUs with bitwise determinism over 800 steps.

In long-running multi-party chats, the same issue is proposed, revised and overturned repeatedly, so semantically similar messages can describe conflicting states. T-Mem treats temporal structure as the core memory representation: retrieved messages act as temporal anchors, the discussion episodes where they cluster are reopened to recover evidence that global ranking missed, and the recovered messages are organized into a query-conditioned evidence forest with Update and Expand relations before low-information spans are compressed. On GroupMemBench and EverMemBench it reaches the highest overall accuracy under both GPT-5.5 and Deepseek-V4-Flash, beating the strongest baseline by 6.50 to 8.59 percentage points.

In long-running multi-party chats, the same issue is proposed, revised and overturned repeatedly, so semantically similar messages can describe conflicting states. T-Mem treats temporal structure as the core memory representation: retrieved messages act as temporal anchors, the discussion episodes where they cluster are reopened to recover evidence that global ranking missed, and the recovered messages are organized into a query-conditioned evidence forest with Update and Expand relations before low-information spans are compressed. On GroupMemBench and EverMemBench it reaches the highest overall accuracy under both GPT-5.5 and Deepseek-V4-Flash, beating the strongest baseline by 6.50 to 8.59 percentage points.

The decision-making part of a Maven-style build tool has been compiled into a fixed weight matrix and used to build ZXing 3.5.3 for real, producing JARs byte-identical to Apache Maven's with all 549 upstream tests passing. The weights come from compiling a program, not from training or gradient descent; the price is speed, with a clean build including tests taking about 1.42 times as long as the same logic in Python.

The decision-making part of a Maven-style build tool has been compiled into a fixed weight matrix and used to build ZXing 3.5.3 for real, producing JARs byte-identical to Apache Maven's with all 549 upstream tests passing. The weights come from compiling a program, not from training or gradient descent; the price is speed, with a clean build including tests taking about 1.42 times as long as the same logic in Python.

A two-stage learning framework turns obstacle point clouds inside the onboard camera frustum into a temporal polar gradient field that no longer separates static from dynamic obstacles, raising success rate by 16 percentage points over the best baseline in dense dynamic scenes while keeping replanning under 4.8 ms.

A two-stage learning framework turns obstacle point clouds inside the onboard camera frustum into a temporal polar gradient field that no longer separates static from dynamic obstacles, raising success rate by 16 percentage points over the best baseline in dense dynamic scenes while keeping replanning under 4.8 ms.

Spatial transcriptomics analysis has long clustered at a single scale, so cells, neighboring cell groups and tissue regions were handled separately. SpatialZoomer applies a set of spectral graph low-pass filters to extract spatial molecular features at the single-cell, niche and domain scales from one dataset, and locates critical scales by partitioning a cross-scale similarity map. In Alzheimer's disease and tumor microenvironment data it identified disease-progression signals at specific scales and spatial context-dependent cell subtypes, matching or exceeding existing methods on single-scale tasks with lower compute and hardware demands.

Spatial transcriptomics analysis has long clustered at a single scale, so cells, neighboring cell groups and tissue regions were handled separately. SpatialZoomer applies a set of spectral graph low-pass filters to extract spatial molecular features at the single-cell, niche and domain scales from one dataset, and locates critical scales by partitioning a cross-scale similarity map. In Alzheimer's disease and tumor microenvironment data it identified disease-progression signals at specific scales and spatial context-dependent cell subtypes, matching or exceeding existing methods on single-scale tasks with lower compute and hardware demands.

In a Mesa implementation of the Schelling segregation model, one agent's symbolic rule for classifying neighbors is replaced by a locally served Qwen3.5 model that classifies through tool calls, while the rest of the dynamics stays fixed. Statistical model checking with MultiVeStA shows the expected ratio of happy agents rising from about 0.65 to about 1.0 within 60 steps in both variants, with point-wise t-tests finding no difference, yet the LLM version takes 8,917 seconds against about 2 seconds. The 0.8b and 2b models show semantic errors, wrong tool-call counts or stuck generation in unit-style checks, while 4b and 9b pass all of them.

In a Mesa implementation of the Schelling segregation model, one agent's symbolic rule for classifying neighbors is replaced by a locally served Qwen3.5 model that classifies through tool calls, while the rest of the dynamics stays fixed. Statistical model checking with MultiVeStA shows the expected ratio of happy agents rising from about 0.65 to about 1.0 within 60 steps in both variants, with point-wise t-tests finding no difference, yet the LLM version takes 8,917 seconds against about 2 seconds. The 0.8b and 2b models show semantic errors, wrong tool-call counts or stuck generation in unit-style checks, while 4b and 9b pass all of them.

DeepSeek released DeepSeek-V4.1-Flash-1017, which it says delivers stronger benchmark performance, faster task completion and shorter outputs than the previous version, a Pareto improvement in performance and cost, now live on its website, app and API. The model uses reinforcement learning with self-feedback to improve end-to-end project development, with demos spanning 3D modeling, a web racing game and a Godot-engine recreation. The Search and Fetch APIs also opened to everyone at $0.75 per 1,000 calls per endpoint.

DeepSeek released DeepSeek-V4.1-Flash-1017, which it says delivers stronger benchmark performance, faster task completion and shorter outputs than the previous version, a Pareto improvement in performance and cost, now live on its website, app and API. The model uses reinforcement learning with self-feedback to improve end-to-end project development, with demos spanning 3D modeling, a web racing game and a Godot-engine recreation. The Search and Fetch APIs also opened to everyone at $0.75 per 1,000 calls per endpoint.

Retinal disease grading has mostly relied on raw color fundus photographs, where low-contrast lesions and thin vessels are easily lost to illumination and noise. F2D-RetinaNet sends an enhanced green channel and a vessel map through two separate branches and fuses them into one grading decision. On APTOS 2019 five-grade diabetic retinopathy staging it reaches 93.32% accuracy, 3.25 points above the RGB-only baseline, and 87.14% on the external DREAM-RFI set after overlapping images were excluded.

Retinal disease grading has mostly relied on raw color fundus photographs, where low-contrast lesions and thin vessels are easily lost to illumination and noise. F2D-RetinaNet sends an enhanced green channel and a vessel map through two separate branches and fuses them into one grading decision. On APTOS 2019 five-grade diabetic retinopathy staging it reaches 93.32% accuracy, 3.25 points above the RGB-only baseline, and 87.14% on the external DREAM-RFI set after overlapping images were excluded.

Simulation of stochastic hybrid systems previously handled only discrete nondeterministic choices, while an action allowed to fire anywhere in a time interval gives a single state infinitely many options. The new method discretizes each action's time interval separately, learns a scheduler with Q-learning over the finite abstraction, and uses precomputed symbolic reachable states to steer simulation and prune unreachable regions in a correct-by-construction way. On the water tank benchmark the simulation reproduces the maximum reachability probabilities reported by an existing method; on a water tank variant with a low-probability underflow objective, guidance makes learning clearly faster than unguided simulation and yields a reachability probability of at least 45%.

Simulation of stochastic hybrid systems previously handled only discrete nondeterministic choices, while an action allowed to fire anywhere in a time interval gives a single state infinitely many options. The new method discretizes each action's time interval separately, learns a scheduler with Q-learning over the finite abstraction, and uses precomputed symbolic reachable states to steer simulation and prune unreachable regions in a correct-by-construction way. On the water tank benchmark the simulation reproduces the maximum reachability probabilities reported by an existing method; on a water tank variant with a low-probability underflow objective, guidance makes learning clearly faster than unguided simulation and yields a reachability probability of at least 45%.

Instead of drawing on a fixed command library, a quadruped exploring unknown space hands the user's high-level intention to a large language model that writes an executable program restricted to a robot function library, while a vision-language model supplies open-ended semantic labels and geometric and semantic thresholds trigger replanning. In Gazebo simulation, Qwen3.5 and Kimi-K2.5 reach about 82% overall success and GPT-5-nano about 46%; finite tasks hold up well, while open-ended tasks vary more.

Instead of drawing on a fixed command library, a quadruped exploring unknown space hands the user's high-level intention to a large language model that writes an executable program restricted to a robot function library, while a vision-language model supplies open-ended semantic labels and geometric and semantic thresholds trigger replanning. In Gazebo simulation, Qwen3.5 and Kimi-K2.5 reach about 82% overall success and GPT-5-nano about 46%; finite tasks hold up well, while open-ended tasks vary more.

nanoHUB has released MCP4MD, which connects LAMMPS and the UMA universal machine-learned potential to an AI assistant, so a user can describe a material property in natural language and have the calculation set up, run and read back. Previously such work meant installing software, writing input files and submitting queue scripts. Every number names the potential behind it and the reference it was fitted to, and arrives with error bars, units and qualifying warnings.

nanoHUB has released MCP4MD, which connects LAMMPS and the UMA universal machine-learned potential to an AI assistant, so a user can describe a material property in natural language and have the calculation set up, run and read back. Previously such work meant installing software, writing input files and submitting queue scripts. Every number names the potential behind it and the reference it was fitted to, and arrives with error bars, units and qualifying warnings.

Single-frame fringe pattern analysis in digital holographic interferometry has relied on manually set Fourier-domain filter windows, which are hard to choose when noise, defocus or overexposure degrade the fringes. An adaptive spectrum filtering network turns Fourier-domain filtering into learnable attention weights that select +1st-order spectral content per input, and is trained on forward-model simulations with added noise and overexposure. On 400 simulated test samples it beats tuned Fourier-transform analysis and an end-to-end ResUNet in pixel error, and after simulation-only training it produces fewer artifacts and outliers than hand-tuned Fourier analysis on unseen experimental interferograms.

Single-frame fringe pattern analysis in digital holographic interferometry has relied on manually set Fourier-domain filter windows, which are hard to choose when noise, defocus or overexposure degrade the fringes. An adaptive spectrum filtering network turns Fourier-domain filtering into learnable attention weights that select +1st-order spectral content per input, and is trained on forward-model simulations with added noise and overexposure. On 400 simulated test samples it beats tuned Fourier-transform analysis and an end-to-end ResUNet in pixel error, and after simulation-only training it produces fewer artifacts and outliers than hand-tuned Fourier analysis on unseen experimental interferograms.

Netflix ran JetBrains' code-indexing engine, the one behind IntelliJ IDEA, headlessly inside Claude Code and Codex so coding agents can query Java/Kotlin symbols and library structure instead of unzipping JARs, decompiling and grepping. On two real codebases, with the same model and task, enabling the engine typically cut token use and completion time by 20–30%, up to about 65% in the best configuration, without lowering correctness.

Netflix ran JetBrains' code-indexing engine, the one behind IntelliJ IDEA, headlessly inside Claude Code and Codex so coding agents can query Java/Kotlin symbols and library structure instead of unzipping JARs, decompiling and grepping. On two real codebases, with the same model and task, enabling the engine typically cut token use and completion time by 20–30%, up to about 65% in the best configuration, without lowering correctness.

ChatGPT, Claude, Gemini and Grok each generated PyTorch helmet-compliance training scripts from three prompt tiers of increasing specificity, run 20 times on 972 balanced images: ten of twelve configurations converged to a 0.957–0.967 accuracy plateau, but the most detailed tier left Gemini and Grok near 0.58, with scripts that ran yet failed to converge. A backbone-by-scheduler ablation traced the failure to the prescribed from-scratch multi-scale network, which collapsed under both schedulers while a pretrained ResNet18 converged under either.

ChatGPT, Claude, Gemini and Grok each generated PyTorch helmet-compliance training scripts from three prompt tiers of increasing specificity, run 20 times on 972 balanced images: ten of twelve configurations converged to a 0.957–0.967 accuracy plateau, but the most detailed tier left Gemini and Grok near 0.58, with scripts that ran yet failed to converge. A backbone-by-scheduler ablation traced the failure to the prescribed from-scratch multi-scale network, which collapsed under both schedulers while a pretrained ResNet18 converged under either.

Metabolomics models trained on one cohort often lose accuracy on another because platform, sample handling and quantification differences shift the data. Using 12,000 plasma NMR samples from the Estonian Biobank to classify heart failure (ICD-10 I50.9), a study evaluated 295 samples re-measured on a different NMR platform and calibrated alignment with up to 108 biospecimens duplicated across both. Harmonization generally beat unaligned data and stabilized as matched pairs grew, yet simple within-domain normalization stayed a strong baseline, reaching balanced accuracy 0.72 and MCC 0.46 with a support vector machine at all 108 pairs.

Metabolomics models trained on one cohort often lose accuracy on another because platform, sample handling and quantification differences shift the data. Using 12,000 plasma NMR samples from the Estonian Biobank to classify heart failure (ICD-10 I50.9), a study evaluated 295 samples re-measured on a different NMR platform and calibrated alignment with up to 108 biospecimens duplicated across both. Harmonization generally beat unaligned data and stabilized as matched pairs grew, yet simple within-domain normalization stayed a strong baseline, reaching balanced accuracy 0.72 and MCC 0.46 with a support vector machine at all 108 pairs.

Abdominal ultrasound simulation has largely relied on patient CT for ray-casting, which limits control over deformation and pathology. A two-stage pipeline first synthesizes a simplified physics-based ultrasound image from anatomical semantic labels, then refines it into a realistic scan with anatomy-guided unpaired translation, so no patient-specific CT volume is needed at inference. In stage one the semantic diffusion model beat Pix2Pix on morphological metrics while Pix2Pix scored better on perceptual ones; training still requires CT-derived segmentations and ray-casting simulations.

Abdominal ultrasound simulation has largely relied on patient CT for ray-casting, which limits control over deformation and pathology. A two-stage pipeline first synthesizes a simplified physics-based ultrasound image from anatomical semantic labels, then refines it into a realistic scan with anatomy-guided unpaired translation, so no patient-specific CT volume is needed at inference. In stage one the semantic diffusion model beat Pix2Pix on morphological metrics while Pix2Pix scored better on perceptual ones; training still requires CT-derived segmentations and ray-casting simulations.

Single-layer diffractive neural networks hit an accuracy ceiling because their wavefront modulation is limited. Edge detection and classification are now combined on one nonlocal metasurface through spin multiplexing: the co-polarized output component performs momentum-space filtering for real-time edge detection, while the cross-polarized component is geometrically phase-modulated to classify. Crescent-shaped nanopillars couple quasi-bound states in the continuum with magnetic dipole resonances, reaching about 55% polarization conversion efficiency. On MNIST, this edge-enhanced architecture raises single-layer accuracy from 64.2% to 80.7%.

Single-layer diffractive neural networks hit an accuracy ceiling because their wavefront modulation is limited. Edge detection and classification are now combined on one nonlocal metasurface through spin multiplexing: the co-polarized output component performs momentum-space filtering for real-time edge detection, while the cross-polarized component is geometrically phase-modulated to classify. Crescent-shaped nanopillars couple quasi-bound states in the continuum with magnetic dipole resonances, reaching about 55% polarization conversion efficiency. On MNIST, this edge-enhanced architecture raises single-layer accuracy from 64.2% to 80.7%.

Pricing American options under the double Heston model is hard because the early exercise boundary is itself unknown and must be solved together with the option value. A deep learning scheme that multiplies the outputs of several sub-networks fixes that boundary first, then predicts the early exercise boundary, the value function and first- and second-order Greeks in a single training run. Under constant and stochastic interest rates, its value functions are compared numerically with asymptotic expansion and least-squares Monte Carlo results.

Pricing American options under the double Heston model is hard because the early exercise boundary is itself unknown and must be solved together with the option value. A deep learning scheme that multiplies the outputs of several sub-networks fixes that boundary first, then predicts the early exercise boundary, the value function and first- and second-order Greeks in a single training run. Under constant and stochastic interest rates, its value functions are compared numerically with asymptotic expansion and least-squares Monte Carlo results.

Reflection AI's Beam reasoning model has left stealth testing under the codename stealth/cat-jam and is now available on AI Gateway under its own name, free for a limited launch period when served by Reflection AI. It targets coding and agentic tasks, ships weights under Apache 2.0, supports tool calling and structured outputs, and carries a 1M-token context window. The model has 501B total parameters but activates only 23B per token and skips tokens a task does not need, so it tends to answer sooner than larger open models.

Reflection AI's Beam reasoning model has left stealth testing under the codename stealth/cat-jam and is now available on AI Gateway under its own name, free for a limited launch period when served by Reflection AI. It targets coding and agentic tasks, ships weights under Apache 2.0, supports tool calling and structured outputs, and carries a 1M-token context window. The model has 501B total parameters but activates only 23B per token and skips tokens a task does not need, so it tends to answer sooner than larger open models.

A multi-agent workflow under investigator direction carried a large electronic health record study from cohort construction and executable statistics through data-quality checks, literature retrieval and reporting, using sequential sacituzumab govitecan (SG) and trastuzumab deruxtecan (T-DXd) orders in breast cancer as the clinical case. From 4.95 million patients, 106 had orders for both drugs, 73 with T-DXd first and 33 with SG first; median second-drug record spans were 63 and 49 days, a median difference of −14 days (95% CI, −41 to 21), descriptive and unadjusted.

A multi-agent workflow under investigator direction carried a large electronic health record study from cohort construction and executable statistics through data-quality checks, literature retrieval and reporting, using sequential sacituzumab govitecan (SG) and trastuzumab deruxtecan (T-DXd) orders in breast cancer as the clinical case. From 4.95 million patients, 106 had orders for both drugs, 73 with T-DXd first and 33 with SG first; median second-drug record spans were 63 and 49 days, a median difference of −14 days (95% CI, −41 to 21), descriptive and unadjusted.

Elorian's visual reasoning model Glimpse is now available on AI Gateway, accepting images and video to follow objects through motion, interpret spatial relationships, and connect an action to its physical consequences. While working, it can return to the visual evidence itself by zooming into a detail, revisiting a short interval of video, or tracking the same feature across frames, supporting uses such as scientific chart analysis, video captioning, visual counting, and robotics data annotation.

Elorian's visual reasoning model Glimpse is now available on AI Gateway, accepting images and video to follow objects through motion, interpret spatial relationships, and connect an action to its physical consequences. While working, it can return to the visual evidence itself by zooming into a detail, revisiting a short interval of video, or tracking the same feature across frames, supporting uses such as scientific chart analysis, video captioning, visual counting, and robotics data annotation.

The open-source AI coding agent OpenCode ships two built-in agents: build, the default with full development access, and plan, which denies file edits by default and asks permission before running bash commands, suited to exploring unfamiliar codebases or planning changes. A general subagent handles complex searches and multistep tasks and can be invoked with @general in a message. Installation covers a shell script, npm, Homebrew, pacman, mise and Nix, plus a desktop app download.

The open-source AI coding agent OpenCode ships two built-in agents: build, the default with full development access, and plan, which denies file edits by default and asks permission before running bash commands, suited to exploring unfamiliar codebases or planning changes. A general subagent handles complex searches and multistep tasks and can be invoked with @general in a message. Installation covers a shell script, npm, Homebrew, pacman, mise and Nix, plus a desktop app download.