AI and science frontiers · 2026-11-08

Lila Sciences used a closed-loop autonomous laboratory to search for acidic oxygen evolution reaction (OER) catalysts, a reaction long dependent on scarce iridium and ruthenium. Its AI model proposed palladium-based recipes while screening 2,942 oxides, although palladium had been treated as a dead end under acidic conditions. After repeated activity tests and more than 1,000 hours of stability testing, the lead modified-palladium candidate performed comparably to ruthenium, and the program identified six palladium-based material families on or near the activity-stability Pareto front.

Lila Sciences used a closed-loop autonomous laboratory to search for acidic oxygen evolution reaction (OER) catalysts, a reaction long dependent on scarce iridium and ruthenium. Its AI model proposed palladium-based recipes while screening 2,942 oxides, although palladium had been treated as a dead end under acidic conditions. After repeated activity tests and more than 1,000 hours of stability testing, the lead modified-palladium candidate performed comparably to ruthenium, and the program identified six palladium-based material families on or near the activity-stability Pareto front.

Test-time training layers replace the recurrent state with a small neural network trained online by gradient descent, which is more expressive than a linear memory but forces sequential computation because each token's gradient depends on every earlier update. ParaTTT instead iteratively refines an estimate of the hidden-layer pre-activation gradients, with each refinement built from a forward, a readout and a backward chunkwise-parallel pass, and the exact online recurrence is a fixed point of that iteration. At 4K context, two refinements reach the error of TTT-MLP's block approximation in less time, and 340M and 1.3B models match or exceed Gated DeltaNet and Mamba-2 on perplexity and recall.

Test-time training layers replace the recurrent state with a small neural network trained online by gradient descent, which is more expressive than a linear memory but forces sequential computation because each token's gradient depends on every earlier update. ParaTTT instead iteratively refines an estimate of the hidden-layer pre-activation gradients, with each refinement built from a forward, a readout and a backward chunkwise-parallel pass, and the exact online recurrence is a fixed point of that iteration. At 4K context, two refinements reach the error of TTT-MLP's block approximation in less time, and 340M and 1.3B models match or exceed Gated DeltaNet and Mamba-2 on perplexity and recall.

Stanford Hospital ran PharmaAgent, an LLM-based vancomycin dosing agent, silently for two months and compared high- and low-autonomy configurations across 403 dosing events in 316 patients. With institutional guidelines, mean absolute percentage error against the treating pharmacist's order was 21.8% for high autonomy versus 37.8% for low autonomy, and recommendations fell within 25% of the order in 73.4% versus 50.9% of events. Yet pharmacists rated 65.8% of 196 reviewed out-of-tolerance recommendations as carrying a safety concern, and neither configuration is ready to dose patients without pharmacist oversight.

Stanford Hospital ran PharmaAgent, an LLM-based vancomycin dosing agent, silently for two months and compared high- and low-autonomy configurations across 403 dosing events in 316 patients. With institutional guidelines, mean absolute percentage error against the treating pharmacist's order was 21.8% for high autonomy versus 37.8% for low autonomy, and recommendations fell within 25% of the order in 73.4% versus 50.9% of events. Yet pharmacists rated 65.8% of 196 reviewed out-of-tolerance recommendations as carrying a safety concern, and neither configuration is ready to dose patients without pharmacist oversight.

When a target cell line has no drug measurements, GRAFT links that line's genetic perturbation responses to drug responses measured in source cells and predicts the transcriptional changes a drug would induce. In leave-one-line-out evaluation across five LINCS cancer lines, GRAFT-D0 reached a mean Top20 Pearson correlation of 0.734, above the source mean of 0.493 and the strongest comparator PRnet at 0.689; across five external cellular contexts the equally weighted mean Top20 correlation was 0.606 against a source mean of 0.416. The method rests on a conditioning state built from how the target's genetic responses differ from source references, and predictions still improve when target genetic measurements are few, though gains vary with data quality and source coverage.

When a target cell line has no drug measurements, GRAFT links that line's genetic perturbation responses to drug responses measured in source cells and predicts the transcriptional changes a drug would induce. In leave-one-line-out evaluation across five LINCS cancer lines, GRAFT-D0 reached a mean Top20 Pearson correlation of 0.734, above the source mean of 0.493 and the strongest comparator PRnet at 0.689; across five external cellular contexts the equally weighted mean Top20 correlation was 0.606 against a source mean of 0.416. The method rests on a conditioning state built from how the target's genetic responses differ from source references, and predictions still improve when target genetic measurements are few, though gains vary with data quality and source coverage.

Single-cell perturbation atlases cover only a fraction of possible experimental combinations, and existing generative models map measured control cells to perturbed cells without a reference for how good a prediction can be. GENESIS instead generates single-cell expression distributions directly from covariates such as cell line or donor, drug or cytokine, dose and batch, training a lightweight multilayer perceptron with the energy score and needing no control cells at inference. On 530 repeated Tahoe-100M plate pairs its Pearson Δ was 0.430 versus 0.432 for experimental repeats and its energy distance 0.158 versus 0.154, neither significantly different; top-100 DEG overlap was 0.289 versus 0.299, 3.4% short.

Single-cell perturbation atlases cover only a fraction of possible experimental combinations, and existing generative models map measured control cells to perturbed cells without a reference for how good a prediction can be. GENESIS instead generates single-cell expression distributions directly from covariates such as cell line or donor, drug or cytokine, dose and batch, training a lightweight multilayer perceptron with the energy score and needing no control cells at inference. On 530 repeated Tahoe-100M plate pairs its Pearson Δ was 0.430 versus 0.432 for experimental repeats and its energy distance 0.158 versus 0.154, neither significantly different; top-100 DEG overlap was 0.289 versus 0.299, 3.4% short.

DiFVM rewrites the complete finite-volume/PISO solution procedure on unstructured polyhedral meshes as vectorized scatter and gather operations over a static graph, giving GPU-native execution and end-to-end automatic differentiation in JAX/XLA. Existing differentiable CFD platforms mostly rely on Cartesian or structured grids and do not directly handle body-conforming complex geometries. On a single NVIDIA L40 GPU, DiFVM is faster than 32-core parallel OpenFOAM in every tested case, the largest with about 10⁶ control volumes; differentiability is demonstrated by recovering a cavity lid velocity and inferring six Windkessel parameters simultaneously.

DiFVM rewrites the complete finite-volume/PISO solution procedure on unstructured polyhedral meshes as vectorized scatter and gather operations over a static graph, giving GPU-native execution and end-to-end automatic differentiation in JAX/XLA. Existing differentiable CFD platforms mostly rely on Cartesian or structured grids and do not directly handle body-conforming complex geometries. On a single NVIDIA L40 GPU, DiFVM is faster than 32-core parallel OpenFOAM in every tested case, the largest with about 10⁶ control volumes; differentiability is demonstrated by recovering a cavity lid velocity and inferring six Windkessel parameters simultaneously.

Autonomous investing agents must learn from outcomes without rewriting their own risk rules. Steward separates editable prompts from an invariant core: nine role-specialised agents decide jointly, code grades every trade nightly, and a quarterly reflector proposes only typed prompt edits that a deterministic gate applies reversibly. Over 102 trading days and 1,632 model calls in a paper account, the portfolio lost 0.3% while the semiconductor sector fell 24.6%, and no injected instruction was followed.

Autonomous investing agents must learn from outcomes without rewriting their own risk rules. Steward separates editable prompts from an invariant core: nine role-specialised agents decide jointly, code grades every trade nightly, and a quarterly reflector proposes only typed prompt edits that a deterministic gate applies reversibly. Over 102 trading days and 1,632 model calls in a paper account, the portfolio lost 0.3% while the semiconductor sector fell 24.6%, and no injected instruction was followed.

Few-step autoregressive video diffusion models gradually lose subject and scene information over long rollouts, and existing distillation targets do not constrain distant history separately. History Forcing builds a context residual from the difference between one autoregressive teacher's predictions with and without extra history, then weights it into the DMD or OPSD distillation target. A 1.3B student reaches a VBench quality score of 85.07 for 15-second generation and a VBench-Long aggregate of 79.82 for 60-second generation, and raises PlayWorld Evolution Quality from 2.35 to 2.70 when the camera leaves and returns.

Few-step autoregressive video diffusion models gradually lose subject and scene information over long rollouts, and existing distillation targets do not constrain distant history separately. History Forcing builds a context residual from the difference between one autoregressive teacher's predictions with and without extra history, then weights it into the DMD or OPSD distillation target. A 1.3B student reaches a VBench quality score of 85.07 for 15-second generation and a VBench-Long aggregate of 79.82 for 60-second generation, and raises PlayWorld Evolution Quality from 2.35 to 2.70 when the camera leaves and returns.

Vision-language-action (VLA) policies emit a whole action sequence per query, but how much of it to execute was previously judged only indirectly, through generation stability or prediction agreement. PAV instead predicts, inside the policy's own task-conditioned representation, what a candidate prefix will lead to, and a verifier trained on trajectory-level outcomes scores whether the task can still be completed. On the four LIBERO suites, PAV averages 94.65% success at 14.0 policy queries per episode, above direct verification without future prediction (84.55%, 22.3 queries) and prediction in DINOv2 features (87.00%, 22.2 queries).

Vision-language-action (VLA) policies emit a whole action sequence per query, but how much of it to execute was previously judged only indirectly, through generation stability or prediction agreement. PAV instead predicts, inside the policy's own task-conditioned representation, what a candidate prefix will lead to, and a verifier trained on trajectory-level outcomes scores whether the task can still be completed. On the four LIBERO suites, PAV averages 94.65% success at 14.0 policy queries per episode, above direct verification without future prediction (84.55%, 22.3 queries) and prediction in DINOv2 features (87.00%, 22.2 queries).

Pre-operative triage of adnexal masses has relied on CA-125, HE4 and ultrasound O-RADS, all with limited sensitivity and specificity. ADAPT-MS instead retrains a classifier for each sample on the proteins actually measured, reading discovery-mode plasma mass spectrometry data directly. Across 623 samples in five independent validation cohorts the classifier reached a summary AUC of 0.835, comparable to O-RADS alone, and rose to 0.922 when combined with O-RADS.

Pre-operative triage of adnexal masses has relied on CA-125, HE4 and ultrasound O-RADS, all with limited sensitivity and specificity. ADAPT-MS instead retrains a classifier for each sample on the proteins actually measured, reading discovery-mode plasma mass spectrometry data directly. Across 623 samples in five independent validation cohorts the classifier reached a summary AUC of 0.835, comparable to O-RADS alone, and rose to 0.922 when combined with O-RADS.

Homomorphic encryption can run a model directly on ciphertext, but the latency introduced by bootstrapping, the costly operation that refreshes a ciphertext's available multiplication depth, keeps encrypted large models impractical, and existing optimizations sit in different, sometimes closed-source frameworks that are difficult and time-consuming to integrate. Perseus is a public Python library spanning the cryptographic backend, data packing and bootstrap placement, with in-kernel key regeneration, complex-Cachemir packing and bootstrap planning driven by coefficient magnitude and packing periodicity on GPUs. On Llama-3-8B decoding its latency is lower than the fastest known implementation, Sylph, which runs on roughly the same hardware but on the closed-source HEaaN.

Homomorphic encryption can run a model directly on ciphertext, but the latency introduced by bootstrapping, the costly operation that refreshes a ciphertext's available multiplication depth, keeps encrypted large models impractical, and existing optimizations sit in different, sometimes closed-source frameworks that are difficult and time-consuming to integrate. Perseus is a public Python library spanning the cryptographic backend, data packing and bootstrap placement, with in-kernel key regeneration, complex-Cachemir packing and bootstrap planning driven by coefficient magnitude and packing periodicity on GPUs. On Llama-3-8B decoding its latency is lower than the fastest known implementation, Sylph, which runs on roughly the same hardware but on the closed-source HEaaN.

Best-of-N inference returns the candidate with the highest proxy reward, so a larger budget reads the extreme tail of the reward distribution rather than its average ranking quality. The work writes the BoN value as a Beta-weighted integral of conditional true reward over within-prompt score quantiles, showing that BoN depends on scores only through their ranks and can decline only when the calibration curve in the tail turns downward. On public candidate pools, negative tail excess concentrates in code generation, and a cross-fitted tail statistic predicts whether a budget cap pays off on held-out data about as well as rank accuracy.

Best-of-N inference returns the candidate with the highest proxy reward, so a larger budget reads the extreme tail of the reward distribution rather than its average ranking quality. The work writes the BoN value as a Beta-weighted integral of conditional true reward over within-prompt score quantiles, showing that BoN depends on scores only through their ranks and can decline only when the calibration curve in the tail turns downward. On public candidate pools, negative tail excess concentrates in code generation, and a cross-fitted tail statistic predicts whether a budget cap pays off on held-out data about as well as rank accuracy.

RI-SINDy removes known Hill-type regulatory terms from the sparse-regression library and solves each coefficient separately from a drift-balance condition, the near-zero time average of the derivative along a bounded trajectory, avoiding the ill-conditioning that arises when regulatory and polynomial terms are nearly collinear over an oscillation. On Hes1 experimental data it gives the lowest trajectory error among Traditional SINDy, SR3 and Nullcline Reconstruction; on synthetic Goodwin and NF-κB models it recovers the correct sparse structure, and in a Goodwin noise sweep it still finds all 6 active terms at 15% noise.

RI-SINDy removes known Hill-type regulatory terms from the sparse-regression library and solves each coefficient separately from a drift-balance condition, the near-zero time average of the derivative along a bounded trajectory, avoiding the ill-conditioning that arises when regulatory and polynomial terms are nearly collinear over an oscillation. On Hes1 experimental data it gives the lowest trajectory error among Traditional SINDy, SR3 and Nullcline Reconstruction; on synthetic Goodwin and NF-κB models it recovers the correct sparse structure, and in a Goodwin noise sweep it still finds all 6 active terms at 15% noise.

Anthropic added dynamic workflows in beta to Claude Managed Agents on the Claude Platform: for work with many pieces, such as reviewing hundreds of documents, an agent can write a program that runs many agents in phases and combines their results, and the server runs it in the background as a workflow run. It is turned on by setting the agent's multiagent field to multiagent_20261001 with workflows enabled; the user can tell the agent in its system prompt when to start a run and follow each run through workflow_run.* events on the session's event stream.

Anthropic added dynamic workflows in beta to Claude Managed Agents on the Claude Platform: for work with many pieces, such as reviewing hundreds of documents, an agent can write a program that runs many agents in phases and combines their results, and the server runs it in the background as a workflow run. It is turned on by setting the agent's multiagent field to multiagent_20261001 with workflows enabled; the user can tell the agent in its system prompt when to start a run and follow each run through workflow_run.* events on the session's event stream.

In electronic-waste disassembly, demonstrations for vision-language-action (VLA) policies are scarce, and supervised fine-tuning alone struggles to cover the state shifts a deployed policy meets. PACER compiles autonomous rollouts, operator corrections and supplementary clean demonstrations into typed action-chunk rows, then weights each row by progress toward the declared target and by the trace's source and role, while keeping the native supervised action loss. On held-out hardware for UR5e motherboard disassembly across five components, the selected policy completed 87/100 approach-and-handoff trials, against 54/100 for uniformly weighted post-SFT and 79/100 for fixed-geometry reweighting.

In electronic-waste disassembly, demonstrations for vision-language-action (VLA) policies are scarce, and supervised fine-tuning alone struggles to cover the state shifts a deployed policy meets. PACER compiles autonomous rollouts, operator corrections and supplementary clean demonstrations into typed action-chunk rows, then weights each row by progress toward the declared target and by the trace's source and role, while keeping the native supervised action loss. On held-out hardware for UR5e motherboard disassembly across five components, the selected policy completed 87/100 approach-and-handoff trials, against 54/100 for uniformly weighted post-SFT and 79/100 for fixed-geometry reweighting.

Nace released Drex Max, a decision model that writes no text: given a screenshot, screencast, scanned document, chart or photo plus a set of typed questions, one forward pass returns a probability for every option supplied. It scores 67.6 on the public Decision Index 0.4.0, above Drex v2.0 at 61.3 and Drex v1.5 at 58.1, and 74.5 on the public Vision Index. Median API latency is 183 ms per decision and cost is $0.06 per 1k decisions.

Nace released Drex Max, a decision model that writes no text: given a screenshot, screencast, scanned document, chart or photo plus a set of typed questions, one forward pass returns a probability for every option supplied. It scores 67.6 on the public Decision Index 0.4.0, above Drex v2.0 at 61.3 and Drex v1.5 at 58.1, and 74.5 on the public Vision Index. Median API latency is 183 ms per decision and cost is $0.06 per 1k decisions.

Oncology guidelines are revised as evidence emerges, so an answer can faithfully cite a real guideline and still cite a superseded version. ASCOBench ties 288 questions to specific versions of ASCO breast and prostate cancer guidelines, with oncologist-reviewed reference answers; about one in four version-sensitive drafts needed revision or rewriting, against fewer than one in ten factual ones. Only 1 of 14 superseded guideline documents states that it has been replaced, and for the strong model Gemini retrieval over several guideline versions made stale answers about four times as frequent as answering without retrieval, while the verification-first SentryLine lowered incorrect answers to 4.2-5.2%.

Oncology guidelines are revised as evidence emerges, so an answer can faithfully cite a real guideline and still cite a superseded version. ASCOBench ties 288 questions to specific versions of ASCO breast and prostate cancer guidelines, with oncologist-reviewed reference answers; about one in four version-sensitive drafts needed revision or rewriting, against fewer than one in ten factual ones. Only 1 of 14 superseded guideline documents states that it has been replaced, and for the strong model Gemini retrieval over several guideline versions made stale answers about four times as frequent as answering without retrieval, while the verification-first SentryLine lowered incorrect answers to 4.2-5.2%.

Long-term memory systems usually store facts as entries and pick them by relevance at query time, which makes facts scattered across sessions and changing over time hard to compose. MaC puts memory construction and retrieval under one code abstraction: facts become executable functions with canonical identifiers and source references, and a query is compiled into a retrieval program that an executor runs after a single planning call. On LOCOMO and LONGMEMEVAL, MaC reaches the highest overall LLM-judge score in all four evaluation settings, improving over the strongest baseline in each by up to 15.91% and 9.06%, while substantially cutting token consumption.

Long-term memory systems usually store facts as entries and pick them by relevance at query time, which makes facts scattered across sessions and changing over time hard to compose. MaC puts memory construction and retrieval under one code abstraction: facts become executable functions with canonical identifiers and source references, and a query is compiled into a retrieval program that an executor runs after a single planning call. On LOCOMO and LONGMEMEVAL, MaC reaches the highest overall LLM-judge score in all four evaluation settings, improving over the strongest baseline in each by up to 15.91% and 9.06%, while substantially cutting token consumption.

Large language models write stories that often lack the rising and falling reader experience that makes narrative engaging. A new framework first learns how reader responses such as suspense or humor evolve in high-quality human-written stories, then uses a small model to predict the response the next segment should elicit and steers an off-the-shelf writer model with that prediction. Across suspense and humor, with GPT-5.5 and Qwen3.5-27B as writers, generated stories improve trajectory-based fidelity to human-written stories by 12%–35% while largely preserving story coherence.

Large language models write stories that often lack the rising and falling reader experience that makes narrative engaging. A new framework first learns how reader responses such as suspense or humor evolve in high-quality human-written stories, then uses a small model to predict the response the next segment should elicit and steers an off-the-shelf writer model with that prediction. Across suspense and humor, with GPT-5.5 and Qwen3.5-27B as writers, generated stories improve trajectory-based fidelity to human-written stories by 12%–35% while largely preserving story coherence.

Benchling released Science Foundries, a way for scientists and AI agents to divide work over the same scientific data and context and run processes such as protein production, lead optimization or process improvement end to end. A user picks a process or goal, then sets which steps agents run and where scientists must approve; agents execute and monitor work, coordinate handoffs and surface key decisions. In a high-throughput antibody example, design, cloning, lab and analysis agents connect to models, a CRO, lab automation and training data, so assay results finished late at night are analyzed before the scientist starts the day.

Benchling released Science Foundries, a way for scientists and AI agents to divide work over the same scientific data and context and run processes such as protein production, lead optimization or process improvement end to end. A user picks a process or goal, then sets which steps agents run and where scientists must approve; agents execute and monitor work, coordinate handoffs and surface key decisions. In a high-throughput antibody example, design, cloning, lab and analysis agents connect to models, a CRO, lab automation and training data, so assay results finished late at night are analyzed before the scientist starts the day.

PRISM-Net splits a product image's holistic features into a line pathway and a color pathway, supervises the two routers zero-shot with frozen SigLIP text prompts, and fuses both with an image-adaptive meta-router. In 5-fold cross-validation over 275,454 e-commerce images in 18 categories and full-scale retrieval of 55,091 queries against a 220,363-image gallery, the SwinV2 backbone gains 5.74 points in line mAP@10, while the color-only configuration gains 9.12 points in color mAP@10; on ConvNeXt-V2 and DINOv3, product retrieval and attribute alignment trade off.

PRISM-Net splits a product image's holistic features into a line pathway and a color pathway, supervises the two routers zero-shot with frozen SigLIP text prompts, and fuses both with an image-adaptive meta-router. In 5-fold cross-validation over 275,454 e-commerce images in 18 categories and full-scale retrieval of 55,091 queries against a 220,363-image gallery, the SwinV2 backbone gains 5.74 points in line mAP@10, while the color-only configuration gains 9.12 points in color mAP@10; on ConvNeXt-V2 and DINOv3, product retrieval and attribute alignment trade off.

Recurrent-memory models compress history into a fixed set of slots, and earlier multi-slot memories wrote every token to the same number of slots, merging whether to store a token with how many slots to update. Factorized Write Routing separates the two: a learned gate first decides whether the current token is written, then a slot router chooses a fixed or token-adaptive number of slots, while queries still read all slots. Across ten seeds on two synthetic long-context tasks with four slots per write, the result reverses with information density: always writing reaches 0.472 long-context accuracy on dense Feature Recall versus 0.337 for the learned gate, while the learned gate reaches 0.222 on sparse-fact Passkey versus 0.048 for always writing.

Recurrent-memory models compress history into a fixed set of slots, and earlier multi-slot memories wrote every token to the same number of slots, merging whether to store a token with how many slots to update. Factorized Write Routing separates the two: a learned gate first decides whether the current token is written, then a slot router chooses a fixed or token-adaptive number of slots, while queries still read all slots. Across ten seeds on two synthetic long-context tasks with four slots per write, the result reverses with information density: always writing reaches 0.472 long-context accuracy on dense Feature Recall versus 0.337 for the learned gate, while the learned gate reaches 0.222 on sparse-fact Passkey versus 0.048 for always writing.

In multimodal federated learning, when clients observe only some modalities, aggregating by which modality is available drops encoders that still shape predictions. RASO instead picks each component's contributors by prediction dependence and local-optimizer membership, keeps momentum on encoders, and synchronizes the shared prediction block to its current aggregate. Across KU-HAR, CREMA-D and other datasets under two fusion architectures, concatenation raises KU-HAR macro-F1 from 41.30 under the availability rule to 49.13, while under masked attention RASO leads on UCI-HAR, KU-HAR and CREMA-D and FedOpt leads on CrisisMMD.

In multimodal federated learning, when clients observe only some modalities, aggregating by which modality is available drops encoders that still shape predictions. RASO instead picks each component's contributors by prediction dependence and local-optimizer membership, keeps momentum on encoders, and synchronizes the shared prediction block to its current aggregate. Across KU-HAR, CREMA-D and other datasets under two fusion architectures, concatenation raises KU-HAR macro-F1 from 41.30 under the availability rule to 49.13, while under masked attention RASO leads on UCI-HAR, KU-HAR and CREMA-D and FedOpt leads on CrisisMMD.

Pretrained generative policies give robot control strong behavior priors, but adapting them efficiently with online reinforcement learning remains hard: noise-space residual methods regulate adaptation through fixed latent feasible regions that may exclude valuable noise edits, while action-space residual methods are more sample-efficient yet still leave room for improvement. CNDO drops predefined latent feasible regions, learning state- and noise-conditioned directions regulated by a soft conservative constraint built from conservative value estimation and gated execution; CANDO adds an action-space residual on top. Experiments report 76.1 NAUC and a 91.4% final success rate for CANDO, 6.6 NAUC and 4.1 percentage points above the best single-space results.

Pretrained generative policies give robot control strong behavior priors, but adapting them efficiently with online reinforcement learning remains hard: noise-space residual methods regulate adaptation through fixed latent feasible regions that may exclude valuable noise edits, while action-space residual methods are more sample-efficient yet still leave room for improvement. CNDO drops predefined latent feasible regions, learning state- and noise-conditioned directions regulated by a soft conservative constraint built from conservative value estimation and gated execution; CANDO adds an action-space residual on top. Experiments report 76.1 NAUC and a 91.4% final success rate for CANDO, 6.6 NAUC and 4.1 percentage points above the best single-space results.

At its San Francisco summit, Sierra announced two new Agent OS models, Curie and Fleming, alongside Tandem Voice, Persona Studio and Ghostwriter. Curie powers the core orchestration loop of a conversation, deciding which path to follow, which tools to call and what replies customers see; Fleming detects when a caller is another agent, so businesses can handle bots and fraud. Sierra says Agent OS automatically ensures compliance, low-latency voice and a better cost basis, leaving companies to focus on building their own agent.

At its San Francisco summit, Sierra announced two new Agent OS models, Curie and Fleming, alongside Tandem Voice, Persona Studio and Ghostwriter. Curie powers the core orchestration loop of a conversation, deciding which path to follow, which tools to call and what replies customers see; Fleming detects when a caller is another agent, so businesses can handle bots and fraud. Sierra says Agent OS automatically ensures compliance, low-latency voice and a better cost basis, leaving companies to focus on building their own agent.

Ranking in Taobao's Guess You Like service no longer generates a list over a truncated single-business candidate set and then hands it to cross-business blending. UGR generates one list over roughly a thousand organic, ad and other-business candidates per request, models historical ad exposure separately during training, and adds current ad value at serving. In an online A/B test on Guess You Like, UGR raised Core IPV by 1.76%, Transaction Volume by 3.29% and GMV by 2.10% over the production baseline.

Ranking in Taobao's Guess You Like service no longer generates a list over a truncated single-business candidate set and then hands it to cross-business blending. UGR generates one list over roughly a thousand organic, ad and other-business candidates per request, models historical ad exposure separately during training, and adds current ad value at serving. In an online A/B test on Guess You Like, UGR raised Core IPV by 1.76%, Transaction Volume by 3.29% and GMV by 2.10% over the production baseline.

Before the European Society of Cardiology Congress 2026 opened, three large language models locked predictions of whether each trial's primary endpoint would be met, using only design information public at registration. Of 48 scorable trials, 28 were positive; Claude, GPT and Gemini separated positive from negative trials with AUCs of 0.80, 0.79 and 0.74, all above chance. Among 26 trials whose effect was reported exactly as locked, the direction was correct in 88% of predictions, and the median multiplicative error of ratio estimates was about 15%.

Before the European Society of Cardiology Congress 2026 opened, three large language models locked predictions of whether each trial's primary endpoint would be met, using only design information public at registration. Of 48 scorable trials, 28 were positive; Claude, GPT and Gemini separated positive from negative trials with AUCs of 0.80, 0.79 and 0.74, all above chance. Among 26 trials whose effect was reported exactly as locked, the direction was correct in 88% of predictions, and the median multiplicative error of ratio estimates was about 15%.

The Simons Foundation announced a partnership with Anthropic and OpenAI under which the three will commit $15 million in combined funding and resources to help the mathematics community navigate the rapid rise of artificial intelligence. The foundation will manage and award funding, including through an open call for proposals, to support workshops and initiatives that help researchers evaluate recent AI-assisted results, learn and use new tools, share perspectives and shape collective standards for AI use. Anthropic and OpenAI will also provide AI tools and computing resources to a select group of researchers supported by the Simons Foundation, extending beyond mathematics to include basic science.

The Simons Foundation announced a partnership with Anthropic and OpenAI under which the three will commit $15 million in combined funding and resources to help the mathematics community navigate the rapid rise of artificial intelligence. The foundation will manage and award funding, including through an open call for proposals, to support workshops and initiatives that help researchers evaluate recent AI-assisted results, learn and use new tools, share perspectives and shape collective standards for AI use. Anthropic and OpenAI will also provide AI tools and computing resources to a select group of researchers supported by the Simons Foundation, extending beyond mathematics to include basic science.

Modal's Sandbox Sidecars have entered public alpha: sidecar containers run beside a main Sandbox on the same host, connected through a low-latency internal network, and can serve custom network proxies, multi-container workloads, or separate an agent harness from its tool execution. The same update adds day-zero support for Kimi K3, Qwen 3.8, GLM 5.3 and GLM 5.3 Flash, and cuts the broad-region pricing multiplier from 1.5x to 1.15x. Sidecars remain an experimental interface, created in the example through an underscore-prefixed call.

Modal's Sandbox Sidecars have entered public alpha: sidecar containers run beside a main Sandbox on the same host, connected through a low-latency internal network, and can serve custom network proxies, multi-container workloads, or separate an agent harness from its tool execution. The same update adds day-zero support for Kimi K3, Qwen 3.8, GLM 5.3 and GLM 5.3 Flash, and cuts the broad-region pricing multiplier from 1.5x to 1.15x. Sidecars remain an experimental interface, created in the example through an underscore-prefixed call.

OpenAI introduced a visual ad format in ChatGPT that shows product inspiration, usage or experiences as images, clearly labeled and kept separate from the image being generated. The format is first tested during image generation for Free and Go users in the US, with an initial group of advertisers starting later this month. OpenAI also expanded conversion-data integrations and attribution partners, and says incrementality measurement is still early, with three partners exploring geo-based experiments for causal impact.

OpenAI introduced a visual ad format in ChatGPT that shows product inspiration, usage or experiences as images, clearly labeled and kept separate from the image being generated. The format is first tested during image generation for Free and Go users in the US, with an initial group of advertisers starting later this month. OpenAI also expanded conversion-data integrations and attribution partners, and says incrementality measurement is still early, with three partners exploring geo-based experiments for causal impact.

DIKT decomposes the probability that a student answers the next exercise correctly into a product of per-skill success probabilities, giving each skill a cognitive state bounded between 0 and 1 that evolves over time. Most deep knowledge-tracing models predict from implicit vectors and cannot say which skill is holding the result back. On ASSIST09, ASSIST17 and Algebra2005, DIKT's AUC exceeds every listed baseline, an average absolute gain of 2.04 percentage points.

DIKT decomposes the probability that a student answers the next exercise correctly into a product of per-skill success probabilities, giving each skill a cognitive state bounded between 0 and 1 that evolves over time. Most deep knowledge-tracing models predict from implicit vectors and cannot say which skill is holding the result back. On ASSIST09, ASSIST17 and Algebra2005, DIKT's AUC exceeds every listed baseline, an average absolute gain of 2.04 percentage points.

A decision method called Triadic Viability Alignment (TVA) uses no instance-level normative labels in the controlled multi-agent Triadic Commons Environment, scoring 3.06 percentage points below a preference-supervised baseline on task success while staying inside a pre-set 10-point non-inferiority margin, and raising joint viability by 83.68 points. It places three roles, the acting subject, affected others and shared conditions, plus non-compensatory hard floors ahead of utility comparison, removing actions predicted to breach a floor. Risk-sensitive control and multi-step model-predictive control nevertheless exceed it on joint viability by 10.42 and 8.68 points, and structured data selection did not reduce data requirements.

A decision method called Triadic Viability Alignment (TVA) uses no instance-level normative labels in the controlled multi-agent Triadic Commons Environment, scoring 3.06 percentage points below a preference-supervised baseline on task success while staying inside a pre-set 10-point non-inferiority margin, and raising joint viability by 83.68 points. It places three roles, the acting subject, affected others and shared conditions, plus non-compensatory hard floors ahead of utility comparison, removing actions predicted to breach a floor. Risk-sensitive control and multi-step model-predictive control nevertheless exceed it on joint viability by 10.42 and 8.68 points, and structured data selection did not reduce data requirements.

Google's Nano Banana 2.1 is now available on AI Gateway as google/gemini-nano-banana-2.1, so developers can call it directly. Compared with earlier Nano Banana models, it improves image generation and editing, covering product recontextualization, mask- and ink-based editing, and factuality, while maintaining Flash-tier latency and cost. It can be used through the AI SDK, the OpenAI-compatible Chat Completions API, or AI CLI.

Google's Nano Banana 2.1 is now available on AI Gateway as google/gemini-nano-banana-2.1, so developers can call it directly. Compared with earlier Nano Banana models, it improves image generation and editing, covering product recontextualization, mask- and ink-based editing, and factuality, while maintaining Flash-tier latency and cost. It can be used through the AI SDK, the OpenAI-compatible Chat Completions API, or AI CLI.

LLM-driven interactive narrative systems have been hard to compare because each experience manager is tied to its own bespoke scenario. An LLM-based narrative platform turns experience managers into interchangeable modules over a shared symbolic world model of facts, operations and goals, and every applied operation carries explicit preconditions and effects, so any session can be replayed as a common state-transition path. On a Little Red Riding Hood scenario, open generation, Monte Carlo repair and autonomous characters each ran 100 simulated sessions, and no manager led on all four properties: player agency, manager controllability, author intent and robustness.

LLM-driven interactive narrative systems have been hard to compare because each experience manager is tied to its own bespoke scenario. An LLM-based narrative platform turns experience managers into interchangeable modules over a shared symbolic world model of facts, operations and goals, and every applied operation carries explicit preconditions and effects, so any session can be replayed as a common state-transition path. On a Little Red Riding Hood scenario, open generation, Monte Carlo repair and autonomous characters each ran 100 simulated sessions, and no manager led on all four properties: player agency, manager controllability, author intent and robustness.

Benchling's Model Hub lets scientists discover, run and track scientific AI models inside Benchling, with each prediction's inputs, model version and timestamp wired back into the R&D record. Previously, running a model meant waiting for DevOps to stand up infrastructure, and results landed on a shared drive disconnected from the sequence that produced them. The hub currently carries structure prediction models including AlphaFold 2, Chai-1, Boltz-2, OpenFold 2, OpenFold 3 and Protenix, plus property prediction models such as Lilly TuneLab, Inductive Bio and open-source DeepSP, with batch prediction and a v3 alpha API.

Benchling's Model Hub lets scientists discover, run and track scientific AI models inside Benchling, with each prediction's inputs, model version and timestamp wired back into the R&D record. Previously, running a model meant waiting for DevOps to stand up infrastructure, and results landed on a shared drive disconnected from the sequence that produced them. The hub currently carries structure prediction models including AlphaFold 2, Chai-1, Boltz-2, OpenFold 2, OpenFold 3 and Protenix, plus property prediction models such as Lilly TuneLab, Inductive Bio and open-source DeepSP, with batch prediction and a v3 alpha API.