AI and science frontiers · 2026-11-04

Learning the continuous-time evolution of nonlinear dynamical systems usually relies on trajectory objectives that simulate the learned dynamics during training and backpropagate through the solver, at a cost growing with the rollout horizon. Velocity matching instead regresses the velocity field directly, using piecewise-linear interpolation of observed trajectories and points sampled uniformly within each interval, so no simulation is needed. On nonlinear ODEs and semi-discretized time-dependent PDEs it attains errors more than an order of magnitude lower than trajectory-based objectives under a common iteration budget, and keeps accuracy under schemes and step sizes not used in training.

Learning the continuous-time evolution of nonlinear dynamical systems usually relies on trajectory objectives that simulate the learned dynamics during training and backpropagate through the solver, at a cost growing with the rollout horizon. Velocity matching instead regresses the velocity field directly, using piecewise-linear interpolation of observed trajectories and points sampled uniformly within each interval, so no simulation is needed. On nonlinear ODEs and semi-discretized time-dependent PDEs it attains errors more than an order of magnitude lower than trajectory-based objectives under a common iteration budget, and keeps accuracy under schemes and step sizes not used in training.

Interactive video world models draw frames from a memory bank stamped with capture times, and a frozen generator reads the newest timestamp as truth, so after an agent undoes an action the only view of the present is an older frame judged outdated. ReDate keeps every frame and re-dates retrieved tokens by valid time read from the action log, so a reverted region reads as one that has just reached a new state, with no training. On FAR's 352 reverted steps the current-state share rises from 0.36 to 0.62 and the outdated share falls from 0.27 to 0.09; on HY-WorldPlay it rises from 0.003 to 0.19.

Interactive video world models draw frames from a memory bank stamped with capture times, and a frozen generator reads the newest timestamp as truth, so after an agent undoes an action the only view of the present is an older frame judged outdated. ReDate keeps every frame and re-dates retrieved tokens by valid time read from the action log, so a reverted region reads as one that has just reached a new state, with no training. On FAR's 352 reverted steps the current-state share rises from 0.36 to 0.62 and the outdated share falls from 0.27 to 0.09; on HY-WorldPlay it rises from 0.003 to 0.19.

LLM watermarks must keep output diversity under a fixed key and must not lower draft acceptance under speculative decoding, and the two have been hard to hold together. TextSeal2 merges two keys' Gaussian values into one virtual key and couples it to fresh noise through a Gaussian copula, so the detector reads full signal on every token; under speculative decoding it decides each round's last token with one more exponential race on the leftovers of the draft races. At equal diversity on three models it detects 8 to 22 percentage points more 100-token answers than TextSeal, and under speculative decoding 18 and 22 points more 200-token answers than TextSeal and textGrain, at unchanged acceptance.

LLM watermarks must keep output diversity under a fixed key and must not lower draft acceptance under speculative decoding, and the two have been hard to hold together. TextSeal2 merges two keys' Gaussian values into one virtual key and couples it to fresh noise through a Gaussian copula, so the detector reads full signal on every token; under speculative decoding it decides each round's last token with one more exponential race on the leftovers of the draft races. At equal diversity on three models it detects 8 to 22 percentage points more 100-token answers than TextSeal, and under speculative decoding 18 and 22 points more 200-token answers than TextSeal and textGrain, at unchanged acceptance.

Probabilistic time-series forecasting must capture temporal structure, express uncertainty and infer quickly; diffusion and flow-matching methods buy flexibility with multi-step iterative sampling whose cost grows with the number of steps. KGO treats prediction as the evolution of structured uncertainty in a Koopman latent space: KoPE extrapolates latent trajectories with a learned linear operator, KoFM transports noise to the future in one step via a matrix exponential, and AUG adapts the deterministic and stochastic mix per variable and horizon. Across 17 ProbTS settings, KGO reports the lowest CRPS on 13 and the lowest NMAE on 16, with at least a 25-fold inference speedup over iterative generative models.

Probabilistic time-series forecasting must capture temporal structure, express uncertainty and infer quickly; diffusion and flow-matching methods buy flexibility with multi-step iterative sampling whose cost grows with the number of steps. KGO treats prediction as the evolution of structured uncertainty in a Koopman latent space: KoPE extrapolates latent trajectories with a learned linear operator, KoFM transports noise to the future in one step via a matrix exponential, and AUG adapts the deterministic and stochastic mix per variable and horizon. Across 17 ProbTS settings, KGO reports the lowest CRPS on 13 and the lowest NMAE on 16, with at least a 25-fold inference speedup over iterative generative models.

LLM agents usually pick tool calls by trial and error, with no guarantee that a plan can execute. CGAP compiles the tool catalog deterministically into an abstract PDDL model that tracks only which fields are available, lets a classical planner order the calls, and has an LLM add entity-specific calls whose arguments and results are checked. Across ToolSandbox, ComplexFuncBench, BFCL multi-turn and τ-bench retail with three LLMs, it beats a standard tool-calling agent, by up to 28.9 points.

LLM agents usually pick tool calls by trial and error, with no guarantee that a plan can execute. CGAP compiles the tool catalog deterministically into an abstract PDDL model that tracks only which fields are available, lets a classical planner order the calls, and has an LLM add entity-specific calls whose arguments and results are checked. Across ToolSandbox, ComplexFuncBench, BFCL multi-turn and τ-bench retail with three LLMs, it beats a standard tool-calling agent, by up to 28.9 points.

Vision-language models can read a task yet still fumble the grasp: on RoboDojo, removing arm execution raises the same model's success rate from 22.5% to 76.2%, locating the bottleneck in physical execution. Auriga separates model decisions from robot execution, with the model keeping an explicit task belief and every motion passing a single admission gateway that previews the trajectory, checks stage tolerances during execution, and stops on divergence. Frozen for evaluation, Auriga reaches 57.7% success and 64.0 Score across 42 tasks, ranking first, and 73.2% on RoboCasa365.

Vision-language models can read a task yet still fumble the grasp: on RoboDojo, removing arm execution raises the same model's success rate from 22.5% to 76.2%, locating the bottleneck in physical execution. Auriga separates model decisions from robot execution, with the model keeping an explicit task belief and every motion passing a single admission gateway that previews the trajectory, checks stage tolerances during execution, and stops on divergence. Frozen for evaluation, Auriga reaches 57.7% success and 64.0 Score across 42 tasks, ranking first, and 73.2% on RoboCasa365.

Prefill jailbreaks force an unsafe opening into the assistant turn so a model that would refuse continues instead. Tracking vocabulary-grounded signals with the Jacobian lens shows refusal weakening faster than harmfulness as the forced prefix deepens, reaching the benign baseline by free generation. J-Anchor, a training-free inference-time defense, anchors each request to its pre-attack assistant-header state and restores only the refusal components that fall below that target during prefilling and decoding, cutting attack success on fixed and deep prefills across four open-weight instruction models while keeping benign refusal and reasoning close to the undefended model.

Prefill jailbreaks force an unsafe opening into the assistant turn so a model that would refuse continues instead. Tracking vocabulary-grounded signals with the Jacobian lens shows refusal weakening faster than harmfulness as the forced prefix deepens, reaching the benign baseline by free generation. J-Anchor, a training-free inference-time defense, anchors each request to its pre-attack assistant-header state and restores only the refusal components that fall below that target during prefilling and decoding, cutting attack success on fixed and deep prefills across four open-weight instruction models while keeping benign refusal and reasoning close to the undefended model.

Co2ODE models only the part of a biochemical network that carries signal in a given cellular context: a conditional mask selects the markers tied to rewiring, and a variational autoencoder compresses the rest into a learned closure term trained jointly with a sparse ODE. On single-cell phosphoproteomic data spanning 30 cancer-relevant overexpressions, it compresses a 30-marker panel into a 10-dimensional state while matching the target-trajectory accuracy of a neural ODE fit to the full panel; adding a prior score shifts the selected markers at almost no cost in fit, indicating these data do not identify the true rewiring.

Co2ODE models only the part of a biochemical network that carries signal in a given cellular context: a conditional mask selects the markers tied to rewiring, and a variational autoencoder compresses the rest into a learned closure term trained jointly with a sparse ODE. On single-cell phosphoproteomic data spanning 30 cancer-relevant overexpressions, it compresses a 30-marker panel into a 10-dimensional state while matching the target-trajectory accuracy of a neural ODE fit to the full panel; adding a prior score shifts the selected markers at almost no cost in fit, indicating these data do not identify the true rewiring.

Retrieval-augmented generation does not improve every task, since irrelevant evidence can distract a generator, and measuring each candidate retriever's gain directly repeats generation for every configuration. EviCast splits a retriever's net utility into the generator's response to controlled evidence states and that retriever's offline distribution over those states; the response is measured once on a small calibration set and reused. On 14 HotpotQA retrieval configurations, one calibration used 96 generation calls, made 14 of 14 correct deployment decisions with a 2.75 F1-point forecast error, and cut diagnostic generation calls by 80% relative to per-configuration direct measurement.

Retrieval-augmented generation does not improve every task, since irrelevant evidence can distract a generator, and measuring each candidate retriever's gain directly repeats generation for every configuration. EviCast splits a retriever's net utility into the generator's response to controlled evidence states and that retriever's offline distribution over those states; the response is measured once on a small calibration set and reused. On 14 HotpotQA retrieval configurations, one calibration used 96 generation calls, made 14 of 14 correct deployment decisions with a 2.75 F1-point forecast error, and cut diagnostic generation calls by 80% relative to per-configuration direct measurement.

Large reasoning models often keep deliberating after enough information is available, raising decoding latency and memory cost. LoCo moves the compression decision to individual reasoning states: it finds candidate overthinking states in the model's own traces, freezes that prefix, regenerates shorter continuations whose answers verify as correct, and learns a continue-or-conclude preference with DPO. Across five math benchmarks, completion length drops 41.7%–45.4% while accuracy holds or improves; trained on math only, the same model cuts length by 23.4%–36.3% on three code benchmarks.

Large reasoning models often keep deliberating after enough information is available, raising decoding latency and memory cost. LoCo moves the compression decision to individual reasoning states: it finds candidate overthinking states in the model's own traces, freezes that prefix, regenerates shorter continuations whose answers verify as correct, and learns a continue-or-conclude preference with DPO. Across five math benchmarks, completion length drops 41.7%–45.4% while accuracy holds or improves; trained on math only, the same model cuts length by 23.4%–36.3% on three code benchmarks.

RoboFFT, from the PKU-Lingchu joint lab and accepted at CoRL2026, turns the forward noising loss already used to pretrain diffusion and flow policies into an approximate probability ratio, so generative robot policies can be post-trained PPO-style from task success or failure feedback. It was tested on Robomimic simulation tasks, the flow action head of vision-language-action model π₀, and four real-robot manipulation tasks; on eggplant grasping at the same data budget, behavior cloning on pooled data reached 14/20, offline post-training 18/20 and online post-training 20/20.

RoboFFT, from the PKU-Lingchu joint lab and accepted at CoRL2026, turns the forward noising loss already used to pretrain diffusion and flow policies into an approximate probability ratio, so generative robot policies can be post-trained PPO-style from task success or failure feedback. It was tested on Robomimic simulation tasks, the flow action head of vision-language-action model π₀, and four real-robot manipulation tasks; on eggplant grasping at the same data budget, behavior cloning on pooled data reached 14/20, offline post-training 18/20 and online post-training 20/20.

Independently trained models hold complementary capabilities, yet knowledge transfer usually shrinks differences in predictions or representations, which erodes that complementarity. Corroborative Learning has each model learn only an increment to its task scores, then reconstructs those increments from the peer's representations on held-out data groups, constraining directions that reconstruct poorly more strongly. Across six language tasks, 24 models and 36 pairs, it improves on capacity-matched independent adaptation on all six tasks, with a mean gain of 0.52 percentage points, while both models still deploy independently at inference.

Independently trained models hold complementary capabilities, yet knowledge transfer usually shrinks differences in predictions or representations, which erodes that complementarity. Corroborative Learning has each model learn only an increment to its task scores, then reconstructs those increments from the peer's representations on held-out data groups, constraining directions that reconstruct poorly more strongly. Across six language tasks, 24 models and 36 pairs, it improves on capacity-matched independent adaptation on all six tasks, with a mean gain of 0.52 percentage points, while both models still deploy independently at inference.

Preference data contain reversed, tied and otherwise noisy comparisons, yet Direct Preference Optimization (DPO) fits every reported comparison as correct, and its quality falls as the share of flipped labels grows. RW-DPO attaches an unknown reliability weight to each comparison, lets it enter the likelihood linearly, and estimates weights jointly with the policy under a minimax-concave penalty, so comparisons the model agrees with stay at weight one and only contradicted ones are adjusted. In DPO fine-tuning of Qwen2.5-1.5B with 10% to 30% randomly flipped labels, its win rate does not fall as the flip rate grows and exceeds that of DPO trained on clean labels, matching the strongest recent method.

Preference data contain reversed, tied and otherwise noisy comparisons, yet Direct Preference Optimization (DPO) fits every reported comparison as correct, and its quality falls as the share of flipped labels grows. RW-DPO attaches an unknown reliability weight to each comparison, lets it enter the likelihood linearly, and estimates weights jointly with the policy under a minimax-concave penalty, so comparisons the model agrees with stay at weight one and only contradicted ones are adjusted. In DPO fine-tuning of Qwen2.5-1.5B with 10% to 30% randomly flipped labels, its win rate does not fall as the flip rate grows and exceeds that of DPO trained on clean labels, matching the strongest recent method.

Web platforms screen posts for large-language-model text at a very low false-positive rate, and invisible edits such as homoglyphs and zero-width spaces push machine text into the human score range. NormDetect normalizes each document with Unicode rules, then anchors a small per-section classifier head to the normalized Binoculars zero-shot score, with the blend weight chosen on training folds under a human false-positive cap. On eight RAID sections and eleven attacks with the Falcon-7B scoring pair, it recalls 55.5% at a 0.1% calibrated false-positive rate, 5.0 points above normalized Binoculars, while every section's human false-positive bound stays at most 1%.

Web platforms screen posts for large-language-model text at a very low false-positive rate, and invisible edits such as homoglyphs and zero-width spaces push machine text into the human score range. NormDetect normalizes each document with Unicode rules, then anchors a small per-section classifier head to the normalized Binoculars zero-shot score, with the blend weight chosen on training folds under a human false-positive cap. On eight RAID sections and eleven attacks with the Falcon-7B scoring pair, it recalls 55.5% at a 0.1% calibrated false-positive rate, 5.0 points above normalized Binoculars, while every section's human false-positive bound stays at most 1%.

DeepWeave places interpretable logic rules and neural predictions inside one probabilistic reasoning program, so the two act as potentially conflicting evidence for each alert. Machine-learning intrusion detection previously needed large labeled attack sets, degraded under deployment shift, and offered analysts little to inspect. With only a few labeled packet observations per attack class, the Anchored DeepWeave variant raises macro-F1 by up to 34.5% over the matched neural baseline and lowers the attack false-positive rate from 0.075 to 0.001.

DeepWeave places interpretable logic rules and neural predictions inside one probabilistic reasoning program, so the two act as potentially conflicting evidence for each alert. Machine-learning intrusion detection previously needed large labeled attack sets, degraded under deployment shift, and offered analysts little to inspect. With only a few labeled packet observations per attack class, the Anchored DeepWeave variant raises macro-F1 by up to 34.5% over the matched neural baseline and lowers the attack false-positive rate from 0.075 to 0.001.

A team from Lingchu Intelligence with Beihang University, Peking University and the Chinese University of Hong Kong introduces FAFM, a frequency-aware flow-matching method that represents robot actions as frequency coefficients tied to real physical time instead of fixed-length discrete sequences, and explicitly constrains the first time derivative while generating actions by flow matching. Demonstrations recorded at 5 Hz, 10 Hz and 20 Hz can then be mapped onto one physical time scale and trained together without downsampling or interpolation. On a Franka arm, integrating FAFM into π₀.₅ raised success from 80% to 95% on grasp-and-place and from 80% to 90% on multi-obstacle avoidance.

A team from Lingchu Intelligence with Beihang University, Peking University and the Chinese University of Hong Kong introduces FAFM, a frequency-aware flow-matching method that represents robot actions as frequency coefficients tied to real physical time instead of fixed-length discrete sequences, and explicitly constrains the first time derivative while generating actions by flow matching. Demonstrations recorded at 5 Hz, 10 Hz and 20 Hz can then be mapped onto one physical time scale and trained together without downsampling or interpolation. On a Franka arm, integrating FAFM into π₀.₅ raised success from 80% to 95% on grasp-and-place and from 80% to 90% on multi-obstacle avoidance.

In streaming deployment a fixed policy's loss can drift with observed history, so certifying a fixed population risk no longer describes the risk actually encountered. StreamLTT certifies average predictable risk, the conditional risk of each decision given the information available before its outcome, averaged along the stream, and conditional centering turns empirical deviations into a martingale sum. Across seven forecasting datasets and seven deployment risks, its variance-adaptive version pushes the certified frontier close to the risk target while empirical family-wise error stays near the prescribed level.

In streaming deployment a fixed policy's loss can drift with observed history, so certifying a fixed population risk no longer describes the risk actually encountered. StreamLTT certifies average predictable risk, the conditional risk of each decision given the information available before its outcome, averaged along the stream, and conditional centering turns empirical deviations into a martingale sum. Across seven forecasting datasets and seven deployment risks, its variance-adaptive version pushes the certified frontier close to the risk target while empirical family-wise error stays near the prescribed level.

Soft robotic grasping lacks multimodal data that coordinates vision, depth, touch, inertial measurement and text, and existing alignment methods apply one strength to every modality pair. SomaAlign first estimates each modality's relative reliability from masked reconstruction, then uses it to set a pair-specific contrastive temperature, and in a third stage tightens intra-sample cross-modal consistency with conflict-aware constraints. Across 30 directed cross-modal retrieval tasks on the Soma dataset the mean Recall@10 is 90.66%, and on the independently collected Soma-Success set, excluded from representation learning, frozen encoders with lightweight probes reach 94.59% post-hoc grasp-outcome classification accuracy.

Soft robotic grasping lacks multimodal data that coordinates vision, depth, touch, inertial measurement and text, and existing alignment methods apply one strength to every modality pair. SomaAlign first estimates each modality's relative reliability from masked reconstruction, then uses it to set a pair-specific contrastive temperature, and in a third stage tightens intra-sample cross-modal consistency with conflict-aware constraints. Across 30 directed cross-modal retrieval tasks on the Soma dataset the mean Recall@10 is 90.66%, and on the independently collected Soma-Success set, excluded from representation learning, frozen encoders with lightweight probes reach 94.59% post-hoc grasp-outcome classification accuracy.

Edge AI development has relied on developers hand-building fixed agent workflows that struggle with heterogeneous data modalities and diverse task pipelines. SE2 distills development into four atomic operations, retrieve, reason, code and test, and lets a coordinator agent pick the next one from the current context, while mobile runtime feedback and reinforcement fine-tuning let the agents keep evolving. On the IoTBench benchmark and three multimodal edge AI applications, its synthesized programs reach up to 26% higher accuracy and 28% lower resource consumption than eight state-of-the-art baselines.

Edge AI development has relied on developers hand-building fixed agent workflows that struggle with heterogeneous data modalities and diverse task pipelines. SE2 distills development into four atomic operations, retrieve, reason, code and test, and lets a coordinator agent pick the next one from the current context, while mobile runtime feedback and reinforcement fine-tuning let the agents keep evolving. On the IoTBench benchmark and three multimodal edge AI applications, its synthesized programs reach up to 26% higher accuracy and 28% lower resource consumption than eight state-of-the-art baselines.

Google AI & Infrastructure, with Mercor, released Physical-Engineering's Last Exam (PLE), an 85-task benchmark and open competition spanning mechanical, electrical, civil and structural engineering that tests whether agents can complete real engineering work under a licensed engineer's review. Each task is a containerized project folder averaging 61 files, and an agent must find governing codes, run calculations and produce deliverables, earning points only after passing every hard safety check. On the private test set all 14 frontier models scored below 50%, with the top configuration reaching a mean of 48.3% on its native harness.

Google AI & Infrastructure, with Mercor, released Physical-Engineering's Last Exam (PLE), an 85-task benchmark and open competition spanning mechanical, electrical, civil and structural engineering that tests whether agents can complete real engineering work under a licensed engineer's review. Each task is a containerized project folder averaging 61 files, and an agent must find governing codes, run calculations and produce deliverables, earning points only after passing every hard safety check. On the private test set all 14 frontier models scored below 50%, with the top configuration reaching a mean of 48.3% on its native harness.

Mercor and Google's AI & Infrastructure research team released Physical-Engineering's Last Exam (PLE), 85 long-horizon tasks testing whether AI agents can do mechanical, electrical, civil and structural engineering work. Each task runs in a containerized project folder of about 61 files, including PDFs, spreadsheets, emails, 2D CAD drawings, 3D BIM models and planted outdated revisions, with no internet access and up to six hours. The best configuration, Claude Opus 5.5 in Claude Code, reaches 48.3% on the hidden test set at an average of 27 million tokens and 63 minutes per task; its designs met 91% of individual constraints on average, but only 41% cleared every hard check.

Mercor and Google's AI & Infrastructure research team released Physical-Engineering's Last Exam (PLE), 85 long-horizon tasks testing whether AI agents can do mechanical, electrical, civil and structural engineering work. Each task runs in a containerized project folder of about 61 files, including PDFs, spreadsheets, emails, 2D CAD drawings, 3D BIM models and planted outdated revisions, with no internet access and up to six hours. The best configuration, Claude Opus 5.5 in Claude Code, reaches 48.3% on the hidden test set at an average of 27 million tokens and 63 minutes per task; its designs met 91% of individual constraints on average, but only 41% cleared every hard check.

A single-cell transcriptome snapshot carries both where a cell sits along a trajectory and which way its expression is moving, and these were usually estimated separately. ManifoldVelo places each cell on a differentiable unspliced/spliced RNA expression curve and matches the curve's derivatives to a transcription–splicing–degradation kinetic field, including between observed cells, so coordinates, curves and kinetics are fitted together. In multipeak cyclic simulations, spliced-only queries gave circular time errors of 0.64% and 7.40% with spliced-velocity cosines of 0.9776 and 0.9421; on WT2, phase error fell 47.0% relative to the recorded VeloCycle result, while VeloCycle did better on WT1 and the smooth cycle.

A single-cell transcriptome snapshot carries both where a cell sits along a trajectory and which way its expression is moving, and these were usually estimated separately. ManifoldVelo places each cell on a differentiable unspliced/spliced RNA expression curve and matches the curve's derivatives to a transcription–splicing–degradation kinetic field, including between observed cells, so coordinates, curves and kinetics are fitted together. In multipeak cyclic simulations, spliced-only queries gave circular time errors of 0.64% and 7.40% with spliced-velocity cosines of 0.9776 and 0.9421; on WT2, phase error fell 47.0% relative to the recorded VeloCycle result, while VeloCycle did better on WT1 and the smooth cycle.

GFlowNets build discrete objects one component at a time, and the order determines which conditional distributions the network must learn. Previously the order was fixed by a preset or uniform backward policy, and standard training struggled to find useful orders. Learning-Order GFlowNet learns positions and symbols jointly without order labels, converging in four of five runs on checksum dependencies, against zero for a uniform backward policy and two of ten for fixed random orders.

GFlowNets build discrete objects one component at a time, and the order determines which conditional distributions the network must learn. Previously the order was fixed by a preset or uniform backward policy, and standard training struggled to find useful orders. Learning-Order GFlowNet learns positions and symbols jointly without order labels, converging in four of five runs on checksum dependencies, against zero for a uniform backward policy and two of ten for fixed random orders.

Sarvam released Prism, a decision model built on Sarvam 105B for enterprise workflows that need fast, calibrated judgments. Across 25 decision models Sarvam tested, Prism averaged 71.8 over three indices with a 44 ms median response, the highest average among models answering under 100 ms; higher-scoring models took roughly two to six times as long. The weights are open so enterprises can train on their own data and run on their own infrastructure.

Sarvam released Prism, a decision model built on Sarvam 105B for enterprise workflows that need fast, calibrated judgments. Across 25 decision models Sarvam tested, Prism averaged 71.8 over three indices with a 44 ms median response, the highest average among models answering under 100 ms; higher-scoring models took roughly two to six times as long. The weights are open so enterprises can train on their own data and run on their own infrastructure.

Aleph Alpha's Kolibri is a hybrid reasoning model whose system prompt carries one sentence selecting no reasoning, low, medium or high effort. Because its supervised fine-tuning traces carried no effort labels, the team annotated each trace post hoc by its reasoning length relative to traces from the same dataset at the same judged difficulty, with a 2B student model reproducing a larger model's difficulty grades. Median reasoning length orders low below medium below high on 23 of 24 benchmarks; low keeps 96% of the high score at 35% of the length, and the released model keeps 98% at 76% of the completion tokens.

Aleph Alpha's Kolibri is a hybrid reasoning model whose system prompt carries one sentence selecting no reasoning, low, medium or high effort. Because its supervised fine-tuning traces carried no effort labels, the team annotated each trace post hoc by its reasoning length relative to traces from the same dataset at the same judged difficulty, with a 2B student model reproducing a larger model's difficulty grades. Median reasoning length orders low below medium below high on 23 of 24 benchmarks; low keeps 96% of the high score at 35% of the length, and the released model keeps 98% at 76% of the completion tokens.

At its first Forge conference, Fireworks released a set of products for companies building specialized intelligence on their own data and tasks, spanning inference, model access, training and enterprise controls. On the inference side, Fireworks Ultra-fast, powered by NVIDIA Blackwell GPUs, is in public preview and demonstrated up to 1,500 tokens per second on DeepSeek V4.1 Flash; in a benchmarked 30-turn agent task, completion time fell from 1 minute to 8 seconds. Background Mode offers a lower-cost runtime at half price for work that can wait.

At its first Forge conference, Fireworks released a set of products for companies building specialized intelligence on their own data and tasks, spanning inference, model access, training and enterprise controls. On the inference side, Fireworks Ultra-fast, powered by NVIDIA Blackwell GPUs, is in public preview and demonstrated up to 1,500 tokens per second on DeepSeek V4.1 Flash; in a benchmarked 30-turn agent task, completion time fell from 1 minute to 8 seconds. Background Mode offers a lower-cost runtime at half price for work that can wait.

NVIDIA has moved the structural description of distributed Kubernetes jobs out of adapters compiled into tools and into a readable definition. Workload Map, a custom resource plus a Go library, gives JobSet, PyTorchJob, RayCluster and other frameworks one structural view. Previously each custom resource needed its own adapter to interpret roles, turn status into phases and edit pod templates without breaking the job, so support for a new type often waited for a release. Definitions state template paths, replica counts and status mappings in YAML, shared by schedulers and platforms; a worked JobSet resolves into one leader and two workers totaling 12 GPUs.

NVIDIA has moved the structural description of distributed Kubernetes jobs out of adapters compiled into tools and into a readable definition. Workload Map, a custom resource plus a Go library, gives JobSet, PyTorchJob, RayCluster and other frameworks one structural view. Previously each custom resource needed its own adapter to interpret roles, turn status into phases and edit pod templates without breaking the job, so support for a new type often waited for a release. Definitions state template paths, replica counts and status mappings in YAML, shared by schedulers and platforms; a worked JobSet resolves into one leader and two workers totaling 12 GPUs.

JustVugg released colibri, an open-source inference engine written in pure C with one file per model family and no GPU required. A mixture-of-experts model uses only a small part of itself for each token, so colibri keeps that part in RAM and reads the remaining experts from disk on demand. In the official example, the 744B-parameter int4 GLM-5.2 MoE model is ready in 32 seconds on CPU with 9.9 GB resident; the minimum is 8 GB of RAM, and the smallest model needs 22 GB of disk.

JustVugg released colibri, an open-source inference engine written in pure C with one file per model family and no GPU required. A mixture-of-experts model uses only a small part of itself for each token, so colibri keeps that part in RAM and reads the remaining experts from disk on demand. In the official example, the 744B-parameter int4 GLM-5.2 MoE model is ready in 32 seconds on CPU with 9.9 GB resident; the minimum is 8 GB of RAM, and the smallest model needs 22 GB of disk.

Collinear AI released CWE-bench v2, a public set of 100 tasks in which an agent gets a checkout of a real open-source repository and must find and fix a security weakness without breaking existing functionality, covering 91 CWEs. Scoring is fully programmatic and reproducible, and a held-out set tracks tasks no model has seen. Among 11 evaluated models, Gemini 4 Argon leads at 67.8% pass@1 and Claude Opus 5.5 follows at 67.3%; 95 tasks are solved by at least one model, yet the leader fails 19% of the set on every attempt.

Collinear AI released CWE-bench v2, a public set of 100 tasks in which an agent gets a checkout of a real open-source repository and must find and fix a security weakness without breaking existing functionality, covering 91 CWEs. Scoring is fully programmatic and reproducible, and a held-out set tracks tasks no model has seen. Among 11 evaluated models, Gemini 4 Argon leads at 67.8% pass@1 and Claude Opus 5.5 follows at 67.3%; 95 tasks are solved by at least one model, yet the leader fails 19% of the set on every attempt.

Trained anchor-based Gaussian-splat scenes stay too large for phones and headsets, and existing post-hoc converters saturate early. This scheme treats a trained model as a read-only artifact, fits a per-scene hash-grid entropy context with per-channel Laplace distributions, and codes attributes into a Morton-ordered, progressively decodable bitstream. Across 13 scenes on three benchmarks, the post-hoc mode averages 6.0x compression at a -0.03 dB quality change and runs unchanged on third-party checkpoints.

Trained anchor-based Gaussian-splat scenes stay too large for phones and headsets, and existing post-hoc converters saturate early. This scheme treats a trained model as a read-only artifact, fits a per-scene hash-grid entropy context with per-channel Laplace distributions, and codes attributes into a Morton-ordered, progressively decodable bitstream. Across 13 scenes on three benchmarks, the post-hoc mode averages 6.0x compression at a -0.03 dB quality change and runs unchanged on third-party checkpoints.

Vercel Security Review is in public beta for Pro and Enterprise teams: once enabled, every push to a connected GitHub repository is automatically reviewed for vulnerabilities, with findings reported in the Vulnerabilities dashboard. Reviews start from push events rather than pull requests, so direct and force pushes to the production branch are reviewed too, and every project affected by one push is reviewed together. Each finding records the project, environment, branch and commit with a summary, a suggested fix and an AI-generated severity rating; Fix with Vercel Agent has the agent propose a fix and open a pull request that nothing merges until you approve.

Vercel Security Review is in public beta for Pro and Enterprise teams: once enabled, every push to a connected GitHub repository is automatically reviewed for vulnerabilities, with findings reported in the Vulnerabilities dashboard. Reviews start from push events rather than pull requests, so direct and force pushes to the production branch are reviewed too, and every project affected by one push is reviewed together. Each finding records the project, environment, branch and commit with a summary, a suggested fix and an AI-generated severity rating; Fix with Vercel Agent has the agent propose a fix and open a pull request that nothing merges until you approve.

Salesforce AI Research updated the memory policy behind Agentforce Memory so an agent treats a customer's later statement as a correction to a known fact rather than a second fact filed beside it. The evaluation ran on the live Agentforce system, using long multi-session conversations from the public ConvoMem benchmark with specific facts planted along the way, then asked the agent questions that depend on those facts. The updated policy recalls 1.8 times as many facts customers state outright as the policy it replaced, and 2.2 times as many facts that change over time, with nothing underneath the policy changed.

Salesforce AI Research updated the memory policy behind Agentforce Memory so an agent treats a customer's later statement as a correction to a known fact rather than a second fact filed beside it. The evaluation ran on the live Agentforce system, using long multi-session conversations from the public ConvoMem benchmark with specific facts planted along the way, then asked the agent questions that depend on those facts. The updated policy recalls 1.8 times as many facts customers state outright as the policy it replaced, and 2.2 times as many facts that change over time, with nothing underneath the policy changed.

Meta added F3 Lightning on top of F3, the framework that authors, compiles and serves the machine-learning features behind ads ranking models, so each feature is compiled, validated, deployed and rolled back independently. Previously the merged graph was both the execution unit and the release unit, so one bad definition blocked every feature, while early per-feature prototypes raised CPU use by dropping shared computation. F3 Lightning rebuilds a shared execution plan at startup, and comparable extraction features now reach production about 3.6x faster than on prior F3 with no increase in infrastructure compute cost.

Meta added F3 Lightning on top of F3, the framework that authors, compiles and serves the machine-learning features behind ads ranking models, so each feature is compiled, validated, deployed and rolled back independently. Previously the merged graph was both the execution unit and the release unit, so one bad definition blocked every feature, while early per-feature prototypes raised CPU use by dropping shared computation. F3 Lightning rebuilds a shared execution plan at startup, and comparable extraction features now reach production about 3.6x faster than on prior F3 with no increase in infrastructure compute cost.

Cloudflare has added Zero Wire, a public vulnerability intelligence feed in Cloudflare Radar's security section, that condenses critical CVEs in Internet-facing software into briefings for security leaders and engineers deciding whether they are affected. Each briefing carries a plain-language explanation, which deployments to investigate, recommended actions, and links to evidence and vendor guidance, plus a pre-compiled prompt for a read-only exposure assessment by an AI agent. Cloudflare positions it as a complement to vendor advisories and vulnerability databases, not a catalog of every CVE.

Cloudflare has added Zero Wire, a public vulnerability intelligence feed in Cloudflare Radar's security section, that condenses critical CVEs in Internet-facing software into briefings for security leaders and engineers deciding whether they are affected. Each briefing carries a plain-language explanation, which deployments to investigate, recommended actions, and links to evidence and vendor guidance, plus a pre-compiled prompt for a read-only exposure assessment by an AI agent. Cloudflare positions it as a complement to vendor advisories and vulnerability databases, not a catalog of every CVE.

Meta's LEAP predicts a training job's throughput on each candidate hardware type at submission time and ranks the candidates by throughput per unit cost for the cluster scheduler. It replaces a lookup table keyed on model type, trainer count, batch size and hardware type, which returned an estimate for only 13% of job attempts. Over 30 nightly retrains the median accuracy is 77%, and on the 946 jobs the old table covered, LEAP predicts 85.1% within ±20% of measured throughput against 53.2% for the table.

Meta's LEAP predicts a training job's throughput on each candidate hardware type at submission time and ranks the candidates by throughput per unit cost for the cluster scheduler. It replaces a lookup table keyed on model type, trainer count, batch size and hardware type, which returned an estimate for only 13% of job attempts. Over 30 nightly retrains the median accuracy is 77%, and on the 946 jobs the old table covered, LEAP predicts 85.1% within ±20% of measured throughput against 53.2% for the table.

Anthropic has released message threads in beta on the Messages API, keeping long conversations and agent loops on the server so each later request sends only the new messages, which the release notes describe as smaller requests and a lower time to first token. Previously such multi-turn calls had to resend the full history each time. The feature uses the message-threads-2026-08-12 beta header, a thread expires 5 minutes after its last request and is deleted no more than 1 hour after last use, and it is available on the Claude API, Claude Platform on AWS, Claude Models on Google Cloud (Anthropic managed) and Claude in Microsoft Foundry.

Anthropic has released message threads in beta on the Messages API, keeping long conversations and agent loops on the server so each later request sends only the new messages, which the release notes describe as smaller requests and a lower time to first token. Previously such multi-turn calls had to resend the full history each time. The feature uses the message-threads-2026-08-12 beta header, a thread expires 5 minutes after its last request and is deleted no more than 1 hour after last use, and it is available on the Claude API, Claude Platform on AWS, Claude Models on Google Cloud (Anthropic managed) and Claude in Microsoft Foundry.

Vercel has added Security Review to Vercel Agent as a public beta: once a team enables it, every push to a connected GitHub repository triggers a scan. Pattern-matching scanners miss logic flaws such as missing authorization checks; this review compares each push against a base, reads the surrounding code, traces whether the change can be reached and rechecks each finding. Because it scans only the diff, a review finishes in seconds and can cover every preview and production deployment. Findings appear in pull request comments, a dashboard and the CLI, and with permission the agent can validate and push fixes.

Vercel has added Security Review to Vercel Agent as a public beta: once a team enables it, every push to a connected GitHub repository triggers a scan. Pattern-matching scanners miss logic flaws such as missing authorization checks; this review compares each push against a base, reads the surrounding code, traces whether the change can be reached and rechecks each finding. Because it scans only the diff, a review finishes in seconds and can cover every preview and production deployment. Findings appear in pull request comments, a dashboard and the CLI, and with permission the agent can validate and push fixes.

Baseten and Postman announced a partnership that splits enterprise AI agent work in two: Postman's Agent Stack prepares APIs, controls which systems agents can reach and governs their actions, while Baseten supplies high-performance inference, multi-cloud GPU capacity and post-training. Baseten says its model APIs placed in the top two providers on Artificial Analysis for throughput and latency across seven 2026 launches, and cites Harvey post-training a 27B open model to match closed-source frontier performance on its legal agent benchmark.

Baseten and Postman announced a partnership that splits enterprise AI agent work in two: Postman's Agent Stack prepares APIs, controls which systems agents can reach and governs their actions, while Baseten supplies high-performance inference, multi-cloud GPU capacity and post-training. Baseten says its model APIs placed in the top two providers on Artificial Analysis for throughput and latency across seven 2026 launches, and cites Harvey post-training a 27B open model to match closed-source frontier performance on its legal agent benchmark.

CovDiff-PaAno replaces cosine matching in patch-based time-series anomaly detection with regularized Mahalanobis scoring and adds raw, first-difference and second-difference views that share one frozen encoder while each keeps its own memory bank and covariance. In a ten-seed controlled comparison on TSB-AD-M, AUC-ROC rose from 0.7612 for reproduced PaAno to 0.7843 and Range-F1 from 0.4043 to 0.4384; on TSB-AD-U the corresponding scores rose from 0.8532 to 0.8626 and from 0.4822 to 0.4924. The encoder parameter count stays at about 0.397M, but inference time grows 3.47-fold and peak inference memory 10.01-fold.

CovDiff-PaAno replaces cosine matching in patch-based time-series anomaly detection with regularized Mahalanobis scoring and adds raw, first-difference and second-difference views that share one frozen encoder while each keeps its own memory bank and covariance. In a ten-seed controlled comparison on TSB-AD-M, AUC-ROC rose from 0.7612 for reproduced PaAno to 0.7843 and Range-F1 from 0.4043 to 0.4384; on TSB-AD-U the corresponding scores rose from 0.8532 to 0.8626 and from 0.4822 to 0.4924. The encoder parameter count stays at about 0.397M, but inference time grows 3.47-fold and peak inference memory 10.01-fold.

Vercel Firewall's AI Bots managed ruleset now lets sites log or block AI crawlers and AI assistants separately. Previously the two kinds of automated access were handled as one bot-traffic category, so a site could not refuse crawlers that collect content for indexing and model training while keeping the page fetches that answer user actions or search queries. Vercel updates the rules as new bots are added, and the ruleset is available on all plans.

Vercel Firewall's AI Bots managed ruleset now lets sites log or block AI crawlers and AI assistants separately. Previously the two kinds of automated access were handled as one bot-traffic category, so a site could not refuse crawlers that collect content for indexing and model training while keeping the page fetches that answer user actions or search queries. Vercel updates the rules as new bots are added, and the ruleset is available on all plans.

Spotify released portal-ai-plugins, a plugin set that brings the Portal CLI's authentication setup, software-catalog search, service briefings, diagnostics and action invocation into coding-agent environments including Claude Code, Codex, Cursor and pi. The workflows query the catalog and documentation in natural language and guard Portal actions with built-in help, dry-run and confirmation; the same marketplace also ships shunt, currently Claude Code only, which routes I/O-heavy work such as bulk file reads to cheaper AiKA worker models.

Spotify released portal-ai-plugins, a plugin set that brings the Portal CLI's authentication setup, software-catalog search, service briefings, diagnostics and action invocation into coding-agent environments including Claude Code, Codex, Cursor and pi. The workflows query the catalog and documentation in natural language and guard Portal actions with built-in help, dry-run and confirmation; the same marketplace also ships shunt, currently Claude Code only, which routes I/O-heavy work such as bulk file reads to cheaper AiKA worker models.