AI and science frontiers · 2026-10-15

A microgram-scale acoustic mode of a high-overtone bulk acoustic wave resonator is prepared in Fock states of up to seven phonons to estimate the displacement of a weak resonant signal of unknown phase. Such non-Gaussian quantum probes had shown metrological advantage only in lighter systems such as trapped ions, while squeezed-state advantages were restricted to a known quadrature. Two independent analyses of the measured phonon-number distributions agree on a Fisher information five times the vacuum value, beyond any classical probe of arbitrary energy; a neural network trained directly on calibration records reproduces the same sensitivity with half the systematic bias.

A microgram-scale acoustic mode of a high-overtone bulk acoustic wave resonator is prepared in Fock states of up to seven phonons to estimate the displacement of a weak resonant signal of unknown phase. Such non-Gaussian quantum probes had shown metrological advantage only in lighter systems such as trapped ions, while squeezed-state advantages were restricted to a known quadrature. Two independent analyses of the measured phonon-number distributions agree on a Fisher information five times the vacuum value, beyond any classical probe of arbitrary energy; a neural network trained directly on calibration records reproduces the same sensitivity with half the systematic bias.

Adult endotracheal tube AI models have been hard to apply to newborns, where annotated imaging is scarce and privacy rules block data sharing. An international retrospective study across 30 neonatal intensive care units in 20 countries and territories used continual learning: the model passes from hospital to hospital, and each site retrains it on 50 local radiographs inside its own environment without sharing patient data. The final model averaged 0.67 cm combined endotracheal tube tip and carina localization error, versus 1.29 cm for hospital-specific fine-tuning and 4.30 cm for the adult model used directly, improving localization at 29 of 30 sites and deep-intubation detection at 27 of 30.

Adult endotracheal tube AI models have been hard to apply to newborns, where annotated imaging is scarce and privacy rules block data sharing. An international retrospective study across 30 neonatal intensive care units in 20 countries and territories used continual learning: the model passes from hospital to hospital, and each site retrains it on 50 local radiographs inside its own environment without sharing patient data. The final model averaged 0.67 cm combined endotracheal tube tip and carina localization error, versus 1.29 cm for hospital-specific fine-tuning and 4.30 cm for the adult model used directly, improving localization at 29 of 30 sites and deep-intubation detection at 27 of 30.

Detecting a watermark in large-model text used to require the generating side's tokenizer or semantic embedding pipeline, which the text itself does not reveal, so text of unknown origin was hard to verify. ByteMark defines its signal on the UTF-8 byte stream of the published text; participating models embed it under one public specification and key, and a verifier needs only the text and the key. In experiments, replacing the detector's tokenizer or embedding model sharply degraded token-level and sentence-level baselines, while ByteMark's detection rate was unchanged, and it stayed detectable through a relay of models and after participating rewriters re-embedded the watermark.

Detecting a watermark in large-model text used to require the generating side's tokenizer or semantic embedding pipeline, which the text itself does not reveal, so text of unknown origin was hard to verify. ByteMark defines its signal on the UTF-8 byte stream of the published text; participating models embed it under one public specification and key, and a verifier needs only the text and the key. In experiments, replacing the detector's tokenizer or embedding model sharply degraded token-level and sentence-level baselines, while ByteMark's detection rate was unchanged, and it stayed detectable through a relay of models and after participating rewriters re-embedded the watermark.

The internal choices inside a visual-head discovery score reshape which heads it returns in vision-language models. On Qwen3-VL-8B-Instruct, fully crossing query type, discovery corpus and the token positions whose attention is read, and testing all 12 head-sets with steerability and necessity interventions, only necessity follows the attention source: query-span head-sets are more necessary than last-token ones, and the two referring plus query-span sets lie 17–19 standard deviations above 250 layer-matched random sets. At the default strength, no discovered head-set steers better than its layer-matched random sets.

The internal choices inside a visual-head discovery score reshape which heads it returns in vision-language models. On Qwen3-VL-8B-Instruct, fully crossing query type, discovery corpus and the token positions whose attention is read, and testing all 12 head-sets with steerability and necessity interventions, only necessity follows the attention source: query-span head-sets are more necessary than last-token ones, and the two referring plus query-span sets lie 17–19 standard deviations above 250 layer-matched random sets. At the default strength, no discovered head-set steers better than its layer-matched random sets.

State-space model decoding previously read and wrote the full recurrent state at every step, making state traffic the memory-bandwidth bottleneck and leaving speculative decoding without cheap rollback. ReplaySSM represents the state as a checkpoint plus a buffer of recent inputs, writes the checkpoint only when the buffer fills, and computes outputs directly on most steps. Implemented in vLLM for Mamba-2 and Gated DeltaNet, it speeds up standard decoding over a tuned baseline at serving batch size 256, reaches about the throughput of vLLM's speculative decoding on two 120B-class models, and separate state precisions reduce output KL divergence and downstream accuracy losses.

State-space model decoding previously read and wrote the full recurrent state at every step, making state traffic the memory-bandwidth bottleneck and leaving speculative decoding without cheap rollback. ReplaySSM represents the state as a checkpoint plus a buffer of recent inputs, writes the checkpoint only when the buffer fills, and computes outputs directly on most steps. Implemented in vLLM for Mamba-2 and Gated DeltaNet, it speeds up standard decoding over a tuned baseline at serving batch size 256, reaches about the throughput of vLLM's speculative decoding on two 120B-class models, and separate state precisions reduce output KL divergence and downstream accuracy losses.

Cardiac MRI reading no longer rests only on ejection fraction and other predefined measurements: a healthy cardiac digital twin generated from each person's age, sex and body habitus yields a deviation score that still improves atrial fibrillation risk stratification among people with preserved ejection fraction.

Cardiac MRI reading no longer rests only on ejection fraction and other predefined measurements: a healthy cardiac digital twin generated from each person's age, sex and body habitus yields a deviation score that still improves atrial fibrillation risk stratification among people with preserved ejection fraction.

Linux readahead infers the next page from recent file offsets, which fails when accesses follow workload structure such as sparse startup reads, B-tree key-order traversal, or tensor-layout model loading. PredictIO separates prediction from execution: the kernel delivers observed accesses and application hints to user-space policies, which emit uniform page-range requests that a kernel backend validates, bounds, and issues as asynchronous prefetch I/O. Across application launch, SQLite range scans, and llama.cpp model loading, latency drops by 5.3–26.3%, scans speed up by up to 2.99, and model loading drops by 8.8–47.9%.

Linux readahead infers the next page from recent file offsets, which fails when accesses follow workload structure such as sparse startup reads, B-tree key-order traversal, or tensor-layout model loading. PredictIO separates prediction from execution: the kernel delivers observed accesses and application hints to user-space policies, which emit uniform page-range requests that a kernel backend validates, bounds, and issues as asynchronous prefetch I/O. Across application launch, SQLite range scans, and llama.cpp model loading, latency drops by 5.3–26.3%, scans speed up by up to 2.99, and model loading drops by 8.8–47.9%.

Alibaba researchers introduce Rollout-Time Intervention, which corrects the decisive error step using reward-side rubrics whenever an entire rollout group fails, letting agent RL learn from successful trajectories the policy could not explore on its own and raising Navigation Pass@1 from GRPO's 35.33 to 57.33.

Alibaba researchers introduce Rollout-Time Intervention, which corrects the decisive error step using reward-side rubrics whenever an entire rollout group fails, letting agent RL learn from successful trajectories the policy could not explore on its own and raising Navigation Pass@1 from GRPO's 35.33 to 57.33.

RemoTAP-Hybrid combines multiple collaborated neural networks with a full physics inversion for global aerosol retrieval from the SPEXone multi-angle polarimeter, improving AOD, Angstrom Exponent and fine-mode AOD accuracy over land and ocean while cutting computation time by 43% and 32%.

RemoTAP-Hybrid combines multiple collaborated neural networks with a full physics inversion for global aerosol retrieval from the SPEXone multi-angle polarimeter, improving AOD, Angstrom Exponent and fine-mode AOD accuracy over land and ocean while cutting computation time by 43% and 32%.

Low-dose CT denoising can now be trained from adjacent slices alone: a matching network estimates inter-slice deformation and corrects the denoiser output, reaching 40.08 dB PSNR on Mayo 2016 and, for the first time, image-domain self-supervised low-dose cone-beam CT denoising.

Low-dose CT denoising can now be trained from adjacent slices alone: a matching network estimates inter-slice deformation and corrects the denoiser output, reaching 40.08 dB PSNR on Mayo 2016 and, for the first time, image-domain self-supervised low-dose cone-beam CT denoising.

In multi-source learning the deployment population's source composition is unknown, and optimizing against the worst source mixture can be overly conservative. PILLAR regularizes adversarial source weights toward a user-specified reference distribution while learning a shared low-rank matrix, solving the resulting nonconvex-concave problem by bifactored gradient descent with mirror ascent, with linear convergence and an error bound that separates low-rank estimation from adversarial reweighting. Simulations and an electronic health record study of suicide-attempt risk show improved worst-source and transported predictive accuracy.

In multi-source learning the deployment population's source composition is unknown, and optimizing against the worst source mixture can be overly conservative. PILLAR regularizes adversarial source weights toward a user-specified reference distribution while learning a shared low-rank matrix, solving the resulting nonconvex-concave problem by bifactored gradient descent with mirror ascent, with linear convergence and an error bound that separates low-rank estimation from adversarial reweighting. Simulations and an electronic health record study of suicide-attempt risk show improved worst-source and transported predictive accuracy.

Step 5 Preview treats agentic workloads as the starting point for model design: it uses a narrow-but-deep stack of 92 transformer blocks, interleaves efficient and sparse full attention at 3:1, and extends the full-attention path to contexts up to 1M tokens with Sparse GQA. Its total parameter count is under one-quarter and its active count under one-third of trillion-scale models such as Qwen3.8-Max and Kimi K3, and it reports long-horizon agentic capability comparable to those counterparts. Agent-native data synthesis and long-rollout RL infrastructure accompany the model, yielding million-level verifiable environments.

Step 5 Preview treats agentic workloads as the starting point for model design: it uses a narrow-but-deep stack of 92 transformer blocks, interleaves efficient and sparse full attention at 3:1, and extends the full-attention path to contexts up to 1M tokens with Sparse GQA. Its total parameter count is under one-quarter and its active count under one-third of trillion-scale models such as Qwen3.8-Max and Kimi K3, and it reports long-horizon agentic capability comparable to those counterparts. Agent-native data synthesis and long-rollout RL infrastructure accompany the model, yielding million-level verifiable environments.

Google released Gemini 4 Argon for developers and enterprises, and in Chrome's roughly 40-million-line codebase it helped complete an architectural refactor across hundreds of files, compressing months of work into weeks.

Google released Gemini 4 Argon for developers and enterprises, and in Chrome's roughly 40-million-line codebase it helped complete an architectural refactor across hundreds of files, compressing months of work into weeks.

SDO, an operator agent for Kubernetes applications, distills each resolved incident into reusable operational memory: parameterized playbooks that record diagnosis and repair steps, non-LLM detectors that decide from cluster state which playbook applies, and an architecture summary that binds playbook roles to the current application. Across 11 open-source microservice applications, SDO cuts resolution time on recurring faults by 52%, with median resolution falling from 584 s to 282 s and tokens from 5.4M to 2.2M; when deploying applications from source, operational memory reduces deployment time and token use by 47.5%.

SDO, an operator agent for Kubernetes applications, distills each resolved incident into reusable operational memory: parameterized playbooks that record diagnosis and repair steps, non-LLM detectors that decide from cluster state which playbook applies, and an architecture summary that binds playbook roles to the current application. Across 11 open-source microservice applications, SDO cuts resolution time on recurring faults by 52%, with median resolution falling from 584 s to 282 s and tokens from 5.4M to 2.2M; when deploying applications from source, operational memory reduces deployment time and token use by 47.5%.

Cognitive differences in psychotic disorders have mostly been grouped by diagnosis or drug dose, leaving the medication's own contribution hard to isolate. A cross-sectional study converted prescriptions for 1,067 patients into individualised occupancy profiles across 49 drugs and 32 receptor targets, then used gradient boosting to predict cognition. Shuffling the receptor profiles within diagnosis roughly halved the pharmacological contribution, with negative signal concentrated in a cluster of alpha2-adrenergic, DRD4, DRD5 and several serotonergic receptors, while dopamine transporter SLC6A3 occupancy tracked better cognition.

Cognitive differences in psychotic disorders have mostly been grouped by diagnosis or drug dose, leaving the medication's own contribution hard to isolate. A cross-sectional study converted prescriptions for 1,067 patients into individualised occupancy profiles across 49 drugs and 32 receptor targets, then used gradient boosting to predict cognition. Shuffling the receptor profiles within diagnosis roughly halved the pharmacological contribution, with negative signal concentrated in a cluster of alpha2-adrenergic, DRD4, DRD5 and several serotonergic receptors, while dopamine transporter SLC6A3 occupancy tracked better cognition.

A ViT-B/3D encoder trained on 113,611 thoraco-abdomino-pelvic CT volumes, TAP-JEPA, matches state-of-the-art CT encoders on thoracic and abdominal classification benchmarks; replaying its exact masking, teacher and predictor shows hidden-target prediction lowers mean Skill from 0.388 to 0.271 for pathology but only from 0.921 to 0.873 for anatomy. Pathology moves 34.7% toward matched controls along the teacher direction while controls move 13.6% toward pathology, with larger shifts farther from visible context, and deleting all final-layer pathology tokens changes the decision margin by at most 0.45% in four of seven scan-level probes.

A ViT-B/3D encoder trained on 113,611 thoraco-abdomino-pelvic CT volumes, TAP-JEPA, matches state-of-the-art CT encoders on thoracic and abdominal classification benchmarks; replaying its exact masking, teacher and predictor shows hidden-target prediction lowers mean Skill from 0.388 to 0.271 for pathology but only from 0.921 to 0.873 for anatomy. Pathology moves 34.7% toward matched controls along the teacher direction while controls move 13.6% toward pathology, with larger shifts farther from visible context, and deleting all final-layer pathology tokens changes the decision margin by at most 0.45% in four of seven scan-level probes.

Protein design has long targeted a single static structure, while natural metamorphic proteins occupy multiple folded states under the same conditions. Using RFdiffusion to partially remodel the termini of 93 native domains yielded 2,291 native-alternative conformation pairs; two sequence-design routes then made one sequence compatible with both backbones, and AlphaFold3 with BioEmu selected state-stabilizing mutations. The designs showed broadened and missing NMR crosspeaks, indicating microsecond-to-millisecond exchange in the metamorphic region; two or three mutations shifted the equilibrium toward either state, yielding crystal and NMR structures consistent with the design models.

Protein design has long targeted a single static structure, while natural metamorphic proteins occupy multiple folded states under the same conditions. Using RFdiffusion to partially remodel the termini of 93 native domains yielded 2,291 native-alternative conformation pairs; two sequence-design routes then made one sequence compatible with both backbones, and AlphaFold3 with BioEmu selected state-stabilizing mutations. The designs showed broadened and missing NMR crosspeaks, indicating microsecond-to-millisecond exchange in the metamorphic region; two or three mutations shifted the equilibrium toward either state, yielding crystal and NMR structures consistent with the design models.

ClawLoop closes task synthesis, training, and failure diagnosis into one loop so harness-based agents improve without an external teacher model, beating the one-shot ClawGym pipeline on SkillsBench, Claw-Eval-Live, and ClawGym-Bench.

ClawLoop closes task synthesis, training, and failure diagnosis into one loop so harness-based agents improve without an external teacher model, beating the one-shot ClawGym pipeline on SkillsBench, Claw-Eval-Live, and ClawGym-Bench.

In head and neck squamous cell carcinoma the partial EMT gene list is expressed by both malignant cells and fibroblasts, yet it has been scored only in bulk tumours. Dividing its 100 genes by the cells that express them shows bulk co-expression is dominated by the stromal arm, so standard pEMT scores mainly measure stroma; a single-cell pseudobulk classifier and a 12-gene malignant core recover the malignant arm in any bulk transcriptome. Among 40 patients given platinum plus cetuximab, both malignant measures were higher in those with long progression-free survival (area under the curve 0.78 and 0.82) and the stromal core was not, but the malignant arm could not be separated from the Basal subtype, which discriminated better (0.93).

In head and neck squamous cell carcinoma the partial EMT gene list is expressed by both malignant cells and fibroblasts, yet it has been scored only in bulk tumours. Dividing its 100 genes by the cells that express them shows bulk co-expression is dominated by the stromal arm, so standard pEMT scores mainly measure stroma; a single-cell pseudobulk classifier and a 12-gene malignant core recover the malignant arm in any bulk transcriptome. Among 40 patients given platinum plus cetuximab, both malignant measures were higher in those with long progression-free survival (area under the curve 0.78 and 0.82) and the stromal core was not, but the malignant arm could not be separated from the Basal subtype, which discriminated better (0.93).

In Bayesian LoRA fine-tuning of LLMs and vision-language models, minibatch gradient errors are heavy-tailed and grow with distance from the final iterate, and a robust estimator combining geometric-median centering with residual clipping, together with adaptive step sizes, retains non-asymptotic Fisher-information guarantees on non-log-concave targets.

In Bayesian LoRA fine-tuning of LLMs and vision-language models, minibatch gradient errors are heavy-tailed and grow with distance from the final iterate, and a robust estimator combining geometric-median centering with residual clipping, together with adaptive step sizes, retains non-asymptotic Fisher-information guarantees on non-log-concave targets.

Eco-RNA turns inference depth into a selectable resource: one 80M-parameter checkpoint applies a shared core repeatedly, lifting aggregate absolute Spearman correlation on the 70-assay RNAGym fitness benchmark from 0.242 to 0.286 and passing Evo 2 40B at four iterations.

Eco-RNA turns inference depth into a selectable resource: one 80M-parameter checkpoint applies a shared core repeatedly, lifting aggregate absolute Spearman correlation on the 70-assay RNAGym fitness benchmark from 0.242 to 0.286 and passing Evo 2 40B at four iterations.

An LLM-directed excavator can execute mechanically feasible motion whose destination or authority no longer matches the operator's approval. The work binds authorization to the exact motion: a one-use lease ties the action, generated motion, mission contract and admission tolerance together, atomic dispatch rechecks current authority and measured joint drift, and revocation requests a carry-pitch quintic stop. Across a 672-case interface suite the gateway admitted all 120 benign cases and produced zero unauthorized dispatches in the attack suite; in synchronous simulation, dump-phase revocation cut maximum closed-site deposition from 1.23 to 0.33 cubic meters.

An LLM-directed excavator can execute mechanically feasible motion whose destination or authority no longer matches the operator's approval. The work binds authorization to the exact motion: a one-use lease ties the action, generated motion, mission contract and admission tolerance together, atomic dispatch rechecks current authority and measured joint drift, and revocation requests a carry-pitch quintic stop. Across a 672-case interface suite the gateway admitted all 120 benign cases and produced zero unauthorized dispatches in the attack suite; in synchronous simulation, dump-phase revocation cut maximum closed-site deposition from 1.23 to 0.33 cubic meters.

GLOBE compresses three-dimensional beam–field states into a compact latent space, learns the dominant evolution with a control-conditioned neural operator, and captures fine-scale residuals with a generative model, then applies parameter-free constraint operators after decoding. On two high-resolution 3D benchmarks, bunched-beam transport and plasma-wakefield acceleration, relative error is 9% and 13% lower than the strongest baseline DPOT, mid-frequency spectral error is 54% and 20% lower, and hours of particle simulation become about a minute per rollout.

GLOBE compresses three-dimensional beam–field states into a compact latent space, learns the dominant evolution with a control-conditioned neural operator, and captures fine-scale residuals with a generative model, then applies parameter-free constraint operators after decoding. On two high-resolution 3D benchmarks, bunched-beam transport and plasma-wakefield acceleration, relative error is 9% and 13% lower than the strongest baseline DPOT, mid-frequency spectral error is 54% and 20% lower, and hours of particle simulation become about a minute per rollout.

Siemens Fuse used NVIDIA NeMo Gym to build an evaluation loop for the Solido Characterization Suite agent that grades final Liberty files, raising the mean composite verification score on standard-cell characterization tasks from about 55% to about 95% while cutting agent turns from 155 to 53 and output tokens from about 81K to 29K. Previously such multistage flows could only check whether individual tool calls succeeded, and an early setup error could propagate silently downstream; now a composite reward combining coverage, validity, relative validation, integrity and the in-tolerance fraction scores the final artifact of each rollout, and those findings guide rewrites of skills and MCP tool descriptions.

Siemens Fuse used NVIDIA NeMo Gym to build an evaluation loop for the Solido Characterization Suite agent that grades final Liberty files, raising the mean composite verification score on standard-cell characterization tasks from about 55% to about 95% while cutting agent turns from 155 to 53 and output tokens from about 81K to 29K. Previously such multistage flows could only check whether individual tool calls succeeded, and an early setup error could propagate silently downstream; now a composite reward combining coverage, validity, relative validation, integrity and the in-tolerance fraction scores the final artifact of each rollout, and those findings guide rewrites of skills and MCP tool descriptions.

In long-horizon autoregressive reinforcement learning, local policy mismatch compounds along the trajectory into state distribution shift, destabilizing training and eventually collapsing it as the horizon grows. Trajectory-Aware Policy Optimization derives a state-distribution correction from the off-policy policy gradient and writes it into the objective, estimates it with a geometric mean to cut variance, and tightens the clipping boundary with the length of the preceding trajectory. Across eight competition-level math and code benchmarks it improves over GRPO by up to 48% on math and 12% on code on average, with training scores rising steadily while KL divergence stays controlled.

In long-horizon autoregressive reinforcement learning, local policy mismatch compounds along the trajectory into state distribution shift, destabilizing training and eventually collapsing it as the horizon grows. Trajectory-Aware Policy Optimization derives a state-distribution correction from the off-policy policy gradient and writes it into the objective, estimates it with a geometric mean to cut variance, and tightens the clipping boundary with the length of the preceding trajectory. Across eight competition-level math and code benchmarks it improves over GRPO by up to 48% on math and 12% on code on average, with training scores rising steadily while KL divergence stays controlled.

Multi-modal image fusion has long traded visual quality against semantic attributes because the two optimization goals conflict. AD-SRF separates them: a transformer-based restorative visual fusion network builds a clear baseline, then an attack-defense game has an attacker mine semantic weaknesses while a defender generates an imperceptible semantic map overlaid back onto the fused image. On MFNet and FMB, it ranks first in four and five of six visual metrics, with segmentation mIoU 1.5 and 1.73 above the runner-up.

Multi-modal image fusion has long traded visual quality against semantic attributes because the two optimization goals conflict. AD-SRF separates them: a transformer-based restorative visual fusion network builds a clear baseline, then an attack-defense game has an attacker mine semantic weaknesses while a defender generates an imperceptible semantic map overlaid back onto the fused image. On MFNet and FMB, it ranks first in four and five of six visual metrics, with segmentation mIoU 1.5 and 1.73 above the runner-up.

Across five vision-centric tasks, adding labeled demonstrations to the context sharply raises accuracy on counterfactual examples that conflict with training priors while leaving prior-aligned accuracy nearly unchanged, indicating the models already have the visual capability and were being overridden by learned regularities.

Across five vision-centric tasks, adding labeled demonstrations to the context sharply raises accuracy on counterfactual examples that conflict with training priors while leaving prior-aligned accuracy nearly unchanged, indicating the models already have the visual capability and were being overridden by learned regularities.

RES-MR splits medication recommendation into two steps, first generating a checkable diagnostic reasoning chain and then adjusting the prescription to each patient's own risk boundary, reaching a 7.7% higher Jaccard than FLAME on MIMIC-III with lower DDI and RIS.

RES-MR splits medication recommendation into two steps, first generating a checkable diagnostic reasoning chain and then adjusting the prescription to each patient's own risk boundary, reaching a 7.7% higher Jaccard than FLAME on MIMIC-III with lower DDI and RIS.

Gemini 4 Argon puts physical understanding, code generation and visual-semantic reasoning into one model so robot tasks can be practiced in simulation first. On the Embodied Reasoning (ERQA) benchmark, which tests physical and spatial problems in 3D environments, it scores 87% using pure reasoning, above other frontier models. It can also reconstruct calibrated simulation twins from robot camera feeds, egocentric human video or hand-drawn sketches, and generate closed-loop control and training data from them.

Gemini 4 Argon puts physical understanding, code generation and visual-semantic reasoning into one model so robot tasks can be practiced in simulation first. On the Embodied Reasoning (ERQA) benchmark, which tests physical and spatial problems in 3D environments, it scores 87% using pure reasoning, above other frontier models. It can also reconstruct calibrated simulation twins from robot camera feeds, egocentric human video or hand-drawn sketches, and generate closed-loop control and training data from them.

Federated aggregation no longer rewrites the global optimization direction to gain fairness. Earlier methods reweight clients, correct conflicting gradients, or treat client losses as symmetric multi-objective common descent, all of which change the global update direction. FedPCD keeps the conventional FedAvg aggregate as the primary objective and individual client updates as secondary constraints, applying only the minimum correction needed when a constraint is violated; across CIFAR-10, CIFAR-100 and Fashion-MNIST it reaches the highest global accuracy in all 18 configurations and the highest worst-10% client accuracy in 17 of them.

Federated aggregation no longer rewrites the global optimization direction to gain fairness. Earlier methods reweight clients, correct conflicting gradients, or treat client losses as symmetric multi-objective common descent, all of which change the global update direction. FedPCD keeps the conventional FedAvg aggregate as the primary objective and individual client updates as secondary constraints, applying only the minimum correction needed when a constraint is violated; across CIFAR-10, CIFAR-100 and Fashion-MNIST it reaches the highest global accuracy in all 18 configurations and the highest worst-10% client accuracy in 17 of them.

Teams of coding agents generally beat a single agent at rebuilding programs, but at four to seven times the cost. A controlled study across 200 ProgramBench programs finds that putting cheaper models only on exploration and planning comes within 2.7 points of an all-strong-model team that delegates every stage, at roughly a fifth of that team's cost, because discovering the original program's behavior is the real bottleneck.

Teams of coding agents generally beat a single agent at rebuilding programs, but at four to seven times the cost. A controlled study across 200 ProgramBench programs finds that putting cheaper models only on exploration and planning comes within 2.7 points of an all-strong-model team that delegates every stage, at roughly a fifth of that team's cost, because discovering the original program's behavior is the real bottleneck.

Claude Platform now exports tracing data for Messages API requests and Claude Managed Agents sessions as OpenTelemetry spans, so developers can view those calls in their own observability backend.

Claude Platform now exports tracing data for Messages API requests and Claude Managed Agents sessions as OpenTelemetry spans, so developers can view those calls in their own observability backend.

On strictly paired vulnerable and patched C/C++ functions, whitened CodeBERT layer-3 embeddings with nearest-neighbour distance separate unfamiliar samples from familiar ones at a mean OOD AUROC of 0.739, and expose a cross-project patch false-positive rate as high as 0.507.

On strictly paired vulnerable and patched C/C++ functions, whitened CodeBERT layer-3 embeddings with nearest-neighbour distance separate unfamiliar samples from familiar ones at a mean OOD AUROC of 0.739, and expose a cross-project patch false-positive rate as high as 0.507.

A binDCT block transform runs directly on a programmable switch, filtering noisy sensor streams on the forwarding path before in-network machine learning, which raises switch-side classification MCC from 0.35 to 0.42 and matches the server-side value.

A binDCT block transform runs directly on a programmable switch, filtering noisy sensor streams on the forwarding path before in-network machine learning, which raises switch-side classification MCC from 0.35 to 0.42 and matches the server-side value.

Adaptive task allocation previously came with guarantees only when worker computation times stay fixed, yet in shared clusters resource contention, background jobs, or network congestion can turn a fast worker into a straggler and invalidate the allocation. Sliding Window Adaptive Task Allocation (SW-ATA) estimates each worker's delay mean from a recent window of observations and picks allocations from lower confidence bounds, so it can track fast workers while the delay distribution changes from round to round. On synthetic drifting-delay distributed stochastic optimization, it clearly beats the original adaptive allocation and uniform allocation, stays close to an oracle that knows the true means, and uses the least worker time.

Adaptive task allocation previously came with guarantees only when worker computation times stay fixed, yet in shared clusters resource contention, background jobs, or network congestion can turn a fast worker into a straggler and invalidate the allocation. Sliding Window Adaptive Task Allocation (SW-ATA) estimates each worker's delay mean from a recent window of observations and picks allocations from lower confidence bounds, so it can track fast workers while the delay distribution changes from round to round. On synthetic drifting-delay distributed stochastic optimization, it clearly beats the original adaptive allocation and uniform allocation, stays close to an oracle that knows the true means, and uses the least worker time.

Time series foundation models now accept covariates, yet how calendar features are encoded has been treated as an implementation detail. Across four models and nine datasets shaped by weekends and holidays, seven calendar encodings were compared in a zero-shot rolling evaluation: encoding choice changed accuracy in every model, with within-model MASE gaps up to 9.9% on non-holidays and 83% on holidays, and the best encoding differed by model. SHAP attribution locates the gains around holidays, where signals cannot be inferred from target history.

Time series foundation models now accept covariates, yet how calendar features are encoded has been treated as an implementation detail. Across four models and nine datasets shaped by weekends and holidays, seven calendar encodings were compared in a zero-shot rolling evaluation: encoding choice changed accuracy in every model, with within-model MASE gaps up to 9.9% on non-holidays and 83% on holidays, and the best encoding differed by model. SHAP attribution locates the gains around holidays, where signals cannot be inferred from target history.

Deployment generalization for speaker verification now has a shared public evaluation: a fixed core organizes 16 public sources into 26 trial sets and 11 scenario groups with 9,150,366 verification trials, frozen target/non-target keys, and an online leaderboard that reports trial-set breakdowns, scenario-group metrics, and macro averages. Across 18 public checkpoints, systems with similar macro EER show different scenario profiles, strength in one scenario does not guarantee strength in another, and trial sets within the same scenario group still expose source and protocol gaps.

Deployment generalization for speaker verification now has a shared public evaluation: a fixed core organizes 16 public sources into 26 trial sets and 11 scenario groups with 9,150,366 verification trials, frozen target/non-target keys, and an online leaderboard that reports trial-set breakdowns, scenario-group metrics, and macro averages. Across 18 public checkpoints, systems with similar macro EER show different scenario profiles, strength in one scenario does not guarantee strength in another, and trial sets within the same scenario group still expose source and protocol gaps.

Evo-RL filters deployment experience by estimated advantage before policy fine-tuning, raising attempt success rate on a real PiperX high-precision screw-insertion task from RECAP's 57.78% to 88.24%.

Evo-RL filters deployment experience by estimated advantage before policy fine-tuning, raising attempt success rate on a real PiperX high-precision screw-insertion task from RECAP's 57.78% to 88.24%.

Inspect Any Crack converts crack segmentation masks into millimeter-scale widths and a continuous severity score, giving 7.33% mean relative width error at 41 caliper-referenced locations so inspection images yield condition judgments that can be checked against standards.

Inspect Any Crack converts crack segmentation masks into millimeter-scale widths and a continuous severity score, giving 7.33% mean relative width error at 41 caliper-referenced locations so inspection images yield condition judgments that can be checked against standards.

TagBO treats candidate rankings produced by a large language model as coarse task-aware priors and fuses them with local structural smoothing into a similarity graph, reaching high-quality candidates in fewer evaluations under a fixed budget on six mixed-variable chemistry and materials optimization tasks.

TagBO treats candidate rankings produced by a large language model as coarse task-aware priors and fuses them with local structural smoothing into a similarity graph, reaching high-quality candidates in fewer evaluations under a fixed budget on six mixed-variable chemistry and materials optimization tasks.

Across 928 archived first-quarter forecasts for 232 assets in four iPulse AI batches, the highest-disagreement quartile had 21.05 percentage points of mean absolute error, yet a no-change baseline rose almost identically to 21.15, so disagreement added no resolved reliability beyond visible asset risk.

Across 928 archived first-quarter forecasts for 232 assets in four iPulse AI batches, the highest-disagreement quartile had 21.05 percentage points of mean absolute error, yet a no-change baseline rose almost identically to 21.15, so disagreement added no resolved reliability beyond visible asset risk.

The spin Hamiltonian parameters of α-RuCl₃ no longer come only from ab initio calculations: Bayesian optimization fits experimental magnetization curves along the b and c axes while optimizing five parameters, the Kitaev interaction K, the off-diagonal interactions Γ and Γ′, the Heisenberg interaction J, and the c-axis g-factor g_c. The cost-minimizing set is (K, Γ, Γ′, J, g_c) = (−6.0, 7.5, −0.3, −1.75, 2.3), with exchange couplings in meV. The cost function is insensitive to the absolute value of K, so magnetization data alone do not fix the energy scale of the Kitaev coupling; it depends only weakly on Γ′ and J, while the optimization favors a large positive Γ.

The spin Hamiltonian parameters of α-RuCl₃ no longer come only from ab initio calculations: Bayesian optimization fits experimental magnetization curves along the b and c axes while optimizing five parameters, the Kitaev interaction K, the off-diagonal interactions Γ and Γ′, the Heisenberg interaction J, and the c-axis g-factor g_c. The cost-minimizing set is (K, Γ, Γ′, J, g_c) = (−6.0, 7.5, −0.3, −1.75, 2.3), with exchange couplings in meV. The cost function is insensitive to the absolute value of K, so magnetization data alone do not fix the energy scale of the Kitaev coupling; it depends only weakly on Γ′ and J, while the optimization favors a large positive Γ.

A clear comparison has emerged for AI substitutes for high-resolution ocean simulation: evolving states on a 1-degree grid and then spatially downscaling recovers only large-scale patterns, while learning dynamical evolution directly on a 0.1-degree grid preserves mesoscale energy, spectral characteristics, and frontal gradients. The work also provides a full-depth 1-degree emulator and a 0.1-degree sea surface temperature emulator, both stable over long-term free-running simulations without significant drift, giving the comparison reliable baselines.

A clear comparison has emerged for AI substitutes for high-resolution ocean simulation: evolving states on a 1-degree grid and then spatially downscaling recovers only large-scale patterns, while learning dynamical evolution directly on a 0.1-degree grid preserves mesoscale energy, spectral characteristics, and frontal gradients. The work also provides a full-depth 1-degree emulator and a 0.1-degree sea surface temperature emulator, both stable over long-term free-running simulations without significant drift, giving the comparison reliable baselines.

The Burer–Monteiro factorized objective for low-rank matrix minimization is shown to be generalized smooth with exponent 2/3, and adaptive gradient descent and SafeAdam built on that property match or beat Dyn-NoLips on SymNMF text and image benchmarks without tuning relative-smoothness kernel parameters.

The Burer–Monteiro factorized objective for low-rank matrix minimization is shown to be generalized smooth with exponent 2/3, and adaptive gradient descent and SafeAdam built on that property match or beat Dyn-NoLips on SymNMF text and image benchmarks without tuning relative-smoothness kernel parameters.

Deciding whether an auxiliary time series genuinely improves prediction of a target has relied mainly on error-magnitude measures such as RMSE or Granger gain. The new construction takes matched target-only and target-plus-source multi-horizon forecasts, computes Shannon entropy of the signed forecast-error distribution at each horizon, and turns the change of that entropy profile across horizons into a directed descriptor. Across four equal-power hidden nonlinear channels the noise-averaged response-sensitivity statistic is 29.00, 27.47, 23.22 and 28.67, versus 14.09, 10.68, 12.67 and 12.71 for mean log-RMSE gain, and the response stays positive in all 108 tested combinations of window length, history length and maximum forecast horizon.

Deciding whether an auxiliary time series genuinely improves prediction of a target has relied mainly on error-magnitude measures such as RMSE or Granger gain. The new construction takes matched target-only and target-plus-source multi-horizon forecasts, computes Shannon entropy of the signed forecast-error distribution at each horizon, and turns the change of that entropy profile across horizons into a directed descriptor. Across four equal-power hidden nonlinear channels the noise-averaged response-sensitivity statistic is 29.00, 27.47, 23.22 and 28.67, versus 14.09, 10.68, 12.67 and 12.71 for mean log-RMSE gain, and the response stays positive in all 108 tested combinations of window length, history length and maximum forecast horizon.

Google Cloud has opened its new frontier model, Gemini 4 Argon, to enterprise users, positioned to take over multi-step work that used to consume hours or days and return finished, usable output. It is trained to read long documents mixing text, tables, charts and images, consolidate findings from many sources, keep going until a long task is done, and pause for confirmation before irreversible steps. Google's comparisons include an overnight autonomous run on its open-source Agent Development Kit that cut in-memory session prep time by 40% and cold imports by 21%, and a Chrome team refactor across hundreds of files in Chromium's 40-million-line codebase finished in weeks instead of months.

Google Cloud has opened its new frontier model, Gemini 4 Argon, to enterprise users, positioned to take over multi-step work that used to consume hours or days and return finished, usable output. It is trained to read long documents mixing text, tables, charts and images, consolidate findings from many sources, keep going until a long task is done, and pause for confirmation before irreversible steps. Google's comparisons include an overnight autonomous run on its open-source Agent Development Kit that cut in-memory session prep time by 40% and cold imports by 21%, and a Chrome team refactor across hundreds of files in Chromium's 40-million-line codebase finished in weeks instead of months.

For concept classes of closed neighborhoods in planar graphs, the non-clashing teaching dimension is now bounded by 2, and by 4 in the positive-only variant, with both bounds tight; the previous bounds were 5 and 7. The proof confirms the conjectured inequality between teaching dimension and VC dimension for this family, yet deciding whether either dimension is at most 1 remains NP-complete, and distinguishing the values 1 and 2 remains NP-hard.

For concept classes of closed neighborhoods in planar graphs, the non-clashing teaching dimension is now bounded by 2, and by 4 in the positive-only variant, with both bounds tight; the previous bounds were 5 and 7. The proof confirms the conjectured inequality between teaching dimension and VC dimension for this family, yet deciding whether either dimension is at most 1 remains NP-complete, and distinguishing the values 1 and 2 remains NP-hard.

Training data for wearable fall detection is compressed into a very small class-balanced synthetic set. Earlier pipelines either used hand-crafted features with shallow classifiers or trained deep networks on the full imbalanced corpus with re-sampling, at high data and compute cost. FallNest aligns synthetic and real samples in both time and frequency domains through gradient direction and embedding distribution, initialized by class-balanced K-Means and augmented with multi-view Fourier views. On SisFall, KFall and MobiFall under subject-wise splits with naturally imbalanced test sets, a 102K-parameter TinyCNN trained on only 32 to 146 synthetic windows reaches 93.5 to 95.8 percent mean accuracy and 0.89 to 0.98 PR-AUC over five seeds.

Training data for wearable fall detection is compressed into a very small class-balanced synthetic set. Earlier pipelines either used hand-crafted features with shallow classifiers or trained deep networks on the full imbalanced corpus with re-sampling, at high data and compute cost. FallNest aligns synthetic and real samples in both time and frequency domains through gradient direction and embedding distribution, initialized by class-balanced K-Means and augmented with multi-view Fourier views. On SisFall, KFall and MobiFall under subject-wise splits with naturally imbalanced test sets, a 102K-parameter TinyCNN trained on only 32 to 146 synthetic windows reaches 93.5 to 95.8 percent mean accuracy and 0.89 to 0.98 PR-AUC over five seeds.

Geopolitical intelligence QA now has a public benchmark of 415 reports and 5,525 question-answer pairs, where document-level retrieval is nearly saturated but even with gold evidence 70% of risk-oriented answers fall short.

Geopolitical intelligence QA now has a public benchmark of 415 reports and 5,525 question-answer pairs, where document-level retrieval is nearly saturated but even with gold evidence 70% of risk-oriented answers fall short.

Dissipative quantum neural networks (DQNNs) previously used unitary perceptrons, so each building block could only realize a restricted class of quantum transformations. A new architecture rewrites the perceptrons as isometries and adds ancilla neurons, making each building block quantum-channel universal, able to implement any completely positive trace-preserving map. On a minimal three-qubit network trained with Choi training, the Hilbert-Schmidt distance reaches a mean diamond distance of 4.55×10⁻⁴ over 100 random qubit-qubit channels, while the quantum Chernoff bound and quantum relative entropy reach only 0.102 and 0.386.

Dissipative quantum neural networks (DQNNs) previously used unitary perceptrons, so each building block could only realize a restricted class of quantum transformations. A new architecture rewrites the perceptrons as isometries and adds ancilla neurons, making each building block quantum-channel universal, able to implement any completely positive trace-preserving map. On a minimal three-qubit network trained with Choi training, the Hilbert-Schmidt distance reaches a mean diamond distance of 4.55×10⁻⁴ over 100 random qubit-qubit channels, while the quantum Chernoff bound and quantum relative entropy reach only 0.102 and 0.386.

When the inference phase of a quantum generative model is delegated to a remote quantum server, the client can now verify the sampling result without revealing inputs, circuit, or outputs. Earlier verifiable blind sampling rested on computational assumptions, whereas this protocol builds on universal blind quantum computing and trap circuits, amortizing verification over rounds and returning an estimate of the corrupted-sample ratio beyond accept or abort. Numerical simulations run on a 4-qubit IQP Born machine for the Bars&Stripes benchmark, where trap detections correlate with wrong samples.

When the inference phase of a quantum generative model is delegated to a remote quantum server, the client can now verify the sampling result without revealing inputs, circuit, or outputs. Earlier verifiable blind sampling rested on computational assumptions, whereas this protocol builds on universal blind quantum computing and trap circuits, amortizing verification over rounds and returning an estimate of the corrupted-sample ratio beyond accept or abort. Numerical simulations run on a 4-qubit IQP Born machine for the Bars&Stripes benchmark, where trap detections correlate with wrong samples.

In rotating detonation engines, the detonation wave propagates fast while the operating mode evolves slowly, and that gap makes direct deep reinforcement learning control hard. Reformulating the control problem in a moving reference frame that follows the detonation-wave pattern makes the wave structure appear quasi-steady to the agent and separates the fast and slow processes. In a one-dimensional reduced-order model, controllers trained in the moving frame learn more reliably than those trained in a stationary frame, remain effective over a broader range of actuation periods, and induce rapid transitions between mode-locked states.

In rotating detonation engines, the detonation wave propagates fast while the operating mode evolves slowly, and that gap makes direct deep reinforcement learning control hard. Reformulating the control problem in a moving reference frame that follows the detonation-wave pattern makes the wave structure appear quasi-steady to the agent and separates the fast and slow processes. In a one-dimensional reduced-order model, controllers trained in the moving frame learn more reliably than those trained in a stationary frame, remain effective over a broader range of actuation periods, and induce rapid transitions between mode-locked states.

Patient-level ultrasound diagnosis has relied on a single attention bottleneck to aggregate multiple views, which dilutes sparse malignant cues and optimizes only global accuracy. C2LUE replaces that bottleneck with multi-query cross-attention, letting several learnable queries extract complementary morphological evidence, and adds a differentiable sensitivity-specificity geometric-mean regularizer during training. On biopsy-proven thyroid and breast ultrasound datasets, patient-level AUC reaches 0.932 and 0.936, with sensitivity rising to 0.874 and 0.876 at a moderate cost in specificity.

Patient-level ultrasound diagnosis has relied on a single attention bottleneck to aggregate multiple views, which dilutes sparse malignant cues and optimizes only global accuracy. C2LUE replaces that bottleneck with multi-query cross-attention, letting several learnable queries extract complementary morphological evidence, and adds a differentiable sensitivity-specificity geometric-mean regularizer during training. On biopsy-proven thyroid and breast ultrasound datasets, patient-level AUC reaches 0.932 and 0.936, with sensitivity rising to 0.874 and 0.876 at a moderate cost in specificity.

SceneForesee turns scene evolution understanding into an evaluable task: a model estimates entity histories from observed video and predicts one scene continuation with future trajectories and Events, plus Visual Evidence from observed motion and a scene description.

SceneForesee turns scene evolution understanding into an evaluable task: a model estimates entity histories from observed video and predicts one scene continuation with future trajectories and Events, plus Visual Evidence from observed motion and a scene description.

LangChain's open source Deep Agents harness applies five context engineering techniques against context rot in long tasks: offloading oversized tool results and user messages to the filesystem, summarizing history near the limit, loading skills on demand, isolating noisy work in subagents, and keeping long-term memory as a plain file. Information is moved out of the active window rather than deleted and can be read back, so memory lives in storage the user controls and survives a model or provider switch.

LangChain's open source Deep Agents harness applies five context engineering techniques against context rot in long tasks: offloading oversized tool results and user messages to the filesystem, summarizing history near the limit, loading skills on demand, isolating noisy work in subagents, and keeping long-term memory as a plain file. Information is moved out of the active window rather than deleted and can be read back, so memory lives in storage the user controls and survives a model or provider switch.

On 57 day-ahead test days for a university teaching building, pretrained TimesFM 3.0 with covariates reached 19.97 kWh hourly mean absolute error, about 43% below the weekly seasonal-naive baseline, without training building-specific weights.

On 57 day-ahead test days for a university teaching building, pretrained TimesFM 3.0 with covariates reached 19.97 kWh hourly mean absolute error, about 43% below the weekly seasonal-naive baseline, without training building-specific weights.

Cloud gaming transmission moves bitrate decisions from whole chunks down to individual game objects: an off-the-shelf semantic segmentation model splits each frame into objects, and a deep reinforcement learning agent uses the player's in-game movement speed, object semantics and motion, and current network conditions to judge which objects the player actually attends to, giving those higher resolution while lowering the rest to save bandwidth. Across simulated 5G, 4G and WiFi scenarios the scheme beats existing ABR algorithms on the proposed QoE metrics, with normalized average QoE at least 20% higher in 5G, and a 34-participant subjective study gives it a mean opinion score of 4.16 versus 2.49 for Pensieve, 3.74 for T-Gaming and 2.86 for MetaABR.

Cloud gaming transmission moves bitrate decisions from whole chunks down to individual game objects: an off-the-shelf semantic segmentation model splits each frame into objects, and a deep reinforcement learning agent uses the player's in-game movement speed, object semantics and motion, and current network conditions to judge which objects the player actually attends to, giving those higher resolution while lowering the rest to save bandwidth. Across simulated 5G, 4G and WiFi scenarios the scheme beats existing ABR algorithms on the proposed QoE metrics, with normalized average QoE at least 20% higher in 5G, and a 34-participant subjective study gives it a mean opinion score of 4.16 versus 2.49 for Pensieve, 3.74 for T-Gaming and 2.86 for MetaABR.

World Labs has signed a definitive agreement to join AMD, bringing the spatial-intelligence company founded in 2024, along with its foundation models and applications, into the chipmaker. The two had worked together since last year on model training and inference optimization on AMD GPUs, which made combining software and hardware, foundation models, and applications a natural fit. Fei-Fei Li will join AMD as Executive Vice President and Chief Scientist, working directly with CEO Lisa Su, while Justin Johnson and Ben Mildenhall continue to lead the World Labs team; the transaction is expected to close by the end of 2026, subject to regulatory approvals and other customary closing conditions.

World Labs has signed a definitive agreement to join AMD, bringing the spatial-intelligence company founded in 2024, along with its foundation models and applications, into the chipmaker. The two had worked together since last year on model training and inference optimization on AMD GPUs, which made combining software and hardware, foundation models, and applications a natural fit. Fei-Fei Li will join AMD as Executive Vice President and Chief Scientist, working directly with CEO Lisa Su, while Justin Johnson and Ben Mildenhall continue to lead the World Labs team; the transaction is expected to close by the end of 2026, subject to regulatory approvals and other customary closing conditions.

Distillation no longer treats every teacher token alike: each token's weight is set by its predicted effect on the student's post-update reward. Earlier token-selection methods relied on proxies such as uncertainty or teacher-student divergence, which need not identify reward-improving supervision. RAD scores each token by differentiating through a temporary optimizer step, then updates token weights by projected ascent, improving mean accuracy on six competition-math benchmarks by 6.5 points for Qwen3-1.7B and 2.5 points for Qwen3-4B-Instruct-2507 at a matched training-response budget, and reaching 54.69% and 58.67% pass@1 on LiveCodeBench-v6.

Distillation no longer treats every teacher token alike: each token's weight is set by its predicted effect on the student's post-update reward. Earlier token-selection methods relied on proxies such as uncertainty or teacher-student divergence, which need not identify reward-improving supervision. RAD scores each token by differentiating through a temporary optimizer step, then updates token weights by projected ascent, improving mean accuracy on six competition-math benchmarks by 6.5 points for Qwen3-1.7B and 2.5 points for Qwen3-4B-Instruct-2507 at a matched training-response budget, and reaching 54.69% and 58.67% pass@1 on LiveCodeBench-v6.

Ship trajectory forecasting no longer needs recurrent networks or repeated sampling. Earlier AIS-based models either used only the target vessel's history or relied on clustering other vessels, and approximated probability intervals through multiple forward passes. Here the vessel's past positions and a local KDE heat map are fed to a Transformer trained with quantile loss, reaching about 0.18 km average displacement error across three Danish water areas while emitting a prediction interval in a single forward pass.

Ship trajectory forecasting no longer needs recurrent networks or repeated sampling. Earlier AIS-based models either used only the target vessel's history or relied on clustering other vessels, and approximated probability intervals through multiple forward passes. Here the vessel's past positions and a local KDE heat map are fed to a Transformer trained with quantile loss, reaching about 0.18 km average displacement error across three Danish water areas while emitting a prediction interval in a single forward pass.

Classifier-free guidance for flow-matching models has applied one shared scale to all latent tokens, even though their velocity predictions evolve differently during sampling. TrajCFG maps historical changes in guided velocity direction to bounded per-token scales, giving stronger guidance to larger directional changes and weaker guidance to more consistent directions, and fuses adjacent-step with interval-normalized longer-span feedback without extra model evaluations. Across text-to-video, image-to-video, text-to-audio-video, text-to-image and image-to-3D generation it improves on base samplers; on text-to-video it raises VBench Total by 0.30 to 0.75 points over the strongest evaluated baseline on each of three backbones.

Classifier-free guidance for flow-matching models has applied one shared scale to all latent tokens, even though their velocity predictions evolve differently during sampling. TrajCFG maps historical changes in guided velocity direction to bounded per-token scales, giving stronger guidance to larger directional changes and weaker guidance to more consistent directions, and fuses adjacent-step with interval-normalized longer-span feedback without extra model evaluations. Across text-to-video, image-to-video, text-to-audio-video, text-to-image and image-to-3D generation it improves on base samplers; on text-to-video it raises VBench Total by 0.30 to 0.75 points over the strongest evaluated baseline on each of three backbones.

vLLM adds a push-based NIXL connector that moves prefill-to-decode KV cache transfer coordination off the critical path. The previous pull mode waited for prefill to finish, relayed source metadata through the router, and had the decode instance allocate blocks and issue the read, a serial cost charged to time-to-first-token. Push mode registers destination blocks at allocation time and has the prefill side issue an NIXL WRITE from a background thread; on Qwen3-32B in a 1P1D P5en plus EFA deployment at 8 QPS and 16k input, TTFT fell from about 1,744 ms to about 540 ms and TPOT from about 10.4 ms to about 7 ms, with results varying by workload and configuration.

vLLM adds a push-based NIXL connector that moves prefill-to-decode KV cache transfer coordination off the critical path. The previous pull mode waited for prefill to finish, relayed source metadata through the router, and had the decode instance allocate blocks and issue the read, a serial cost charged to time-to-first-token. Push mode registers destination blocks at allocation time and has the prefill side issue an NIXL WRITE from a background thread; on Qwen3-32B in a 1P1D P5en plus EFA deployment at 8 QPS and 16k input, TTFT fell from about 1,744 ms to about 540 ms and TPOT from about 10.4 ms to about 7 ms, with results varying by workload and configuration.

Gait attribute recognition now has a large multimodal benchmark: RA-GAR covers 1,001 subjects and 242,729 walking sequences annotated with 15 static and motion gait attributes, addressing the shortage of fine-grained annotation. The accompanying Gait-VL matches attribute-specific multi-frame visual evidence against textual class prototypes instead of representing candidate classes through learned classifier output weights. It reaches state-of-the-art attribute recognition on RA-GAR, and replacing generic upstream vision models with the attribute-pretrained encoder yields consistent gains across several identity recognition pipelines and benchmarks.

Gait attribute recognition now has a large multimodal benchmark: RA-GAR covers 1,001 subjects and 242,729 walking sequences annotated with 15 static and motion gait attributes, addressing the shortage of fine-grained annotation. The accompanying Gait-VL matches attribute-specific multi-frame visual evidence against textual class prototypes instead of representing candidate classes through learned classifier output weights. It reaches state-of-the-art attribute recognition on RA-GAR, and replacing generic upstream vision models with the attribute-pretrained encoder yields consistent gains across several identity recognition pipelines and benchmarks.

Stance-detection prompt optimization now splits the prompt into six functional modules refined one at a time, with misclassification feedback routed by gold label, and on SemEval-2016 four of five open-weight backbones reach higher mean Macro-F1 than pooling all errors together.

Stance-detection prompt optimization now splits the prompt into six functional modules refined one at a time, with misclassification feedback routed by gold label, and on SemEval-2016 four of five open-weight backbones reach higher mean Macro-F1 than pooling all errors together.

For time-stamped speaker-attributed transcription of two-person smart-glasses conversations, extracting or separating each target speaker before decoding lowers error rates, and the gain grows with the speech overlap ratio on the validation set.

For time-stamped speaker-attributed transcription of two-person smart-glasses conversations, extracting or separating each target speaker before decoding lowers error rates, and the gain grows with the speech overlap ratio on the validation set.

Hydrodynamic design of blood-contacting medical devices has long relied on meshed CFD, where mesh structure and compute cost slow optimization. A mesh-free physics-informed neural network writes the axisymmetric mass-conservation equation into its loss function and is trained on experimental PIV/LDV data over the FDA reference nozzle, spanning laminar, transitional and fully turbulent regimes (Re = 500–6500), reconstructing flow fields and turbulent kinetic energy at high resolution and yielding an empirical correlation for peak TKE as a rapid hydrodynamic indicator.

Hydrodynamic design of blood-contacting medical devices has long relied on meshed CFD, where mesh structure and compute cost slow optimization. A mesh-free physics-informed neural network writes the axisymmetric mass-conservation equation into its loss function and is trained on experimental PIV/LDV data over the FDA reference nozzle, spanning laminar, transitional and fully turbulent regimes (Re = 500–6500), reconstructing flow fields and turbulent kinetic energy at high resolution and yielding an empirical correlation for peak TKE as a rapid hydrodynamic indicator.

Panvascular intervention research is starting to write traceability into an agent's output contract: ScienceBuddy-PV fixes a shared convolutional model on the public ARCADE coronary angiography benchmark and evaluates only an evidence gate and action audit, where the gate accepts 76.7% of stenosis-frame candidates against 100% for a score-only comparator. On the locked test set, stenosis Dice is 0.231 and AUROC is 0.957 at a pixel prevalence of 0.00975, while reference lesion candidates are accepted 44.3% and rejected 55.7% of the time. The result is an image-level technical validation; recursive harness optimization and model learning remain protocols rather than executed experiments.

Panvascular intervention research is starting to write traceability into an agent's output contract: ScienceBuddy-PV fixes a shared convolutional model on the public ARCADE coronary angiography benchmark and evaluates only an evidence gate and action audit, where the gate accepts 76.7% of stenosis-frame candidates against 100% for a score-only comparator. On the locked test set, stenosis Dice is 0.231 and AUROC is 0.957 at a pixel prevalence of 0.00975, while reference lesion candidates are accepted 44.3% and rejected 55.7% of the time. The result is an image-level technical validation; recursive harness optimization and model learning remain protocols rather than executed experiments.

Influential-node identification in complex networks now has a cross-scale approach. Centrality-based and learning-based methods often integrate structural information across network scales only to a limited degree, so local, community-level and global patterns are not used together. CSBA-Net extracts structural representations at the three scales, models dependencies between them with a BiGRU and multi-head self-attention, and learns node influence scores by regression against SIR diffusion simulations; on 24 real and 12 synthetic networks its ranking metrics exceed the compared centrality, CNN and GNN baselines.

Influential-node identification in complex networks now has a cross-scale approach. Centrality-based and learning-based methods often integrate structural information across network scales only to a limited degree, so local, community-level and global patterns are not used together. CSBA-Net extracts structural representations at the three scales, models dependencies between them with a BiGRU and multi-head self-attention, and learns node influence scores by regression against SIR diffusion simulations; on 24 real and 12 synthetic networks its ranking metrics exceed the compared centrality, CNN and GNN baselines.

ElevenLabs runs its own customer support on ElevenAgents, resolving over 22,000 tickets a month without needing to involve a person, with 65% of last month's tickets resolved end to end and first response down from hours to 60 seconds.

ElevenLabs runs its own customer support on ElevenAgents, resolving over 22,000 tickets a month without needing to involve a person, with 65% of last month's tickets resolved end to end and first response down from hours to 60 seconds.

REA connects an agent to a set of local analysis tools for inspecting native binaries, JavaScript and Electron apps, .NET assemblies and websites, so the agent can explain how a feature works without source code and return the evidence and limitations behind each conclusion. It registers with agents such as Claude Code, Codex, Cursor and Gemini CLI through MCP, offers the same workflows from the terminal, can reuse an existing Hopper, Ghidra or IDA installation for native analysis, and needs no native analysis engine for static JavaScript analysis.

REA connects an agent to a set of local analysis tools for inspecting native binaries, JavaScript and Electron apps, .NET assemblies and websites, so the agent can explain how a feature works without source code and return the evidence and limitations behind each conclusion. It registers with agents such as Claude Code, Codex, Cursor and Gemini CLI through MCP, offers the same workflows from the terminal, can reuse an existing Hopper, Ghidra or IDA installation for native analysis, and needs no native analysis engine for static JavaScript analysis.

In its June 2026 update, Skyvern renamed Workflows to Agents across the product, public API, and docs, moving the API under /agents and accepting an agent_id alias while existing workflow_id inputs and /workflows endpoints keep working.

In its June 2026 update, Skyvern renamed Workflows to Agents across the product, public API, and docs, moving the API under /agents and accepting an agent_id alias while existing workflow_id inputs and /workflows endpoints keep working.

The Claude Platform now exports OpenTelemetry traces from Managed Agents, the Messages API and Claude SDKs, so each agent turn, model call and tool use arrives in existing observability tools with latency, tokens and IDs.

The Claude Platform now exports OpenTelemetry traces from Managed Agents, the Messages API and Claude SDKs, so each agent turn, model call and tool use arrives in existing observability tools with latency, tokens and IDs.

Google has opened Gemini 4 Argon, a frontier model for complex workflows, to developers, covering real-world software engineering, enterprise knowledge work, and cybersecurity defense. It is designed for long-horizon tasks that need deep reasoning and structural precision, such as code migrations and frontier research, and handles multi-step problems. Developers can start building through the paid tier of the Gemini API in Google AI Studio and Android Studio, while Google AI Ultra subscribers can try it in AI Studio and the agent-first development platform Google Antigravity.

Google has opened Gemini 4 Argon, a frontier model for complex workflows, to developers, covering real-world software engineering, enterprise knowledge work, and cybersecurity defense. It is designed for long-horizon tasks that need deep reasoning and structural precision, such as code migrations and frontier research, and handles multi-step problems. Developers can start building through the paid tier of the Gemini API in Google AI Studio and Android Studio, while Google AI Ultra subscribers can try it in AI Studio and the agent-first development platform Google Antigravity.

At the 2026 OCP Global Summit, AMD showed open full-stack infrastructure for agentic AI, with the Helios rack ramping production a year after its 2025 OCP debut and customer deliveries, deployments and workload validations underway, including initial systems installed and running at OpenAI, Meta, Microsoft and Oracle; the Helion rack lets customers tailor CPU core density, performance and power to agentic workloads in CPU-only and integrated CPU/GPU configurations, is planned for November 2026, and AMD plans to contribute its architecture to OCP.

At the 2026 OCP Global Summit, AMD showed open full-stack infrastructure for agentic AI, with the Helios rack ramping production a year after its 2025 OCP debut and customer deliveries, deployments and workload validations underway, including initial systems installed and running at OpenAI, Meta, Microsoft and Oracle; the Helion rack lets customers tailor CPU core density, performance and power to agentic workloads in CPU-only and integrated CPU/GPU configurations, is planned for November 2026, and AMD plans to contribute its architecture to OCP.

The open-source project PentAGI hands penetration testing tasks to a fully autonomous multi-agent system whose agents collaborate through a graph structure and can be self-hosted. Previously such complex attack workflows relied mainly on manual step-by-step operation or a single model, while this project ships a complete system with a Go backend, React frontend and GraphQL interface, plus Docker Compose deployment. It is released under the MIT license, and the repository has gathered about 29.3k stars and 4.4k forks.

The open-source project PentAGI hands penetration testing tasks to a fully autonomous multi-agent system whose agents collaborate through a graph structure and can be self-hosted. Previously such complex attack workflows relied mainly on manual step-by-step operation or a single model, while this project ships a complete system with a Go backend, React frontend and GraphQL interface, plus Docker Compose deployment. It is released under the MIT license, and the repository has gathered about 29.3k stars and 4.4k forks.

Sakana AI launched My Fish for Sakana Chat, turning its previously behind-the-scenes memory into a viewable interface where users can see what it remembers, when those memories were used, and how they connect.

Sakana AI launched My Fish for Sakana Chat, turning its previously behind-the-scenes memory into a viewable interface where users can see what it remembers, when those memories were used, and how they connect.

Mastra has released a Metrics API that lets developers query token usage, model cost, latency and error rates for production projects through the client SDK, CLI or observability endpoints. Previously these metrics were only viewable in the Mastra platform and Studio, with no way to construct custom queries. Queries can now combine time ranges, status filters and grouping to produce totals, breakdowns, time series and percentile data for custom reports, dashboards, downstream delivery and alerts when cost or error rates rise.

Mastra has released a Metrics API that lets developers query token usage, model cost, latency and error rates for production projects through the client SDK, CLI or observability endpoints. Previously these metrics were only viewable in the Mastra platform and Studio, with no way to construct custom queries. Queries can now combine time ranges, status filters and grouping to produce totals, breakdowns, time series and percentile data for custom reports, dashboards, downstream delivery and alerts when cost or error rates rise.