AI and science frontiers · 2026-10-27

SGLang closed a cross-request KV cache exposure in disaggregated prefill/decode serving: after a request aborts mid-transfer, RDMA writes it already launched can still land and overwrite another request's cache. Decode now keeps the aborted request's slots until every notified prefill rank acknowledges a drain or a bounded timeout fires, default-on for Mooncake and NIXL from v0.5.21, with throughput deltas within ±2% on 2×H100.

SGLang closed a cross-request KV cache exposure in disaggregated prefill/decode serving: after a request aborts mid-transfer, RDMA writes it already launched can still land and overwrite another request's cache. Decode now keeps the aborted request's slots until every notified prefill rank acknowledges a drain or a bounded timeout fires, default-on for Mooncake and NIXL from v0.5.21, with throughput deltas within ±2% on 2×H100.

As the training set grows, diffusion models move from memorizing training samples to generalizing, and between these phases an intermediate kind of attractor appears that matches no training image yet recurs in generated samples. Reading generation as retrieval in a dense associative memory predicts such spurious states once storage capacity is exceeded; on MNIST, FASHION-MNIST, CIFAR-10 and LSUN-CHURCH the spurious fraction rises and then falls at the memorization-to-generalization boundary, with basin volumes between those of memorized and generalized samples.

As the training set grows, diffusion models move from memorizing training samples to generalizing, and between these phases an intermediate kind of attractor appears that matches no training image yet recurs in generated samples. Reading generation as retrieval in a dense associative memory predicts such spurious states once storage capacity is exceeded; on MNIST, FASHION-MNIST, CIFAR-10 and LSUN-CHURCH the spurious fraction rises and then falls at the memorization-to-generalization boundary, with basin volumes between those of memorized and generalized samples.

Cell proportions from bulk RNA sequencing are inferred through a single-cell reference, and that reference can change the conclusion drawn from them. In 981 PBMC donors the CD8 T-cell age association reversed in all 80 six-type references and the CD4 slope was inflated about 6.3-fold; in 48 islet donors a borderline alpha-cell association became significant in 54% to 100% of references. The error recurred across four estimator classes, its direction was set by the estimator, and fit accuracy did not flag it.

Cell proportions from bulk RNA sequencing are inferred through a single-cell reference, and that reference can change the conclusion drawn from them. In 981 PBMC donors the CD8 T-cell age association reversed in all 80 six-type references and the CD4 slope was inflated about 6.3-fold; in 48 islet donors a borderline alpha-cell association became significant in 54% to 100% of references. The error recurred across four estimator classes, its direction was set by the estimator, and fit accuracy did not flag it.

In the symmetric-privacy setting where both the query and the database must stay hidden, HNSecW turns HNSW graph-index approximate nearest neighbor search into a fully data-independent procedure that runs under secure multi-party computation, giving single-query online communication and local computation logarithmic complexity. Earlier protocols of this kind mostly used cluster-based pipelines whose online communication grows linearly with database size. On million- to billion-scale benchmarks, HNSecW keeps recall comparable to plaintext HNSW and reaches up to orders-of-magnitude end-to-end speedup over 2PC PANTHER on a wide-area network.

In the symmetric-privacy setting where both the query and the database must stay hidden, HNSecW turns HNSW graph-index approximate nearest neighbor search into a fully data-independent procedure that runs under secure multi-party computation, giving single-query online communication and local computation logarithmic complexity. Earlier protocols of this kind mostly used cluster-based pipelines whose online communication grows linearly with database size. On million- to billion-scale benchmarks, HNSecW keeps recall comparable to plaintext HNSW and reaches up to orders-of-magnitude end-to-end speedup over 2PC PANTHER on a wide-area network.

Prediction intervals for engineering regression are usually calibrated on random splits, while deployment happens outside training conditions. A benchmark splits seven public engineering datasets into in-distribution, soft-shift and hard-shift rungs along physically meaningful shift variables, and compares 12 uncertainty quantification methods from five families tuned on source data only. In distribution, conformal methods cover at least 0.946 on every dataset; under hard shift, mean coverage falls to 0.40–0.75 for the 11 methods that always return an interval, and no systematic difference in method ranks is detected.

Prediction intervals for engineering regression are usually calibrated on random splits, while deployment happens outside training conditions. A benchmark splits seven public engineering datasets into in-distribution, soft-shift and hard-shift rungs along physically meaningful shift variables, and compares 12 uncertainty quantification methods from five families tuned on source data only. In distribution, conformal methods cover at least 0.946 on every dataset; under hard shift, mean coverage falls to 0.40–0.75 for the 11 methods that always return an interval, and no systematic difference in method ranks is detected.

Enzyme kinetics prediction has relied on absolute regression over heterogeneous data, where the same enzyme-substrate system can differ by orders of magnitude across studies and models drift toward the mean. DeltaKcat recasts the task as pairwise relative learning: within one study it fixes either the enzyme or the substrate and a siamese network predicts the kinetic difference between two enzyme-substrate complexes, after which an experimentally characterized reference complex restores absolute values. Under a cold-start split, absolute kcat prediction reaches R² 0.71 and PCC 0.85, against R² 0.14 to 0.48 for DLKcat, CataPro and UniKP; on external adenylate kinase data excluded from training, zero-shot PCC is 0.40.

Enzyme kinetics prediction has relied on absolute regression over heterogeneous data, where the same enzyme-substrate system can differ by orders of magnitude across studies and models drift toward the mean. DeltaKcat recasts the task as pairwise relative learning: within one study it fixes either the enzyme or the substrate and a siamese network predicts the kinetic difference between two enzyme-substrate complexes, after which an experimentally characterized reference complex restores absolute values. Under a cold-start split, absolute kcat prediction reaches R² 0.71 and PCC 0.85, against R² 0.14 to 0.48 for DLKcat, CataPro and UniKP; on external adenylate kinase data excluded from training, zero-shot PCC is 0.40.

Facial emotion recognition for human-robot interaction is usually judged by clean-set accuracy, yet a robot that maps an emotion estimate to an action also needs to know whether that estimate can be trusted. Comparing five pipelines on a five-class FER2013 subset, the strongest deep residual CNN reaches 0.760 accuracy and 0.753 macro-F1 but collapses with every other trained architecture under gradient attacks: FGSM drops all four below 0.16 accuracy at epsilon 0.01, and PGD drives them near zero at epsilon 0.02 and above. The attacked predictions stay confident, with mean confidence 0.964 under PGD and 96.0% still above a 0.70 threshold after temperature scaling, so thresholding the maximum softmax probability does not separate legitimate from adversarial predictions.

Facial emotion recognition for human-robot interaction is usually judged by clean-set accuracy, yet a robot that maps an emotion estimate to an action also needs to know whether that estimate can be trusted. Comparing five pipelines on a five-class FER2013 subset, the strongest deep residual CNN reaches 0.760 accuracy and 0.753 macro-F1 but collapses with every other trained architecture under gradient attacks: FGSM drops all four below 0.16 accuracy at epsilon 0.01, and PGD drives them near zero at epsilon 0.02 and above. The attacked predictions stay confident, with mean confidence 0.964 under PGD and 96.0% still above a 0.70 threshold after temperature scaling, so thresholding the maximum softmax probability does not separate legitimate from adversarial predictions.

Automated blood and bone marrow smear analysis had been confined to manually selected regions of interest because gigapixel whole-slide scanning is computationally heavy. HemAutomaton uses a neural cellular automaton to locate white blood cells across entire slides without manual region selection, matching Faster R-CNN in-domain detection and reaching a mean cross-domain F1 of 0.79 versus 0.65, with 3.3k trainable parameters against 41.4M. Across three blood cancer cohorts it extracted 108 million single-cell images and performed automated differential counting through few-shot classification with a pretrained hematology foundation model.

Automated blood and bone marrow smear analysis had been confined to manually selected regions of interest because gigapixel whole-slide scanning is computationally heavy. HemAutomaton uses a neural cellular automaton to locate white blood cells across entire slides without manual region selection, matching Faster R-CNN in-domain detection and reaching a mean cross-domain F1 of 0.79 versus 0.65, with 3.3k trainable parameters against 41.4M. Across three blood cancer cohorts it extracted 108 million single-cell images and performed automated differential counting through few-shot classification with a pretrained hematology foundation model.

Predicting cancer treatment side effects has long relied on static nomograms built from a handful of variables. SPARC instead models longitudinal sequences of blood counts, metabolic panels and vital signs, pretraining on 101,920,379 measurements from 224,047 patients and then fine-tuning for mortality, anemia, neutropenia, febrile neutropenia, acute kidney injury and liver toxicity. It matched or led comparison models on held-out, prospective and multi-hospital cohorts, and its predicted risk strata identified patients most likely to benefit from G-CSF prophylaxis.

Predicting cancer treatment side effects has long relied on static nomograms built from a handful of variables. SPARC instead models longitudinal sequences of blood counts, metabolic panels and vital signs, pretraining on 101,920,379 measurements from 224,047 patients and then fine-tuning for mortality, anemia, neutropenia, febrile neutropenia, acute kidney injury and liver toxicity. It matched or led comparison models on held-out, prospective and multi-hospital cohorts, and its predicted risk strata identified patients most likely to benefit from G-CSF prophylaxis.

Ewing sarcoma is genomically near-uniform yet patients fare very differently. Single-cell and spatial transcriptomics of tumors and their microenvironment in 135 patients identified five tumor-cell transcriptional states and eight niches defined by neighboring cell configurations, and niche composition sorted patients into Fusion Active, Proliferative, Immune Reactive and TME-enriched types. The Immune Reactive type, marked by interferon-activated tumor cells, had favorable survival, and the survival differences reproduced in two independent bulk transcriptomic cohorts.

Ewing sarcoma is genomically near-uniform yet patients fare very differently. Single-cell and spatial transcriptomics of tumors and their microenvironment in 135 patients identified five tumor-cell transcriptional states and eight niches defined by neighboring cell configurations, and niche composition sorted patients into Fusion Active, Proliferative, Immune Reactive and TME-enriched types. The Immune Reactive type, marked by interferon-activated tumor cells, had favorable survival, and the survival differences reproduced in two independent bulk transcriptomic cohorts.

Wearable accelerometers record activity around the clock, but summary measures and population-wide clustering can miss disease-associated behavioral differences. LocPA adjusts hourly activity profiles for age, sex and body mass index, then tests local disease enrichment on a similarity network and groups enriched neighborhoods into patterns. In UK Biobank 24-hour accelerometry, a low-activity and a sleep-disrupted phenotype recurred across six diseases; in hypertension and diabetes the sleep-disrupted group carried greater disease burden, while average acceleration and moderate-to-vigorous activity time did not follow that ordering.

Wearable accelerometers record activity around the clock, but summary measures and population-wide clustering can miss disease-associated behavioral differences. LocPA adjusts hourly activity profiles for age, sex and body mass index, then tests local disease enrichment on a similarity network and groups enriched neighborhoods into patterns. In UK Biobank 24-hour accelerometry, a low-activity and a sleep-disrupted phenotype recurred across six diseases; in hypertension and diabetes the sleep-disrupted group carried greater disease burden, while average acceleration and moderate-to-vigorous activity time did not follow that ordering.

On four trailer-editing task types built from 49 real films and their released trailers, decaying an LLM-judged turn-level process reward to zero within the same GRPO setup gives a 45.4% relative gain over outcome-only GRPO on Fine Cut, while the fixed-weight variant scores below outcome-only GRPO on the three shorter task types and truncates 44.1% of rollouts. Process rewards do encourage visual inspection and reasoning early, but continued optimization drives redundant tool calls without a committed result; token-level on-policy self-distillation proves highly sensitive to the teacher's privileged information.

On four trailer-editing task types built from 49 real films and their released trailers, decaying an LLM-judged turn-level process reward to zero within the same GRPO setup gives a 45.4% relative gain over outcome-only GRPO on Fine Cut, while the fixed-weight variant scores below outcome-only GRPO on the three shorter task types and truncates 44.1% of rollouts. Process rewards do encourage visual inspection and reasoning early, but continued optimization drives redundant tool calls without a committed result; token-level on-policy self-distillation proves highly sensitive to the teacher's privileged information.

Across 47 public PHM datasets harmonized to one tensor contract, an 801,808-parameter channel-independent Transformer encoder pretrained with masked patch prediction raises fault-classification macro-F1 from 0.350 to 0.812 on the fully withheld CWRU set and from 0.569 to 0.951 on HSG18, with frozen linear probing alone reaching 0.705 and 0.906; the same pretraining does not reliably beat a from-scratch baseline on CMAPSS remaining-useful-life estimation. Simulated federated pretraining over ten clients grouped by physical domain retains part of the representation but trails the centralized encoder at matched budget, and the self-supervised loss does not predict downstream utility.

Across 47 public PHM datasets harmonized to one tensor contract, an 801,808-parameter channel-independent Transformer encoder pretrained with masked patch prediction raises fault-classification macro-F1 from 0.350 to 0.812 on the fully withheld CWRU set and from 0.569 to 0.951 on HSG18, with frozen linear probing alone reaching 0.705 and 0.906; the same pretraining does not reliably beat a from-scratch baseline on CMAPSS remaining-useful-life estimation. Simulated federated pretraining over ten clients grouped by physical domain retains part of the representation but trails the centralized encoder at matched budget, and the self-supervised loss does not predict downstream utility.

One-shot federated learning finishes training in a single communication round, but the resulting model is frozen once devices join or leave, leaving full retraining as the fallback. F-SOSA draws a fixed-size uniform random client subset so outputs under neighboring device sets stay bounded in total variation, then rewrites curvature-aware fusion as a projection over client-additive sufficient statistics, enabling F-SOSA-U for withdrawing a contribution and F-SOSA-I for adding one. In a SUMO urban mobility simulation, both match retraining accuracy while cutting cumulative time and communication by 36%–44% and 22%–33%.

One-shot federated learning finishes training in a single communication round, but the resulting model is frozen once devices join or leave, leaving full retraining as the fallback. F-SOSA draws a fixed-size uniform random client subset so outputs under neighboring device sets stay bounded in total variation, then rewrites curvature-aware fusion as a projection over client-additive sufficient statistics, enabling F-SOSA-U for withdrawing a contribution and F-SOSA-I for adding one. In a SUMO urban mobility simulation, both match retraining accuracy while cutting cumulative time and communication by 36%–44% and 22%–33%.

Evaluations of long-horizon web-search agents have leaned on simulated trajectories or on trails reconstructed from final answers, which makes the search process itself hard to inspect. LongSearchEval runs 16 models on 100 tasks under one shared agent controller and five restricted web tools, retains 1,600 real trajectories, verifies answer groups against the webpages the agent itself cited, then links flagged failure nodes to groups judged incorrect. Human review agrees with 94.50% of automatic judgments on 946 answer groups, and 311 of 317 sampled round-to-group links pass a feature-correspondence check.

Evaluations of long-horizon web-search agents have leaned on simulated trajectories or on trails reconstructed from final answers, which makes the search process itself hard to inspect. LongSearchEval runs 16 models on 100 tasks under one shared agent controller and five restricted web tools, retains 1,600 real trajectories, verifies answer groups against the webpages the agent itself cited, then links flagged failure nodes to groups judged incorrect. Human review agrees with 94.50% of automatic judgments on 946 answer groups, and 311 of 317 sampled round-to-group links pass a feature-correspondence check.

AI agent skills run on implicit trust and get little vetting before installation. NVIDIA's open-source SkillSpector scans a repository, zip file, directory or single file before install, covering 71 vulnerability patterns in 17 categories with fast static analysis and an optional LLM semantic pass. In the 31,132-skill analyzed subset of its research dataset, 26.1% of skills contain vulnerabilities and 5.2% show likely malicious intent.

AI agent skills run on implicit trust and get little vetting before installation. NVIDIA's open-source SkillSpector scans a repository, zip file, directory or single file before install, covering 71 vulnerability patterns in 17 categories with fast static analysis and an optional LLM semantic pass. In the 31,132-skill analyzed subset of its research dataset, 26.1% of skills contain vulnerabilities and 5.2% show likely malicious intent.

Olive short tandem repeats (STRs, stretches whose repeat-unit copy number varies) had been read mainly through small microsatellite panels, leaving genome-wide variation and interpretation open. Across 212 Iranian and Mediterranean olive accessions, the study catalogued 117,119 polymorphic STR loci and passed natural allele series through three plant genomic language models, AgroNT, PlantCAD2 and GPN, measuring how far embeddings move as repeat length departs from the reference allele. On 62,847 loci, the three models shared 7,417 embedding-sensitive loci, and at most tested loci STR responses exceeded length-matched nonrepetitive indels.

Olive short tandem repeats (STRs, stretches whose repeat-unit copy number varies) had been read mainly through small microsatellite panels, leaving genome-wide variation and interpretation open. Across 212 Iranian and Mediterranean olive accessions, the study catalogued 117,119 polymorphic STR loci and passed natural allele series through three plant genomic language models, AgroNT, PlantCAD2 and GPN, measuring how far embeddings move as repeat length departs from the reference allele. On 62,847 loci, the three models shared 7,417 embedding-sensitive loci, and at most tested loci STR responses exceeded length-matched nonrepetitive indels.

Generative models retrained on their own output narrow their distribution and lose the tails first, and training pipelines have lacked a way to correct this while they run. HOMEOSTASIS fuses signed, hash-chained provenance records with three model vital signs, predictive entropy, calibration error and representational drift, into a Lyapunov-style collapse-risk score, and a PID controller adjusts the human-data mixing ratio and quarantine decisions. On a five-dimensional Gaussian recursion over 30 seeds and 20 generations, closed-loop control lowered final collapse risk from 2.67 to 0.48 and held tail coverage at 97.4% against 66.2% without control, while provenance-gated quarantine showed no detectable benefit beyond control alone.

Generative models retrained on their own output narrow their distribution and lose the tails first, and training pipelines have lacked a way to correct this while they run. HOMEOSTASIS fuses signed, hash-chained provenance records with three model vital signs, predictive entropy, calibration error and representational drift, into a Lyapunov-style collapse-risk score, and a PID controller adjusts the human-data mixing ratio and quarantine decisions. On a five-dimensional Gaussian recursion over 30 seeds and 20 generations, closed-loop control lowered final collapse risk from 2.67 to 0.48 and held tail coverage at 97.4% against 66.2% without control, while provenance-gated quarantine showed no detectable benefit beyond control alone.

Sarvam built a knowledge fabric for enterprise agents that compiles documents once into a navigable knowledge tree and a queryable domain ontology, so agents locate material by directory and filter by entity instead of re-retrieving passages for every question. On Harvey's LAB law-firm benchmark with GLM 5.3 as the agent model, answers met 84.2% of expert rubric criteria versus 55.1% for agentic RAG, while average task cost fell from $0.64 to $0.28; on Sierra's tau-knowledge banking benchmark with GLM 5.2, the fabric matched baseline accuracy using 64% fewer tokens.

Sarvam built a knowledge fabric for enterprise agents that compiles documents once into a navigable knowledge tree and a queryable domain ontology, so agents locate material by directory and filter by entity instead of re-retrieving passages for every question. On Harvey's LAB law-firm benchmark with GLM 5.3 as the agent model, answers met 84.2% of expert rubric criteria versus 55.1% for agentic RAG, while average task cost fell from $0.64 to $0.28; on Sierra's tau-knowledge banking benchmark with GLM 5.2, the fabric matched baseline accuracy using 64% fewer tokens.

Single-cell image classification struggles to name species on unseen strains. SC-MAPS aligns phase-contrast, bright-field and three fluorescence channels on 916,435 cells from 13 species and fuses them with deep learning; on held-out strains single-cell species accuracy is 65.4%, pooling species probabilities into eight AST-routing groups raises it to 77.1%, and combining five cells from one sample reaches 91.1%.

Single-cell image classification struggles to name species on unseen strains. SC-MAPS aligns phase-contrast, bright-field and three fluorescence channels on 916,435 cells from 13 species and fuses them with deep learning; on held-out strains single-cell species accuracy is 65.4%, pooling species probabilities into eight AST-routing groups raises it to 77.1%, and combining five cells from one sample reaches 91.1%.

When an institutional AI agent proposes a tool call, the same refund may come from an authorised tax officer or from an instruction hidden in the petition it was asked to summarise. This authorisation guard keeps separate neutrosophic triples for effect and for purpose, discounts evidence by channel trust, fuses it, and then authorises, escalates or denies. On 1,054 InjecAgent attack cases, retaining ambivalence about an unrequested call lowered worst-case direct-harm pass-through after channel mislabelling from 100% to 34.5%, and to 25.0% on synthetic e-government scenarios.

When an institutional AI agent proposes a tool call, the same refund may come from an authorised tax officer or from an instruction hidden in the petition it was asked to summarise. This authorisation guard keeps separate neutrosophic triples for effect and for purpose, discounts evidence by channel trust, fuses it, and then authorises, escalates or denies. On 1,054 InjecAgent attack cases, retaining ambivalence about an unrequested call lowered worst-case direct-harm pass-through after channel mislabelling from 100% to 34.5%, and to 25.0% on synthetic e-government scenarios.

Greedy State Boundary Search (GSBS) finds transitions between stable neural activity patterns, but it rebuilt state templates and recomputed correlations for every candidate boundary, making analyses across recordings and settings expensive. HP-GSBS rewrites the matching objective as additive interval contributions and computes t-distance from reusable correlation statistics, adding streaming candidate reduction, sequential local refinement and active-block reuse across CPU, PyTorch CUDA and Triton paths. Across seven controlled workloads with two input realizations each, median same-input speedups range from 154 to 1,730, and all 168 optimized executions reproduce complete reference boundary histories and selected state counts, with a maximum absolute t-distance difference of 1.88×10⁻¹¹.

Greedy State Boundary Search (GSBS) finds transitions between stable neural activity patterns, but it rebuilt state templates and recomputed correlations for every candidate boundary, making analyses across recordings and settings expensive. HP-GSBS rewrites the matching objective as additive interval contributions and computes t-distance from reusable correlation statistics, adding streaming candidate reduction, sequential local refinement and active-block reuse across CPU, PyTorch CUDA and Triton paths. Across seven controlled workloads with two input realizations each, median same-input speedups range from 154 to 1,730, and all 168 optimized executions reproduce complete reference boundary histories and selected state counts, with a maximum absolute t-distance difference of 1.88×10⁻¹¹.

When an AI agent needs an answer from a real folder of reports, ledgers and contracts, it searches, opens a file, reads pages and searches again, sometimes for twenty or thirty steps. Drex for Search moves that decision earlier: before the agent reads anything, the decision model Drex decides which folder, file and page can hold the answer, and inside a spreadsheet which table and column, so the agent reads only those places and returns the source. Across 6 professional-work benchmarks and 529 questions, coding agents using it took 62% fewer steps and 56% fewer input tokens, raised accuracy from 63.1% to 67.1%, and cut median time per question from 80.8 s to 35.5 s.

When an AI agent needs an answer from a real folder of reports, ledgers and contracts, it searches, opens a file, reads pages and searches again, sometimes for twenty or thirty steps. Drex for Search moves that decision earlier: before the agent reads anything, the decision model Drex decides which folder, file and page can hold the answer, and inside a spreadsheet which table and column, so the agent reads only those places and returns the source. Across 6 professional-work benchmarks and 529 questions, coding agents using it took 62% fewer steps and 56% fewer input tokens, raised accuracy from 63.1% to 67.1%, and cut median time per question from 80.8 s to 35.5 s.

For the 2026–2027 U.S. influenza season, vaccine effectiveness has been forecast by component before the season starts: the topological model gives 40.3%, 56.9% and 47.1% against A(H1N1)pdm09, A(H3N2) and B/Victoria, and the geometric model gives 34.0%, 56.8% and 51.1%. The inputs are 87 hemagglutinin sequences collected in August–September compared with that season's vaccine strains through two mathematical distances, fed into linear regressions fitted on 26 historical season-component observations. These are component point forecasts, not an overall seasonal estimate, with 95% prediction intervals roughly 52–58 percentage points wide, to be checked against final US Flu VE Network estimates.

For the 2026–2027 U.S. influenza season, vaccine effectiveness has been forecast by component before the season starts: the topological model gives 40.3%, 56.9% and 47.1% against A(H1N1)pdm09, A(H3N2) and B/Victoria, and the geometric model gives 34.0%, 56.8% and 51.1%. The inputs are 87 hemagglutinin sequences collected in August–September compared with that season's vaccine strains through two mathematical distances, fed into linear regressions fitted on 26 historical season-component observations. These are component point forecasts, not an overall seasonal estimate, with 95% prediction intervals roughly 52–58 percentage points wide, to be checked against final US Flu VE Network estimates.

Large language models extract cause-effect pairs from nuclear Licensee Event Reports as flat assertions, leaving risk analysts unable to separate reliable relations from spurious ones. A study stacks four signals, token-level probabilities, self-consistency across stochastic decoding runs, disagreement among three independently fine-tuned models, and conformal prediction, into a calibrated confidence score per pair. On a held-out test set the integrated score reached an expected calibration error of 0.067 and an AUROC of 0.84, above every individual signal, and uncertainty-guided review found 82% of extraction errors within the first 20% of relations reviewed, against 24% for random selection.

Large language models extract cause-effect pairs from nuclear Licensee Event Reports as flat assertions, leaving risk analysts unable to separate reliable relations from spurious ones. A study stacks four signals, token-level probabilities, self-consistency across stochastic decoding runs, disagreement among three independently fine-tuned models, and conformal prediction, into a calibrated confidence score per pair. On a held-out test set the integrated score reached an expected calibration error of 0.067 and an AUROC of 0.84, above every individual signal, and uncertainty-guided review found 82% of extraction errors within the first 20% of relations reviewed, against 24% for random selection.

Judging how immersed a VR football viewer is at a given moment is hard from pictures alone, since video cannot tell watching from disengagement, and naive concatenation of four signals carries low-quality channel noise into the decision. CMAFNet treats video as a semantic hub, runs bidirectional cross-attention with audio, EEG and eye tracking, weights each modality per sample by real-time signal quality, and uses an ordinal regression loss to keep immersion levels ordered. On 86,400 samples cut from 720 sixty-second VR match clips, four-level classification reaches 81.6% overall accuracy and 0.274 mean absolute error, 3.4 points above and 13.8% below the strongest 2024 baseline, MT-AMW.

Judging how immersed a VR football viewer is at a given moment is hard from pictures alone, since video cannot tell watching from disengagement, and naive concatenation of four signals carries low-quality channel noise into the decision. CMAFNet treats video as a semantic hub, runs bidirectional cross-attention with audio, EEG and eye tracking, weights each modality per sample by real-time signal quality, and uses an ordinal regression loss to keep immersion levels ordered. On 86,400 samples cut from 720 sixty-second VR match clips, four-level classification reaches 81.6% overall accuracy and 0.274 mean absolute error, 3.4 points above and 13.8% below the strongest 2024 baseline, MT-AMW.

OpenAI released GPT-6.1 Terra, a fast, efficient model combining detailed visual reasoning with precise computer use, priced 50% below GPT-5.6 Terra and GPT-6.1 Sol on the standard API. On Agents' Last Exam at medium reasoning effort it scores 15 percentage points above GPT-5.6 Terra at about one-third the cost, and on Terminal-Bench 4.0 at maximum effort it solves 50.3% of tasks versus GPT-5.6 Terra's published 21.5%. OpenAI notes that GPT-6.1 Sol and GPT-6 Astra remain the better choice for the most complex coding tasks.

OpenAI released GPT-6.1 Terra, a fast, efficient model combining detailed visual reasoning with precise computer use, priced 50% below GPT-5.6 Terra and GPT-6.1 Sol on the standard API. On Agents' Last Exam at medium reasoning effort it scores 15 percentage points above GPT-5.6 Terra at about one-third the cost, and on Terminal-Bench 4.0 at maximum effort it solves 50.3% of tasks versus GPT-5.6 Terra's published 21.5%. OpenAI notes that GPT-6.1 Sol and GPT-6 Astra remain the better choice for the most complex coding tasks.

Discrete flow matching (DFM) decouples graph-generation training from sampling, but existing graph DFM samplers often rely on task-specific heuristics and need sampling budgets comparable to diffusion, so the speedup potential is only partly realized. H-GLow decomposes the continuous-time Markov chain generator into hazards and conditional jump distributions, then uses both to plan graph edits and adjust sampling time to graph activity. Adaptable to pretrained graph DFM models, it reports better generation quality with substantially fewer sampling steps on molecular and synthetic graph benchmarks; a companion graph DFM model predicts hazards and conditional jumps directly, with a training objective that decomposes exactly into hazard matching and hazard-weighted jump prediction.

Discrete flow matching (DFM) decouples graph-generation training from sampling, but existing graph DFM samplers often rely on task-specific heuristics and need sampling budgets comparable to diffusion, so the speedup potential is only partly realized. H-GLow decomposes the continuous-time Markov chain generator into hazards and conditional jump distributions, then uses both to plan graph edits and adjust sampling time to graph activity. Adaptable to pretrained graph DFM models, it reports better generation quality with substantially fewer sampling steps on molecular and synthetic graph benchmarks; a companion graph DFM model predicts hazards and conditional jumps directly, with a training objective that decomposes exactly into hazard matching and hazard-weighted jump prediction.

Multi-pinhole SPECT gains sensitivity but stacks overlapping projections with attenuation, scatter, quantum noise and aperture effects, and attenuation correction usually needs an attenuation coefficient map from X-ray CT. A U-Net++ convolutional network takes Monte Carlo-simulated overlapping projections as input and ray-traced ideal projections as target, separating overlaps and correcting the degradations together before maximum-likelihood expectation-maximization reconstruction. On 60 public brain cases under quasi-stationary acquisition, it reached NMSE 0.18±0.02, PSNR 18.42±0.69 dB and SSIM 0.81±0.04, better than 11-pinhole reconstruction and fan-beam reconstruction with 20-fold longer acquisition.

Multi-pinhole SPECT gains sensitivity but stacks overlapping projections with attenuation, scatter, quantum noise and aperture effects, and attenuation correction usually needs an attenuation coefficient map from X-ray CT. A U-Net++ convolutional network takes Monte Carlo-simulated overlapping projections as input and ray-traced ideal projections as target, separating overlaps and correcting the degradations together before maximum-likelihood expectation-maximization reconstruction. On 60 public brain cases under quasi-stationary acquisition, it reached NMSE 0.18±0.02, PSNR 18.42±0.69 dB and SSIM 0.81±0.04, better than 11-pinhole reconstruction and fan-beam reconstruction with 20-fold longer acquisition.

Synthetic electronic health records have long traded clinical fidelity against formal privacy guarantees. SynthEHR-DP combines a mixed-type encoder, a temporal transformer encoder and a conditional variational autoencoder, trained under differentially private stochastic gradient descent with Rényi privacy accounting, then applies a post-hoc clinical constraint layer that spends no privacy budget. On MIMIC-IV at a total budget of ε=8 and δ=10⁻⁵, membership inference AUC is 0.52, downstream TSTR AUROC is 0.786, 4.2% below the non-private TRTR oracle of 0.821, and the clinical validity score is 0.961.

Synthetic electronic health records have long traded clinical fidelity against formal privacy guarantees. SynthEHR-DP combines a mixed-type encoder, a temporal transformer encoder and a conditional variational autoencoder, trained under differentially private stochastic gradient descent with Rényi privacy accounting, then applies a post-hoc clinical constraint layer that spends no privacy budget. On MIMIC-IV at a total budget of ε=8 and δ=10⁻⁵, membership inference AUC is 0.52, downstream TSTR AUROC is 0.786, 4.2% below the non-private TRTR oracle of 0.821, and the clinical validity score is 0.961.

Assistive agents in shared control must act on the human's intent while staying legible enough for the human to respond in real time. Dream2Assist adds intent inference to a recurrent state-space world model and trains the assistive policy through fictitious co-play, anticipating discrete human objectives without overriding human agency. Three human-subjects studies and a preliminary on-vehicle deployment span high-speed racing (N=30) and urban driving with richer intent structure (N=10), where intent conditioning yields measurable improvements in torque alignment and collision rate.

Assistive agents in shared control must act on the human's intent while staying legible enough for the human to respond in real time. Dream2Assist adds intent inference to a recurrent state-space world model and trains the assistive policy through fictitious co-play, anticipating discrete human objectives without overriding human agency. Three human-subjects studies and a preliminary on-vehicle deployment span high-speed racing (N=30) and urban driving with richer intent structure (N=10), where intent conditioning yields measurable improvements in torque alignment and collision rate.

Brain age prediction no longer relies only on diffusion tensor scalar metrics: white matter fiber orientation distributions (WM-FODs) enter directly as spherical harmonic coefficients, processed by an SE(3) equivariant CNN that preserves their transformation properties under rotation. On 692 participants aged 36-90 from the Human Connectome Project Aging dataset, this combined framework outperformed the evaluated DTI-based models and conventional CNNs; averaging five cross-validation models gave a mean absolute error of 3.54 years and a Pearson correlation of 0.958. The advantage depended on combining WM-FODs with equivariant processing, and accuracy was highest at maximum spherical harmonic order Lmax = 2.

Brain age prediction no longer relies only on diffusion tensor scalar metrics: white matter fiber orientation distributions (WM-FODs) enter directly as spherical harmonic coefficients, processed by an SE(3) equivariant CNN that preserves their transformation properties under rotation. On 692 participants aged 36-90 from the Human Connectome Project Aging dataset, this combined framework outperformed the evaluated DTI-based models and conventional CNNs; averaging five cross-validation models gave a mean absolute error of 3.54 years and a Pearson correlation of 0.958. The advantage depended on combining WM-FODs with equivariant processing, and accuracy was highest at maximum spherical harmonic order Lmax = 2.

In glioma Raman spectra, a single spectrum is enough for a model to tell which patient it came from, a sample-specific bias unrelated to biology that can confound tumor typing. On the 46-sample APOLLO dataset, a learnable feature-importance layer is trained adversarially against a two-headed classifier to suppress patient-identity signal while keeping IDH mutation information. Patient-ID balanced accuracy falls from 68% to 12%, while mutant-versus-wild-type classification stays at 0.8 balanced accuracy.

In glioma Raman spectra, a single spectrum is enough for a model to tell which patient it came from, a sample-specific bias unrelated to biology that can confound tumor typing. On the 46-sample APOLLO dataset, a learnable feature-importance layer is trained adversarially against a two-headed classifier to suppress patient-identity signal while keeping IDH mutation information. Patient-ID balanced accuracy falls from 68% to 12%, while mutant-versus-wild-type classification stays at 0.8 balanced accuracy.

A neural collocation scheme for semilinear time-fractional reaction–diffusion equations rewrites the Caputo memory term as cubic-Hermite product integration, using the network state and its autodifferentiated time slope as endpoint data on each time panel while the weakly singular weights are precomputed once before training. On the scalar benchmark v(t)=t⁴.2 the Hermite memory approximation is three to six orders of magnitude more accurate than the classical L1 formula on the tested meshes, and for a manufactured problem with α=0.7 a small CPU-trained network reaches a relative L2 error of 1.98×10⁻⁵ and a maximum error of 3.80×10⁻⁵.

A neural collocation scheme for semilinear time-fractional reaction–diffusion equations rewrites the Caputo memory term as cubic-Hermite product integration, using the network state and its autodifferentiated time slope as endpoint data on each time panel while the weakly singular weights are precomputed once before training. On the scalar benchmark v(t)=t⁴.2 the Hermite memory approximation is three to six orders of magnitude more accurate than the classical L1 formula on the tested meshes, and for a manufactured problem with α=0.7 a small CPU-trained network reaches a relative L2 error of 1.98×10⁻⁵ and a maximum error of 3.80×10⁻⁵.

Predicting the bispectral index (BIS) during balanced anesthesia with sevoflurane or desflurane alongside propofol and remifentanil has relied on fixed population pharmacodynamic parameters. A computational modelling study merged all three drugs into one response-surface model and fitted per-patient sensitivity parameters from 20 real surgical records using artifact-resistant global optimization. With personalized parameters the cohort mean MDAPE was 11.77%, below the 15% clinical viability threshold; a population baseline aggregated from the same cohort gave 16.98%.

Predicting the bispectral index (BIS) during balanced anesthesia with sevoflurane or desflurane alongside propofol and remifentanil has relied on fixed population pharmacodynamic parameters. A computational modelling study merged all three drugs into one response-surface model and fitted per-patient sensitivity parameters from 20 real surgical records using artifact-resistant global optimization. With personalized parameters the cohort mean MDAPE was 11.77%, below the 15% clinical viability threshold; a population baseline aggregated from the same cohort gave 16.98%.

In real-time prediction over streaming data, an investigator often holds several generalized linear models at once, while model averaging is typically a convex optimization over all historical data. ReMAQ replaces that procedure with a quadratic approximation, so when a new batch arrives the averaged prediction is updated sequentially by solving one quadratic program, without accessing the entire history. Under regularity conditions its weights are asymptotically equivalent to oracle model averaging built on the inaccessible full data, which establishes asymptotic optimality; simulation studies and an application to the Airline On-Time Performance Dataset assess the method.

In real-time prediction over streaming data, an investigator often holds several generalized linear models at once, while model averaging is typically a convex optimization over all historical data. ReMAQ replaces that procedure with a quadratic approximation, so when a new batch arrives the averaged prediction is updated sequentially by solving one quadratic program, without accessing the entire history. Under regularity conditions its weights are asymptotically equivalent to oracle model averaging built on the inaccessible full data, which establishes asymptotic optimality; simulation studies and an application to the Airline On-Time Performance Dataset assess the method.

Beam training for modular extremely large antenna arrays in hybrid near- and far-field environments no longer keeps a fixed candidate codebook. Prior subarray schemes relied on predefined codebooks or activation patterns and adapted poorly to spatially non-stationary channels; this scheme builds a parallax-aware initial codebook covering each subarray's angular shift, then learns online with a multi-armed bandit, and once a statistical steady state is detected it projects beams across subarrays and prunes noise-only sectors with a likelihood ratio test, compressing the action space. In simulation the scheme's effective spectral efficiency exceeds three comparison schemes across the evaluated SNR range, with markedly reduced training overhead.

Beam training for modular extremely large antenna arrays in hybrid near- and far-field environments no longer keeps a fixed candidate codebook. Prior subarray schemes relied on predefined codebooks or activation patterns and adapted poorly to spatially non-stationary channels; this scheme builds a parallax-aware initial codebook covering each subarray's angular shift, then learns online with a multi-armed bandit, and once a statistical steady state is detected it projects beams across subarrays and prunes noise-only sectors with a likelihood ratio test, compressing the action space. In simulation the scheme's effective spectral efficiency exceeds three comparison schemes across the evaluated SNR range, with markedly reduced training overhead.

On evidence-sufficiency judgment, a domain with no decidable correctness reward, a boundary-keeping recipe of supervised fine-tuning followed by GRPO on an LLM-judge reward raised spontaneous correct refusal under action pressure from 25% in the instruction-tuned baseline to 100% (40 per arm; Wilson 95% CI lower bound 91.2%, non-overlapping with the baseline interval [14.2%, 40.2%]), while a scrambled-pairing placebo fell to 5%. The same recipe's automatic false refusal on 281 well-formed claims was 1.1% (Wilson upper bound 3.09%), so the effect is boundary-specific rather than blanket refusal; SFT saturates the face, leaving the GRPO increment not separately identifiable, and ratings come from model sessions, not human experts.

On evidence-sufficiency judgment, a domain with no decidable correctness reward, a boundary-keeping recipe of supervised fine-tuning followed by GRPO on an LLM-judge reward raised spontaneous correct refusal under action pressure from 25% in the instruction-tuned baseline to 100% (40 per arm; Wilson 95% CI lower bound 91.2%, non-overlapping with the baseline interval [14.2%, 40.2%]), while a scrambled-pairing placebo fell to 5%. The same recipe's automatic false refusal on 281 well-formed claims was 1.1% (Wilson upper bound 3.09%), so the effect is boundary-specific rather than blanket refusal; SFT saturates the face, leaving the GRPO increment not separately identifiable, and ratings come from model sessions, not human experts.

Parkinson's disease and essential tremor look alike early, with misdiagnosis rates of 20% to 30% even at specialized clinics, and the white-matter tracts mapped by conventional diffusion tensor imaging do not capture microstructural change inside gray matter. The study instead builds diffusion similarity networks, linking brain regions by the statistical similarity of their diffusion profiles, then filters connections with iterative permutation sampling recursive feature elimination and classifies with a graph neural network. In five-fold cross-validation on a local cohort of 100 people, the pipeline separated Parkinson's disease from essential tremor with an AUC of 0.94, and on a separate 40-person hold-out set it reached 87.5% balanced accuracy.

Parkinson's disease and essential tremor look alike early, with misdiagnosis rates of 20% to 30% even at specialized clinics, and the white-matter tracts mapped by conventional diffusion tensor imaging do not capture microstructural change inside gray matter. The study instead builds diffusion similarity networks, linking brain regions by the statistical similarity of their diffusion profiles, then filters connections with iterative permutation sampling recursive feature elimination and classifies with a graph neural network. In five-fold cross-validation on a local cohort of 100 people, the pipeline separated Parkinson's disease from essential tremor with an AUC of 0.94, and on a separate 40-person hold-out set it reached 87.5% balanced accuracy.

Pine wilt disease is hardest to catch before needles visibly discolour, and RGB drone imagery lacks the cues. A multi-scale vision transformer, LMS-ViT, was built for drone hyperspectral data, combining a spectral-spatial fusion module, a Gaussian self-attention mechanism and a multi-scale feature pyramid to model spectral differences between central pixels and neighbours. In a Japanese black pine stand in Shandong, it identified green attack and early stage trees at 87.40% and 88.21% accuracy, with 87.90% overall accuracy, a Kappa of 83.25% and 2.45 M parameters.

Pine wilt disease is hardest to catch before needles visibly discolour, and RGB drone imagery lacks the cues. A multi-scale vision transformer, LMS-ViT, was built for drone hyperspectral data, combining a spectral-spatial fusion module, a Gaussian self-attention mechanism and a multi-scale feature pyramid to model spectral differences between central pixels and neighbours. In a Japanese black pine stand in Shandong, it identified green attack and early stage trees at 87.40% and 88.21% accuracy, with 87.90% overall accuracy, a Kappa of 83.25% and 2.45 M parameters.

Hybrid Bayesian network learning separates skeleton construction from score-based search, and both the search strategy (hill-climbing or tabu) and the scoring function (BDe, LL or BIC) change the learned structure. PCTABU pairs the PC skeleton with tabu search; against PCHC on eight real benchmark networks it did better on the sparser Alarm, Child and Hailfinder, tied on the other five, and LL scored worst for both algorithms while BDe and BIC performed similarly.

Hybrid Bayesian network learning separates skeleton construction from score-based search, and both the search strategy (hill-climbing or tabu) and the scoring function (BDe, LL or BIC) change the learned structure. PCTABU pairs the PC skeleton with tabu search; against PCHC on eight real benchmark networks it did better on the sparser Alarm, Child and Hailfinder, tied on the other five, and LL scored worst for both algorithms while BDe and BIC performed similarly.

Automated history-taking in telemedicine must collect a structured record over many turns while adapting to what the patient answers. MedAgent-RAG keeps an explicit clinical state of filled slots, a ranked differential diagnosis, questions already asked and red flags detected, selects every question from that state, and grounds its final assessment in clinical guidelines retrieved by a parent–child retrieval-augmented pipeline. Evaluated on the MediTOD corpus with a language-model patient simulator, it raises BERTScore F1 from 0.68 for a single-shot baseline built on the same language model to 0.84, at a cost of 15.8 turns on average.

Automated history-taking in telemedicine must collect a structured record over many turns while adapting to what the patient answers. MedAgent-RAG keeps an explicit clinical state of filled slots, a ranked differential diagnosis, questions already asked and red flags detected, selects every question from that state, and grounds its final assessment in clinical guidelines retrieved by a parent–child retrieval-augmented pipeline. Evaluated on the MediTOD corpus with a language-model patient simulator, it raises BERTScore F1 from 0.68 for a single-shot baseline built on the same language model to 0.84, at a cost of 15.8 turns on average.

Exploration in physical control often relies on immediate prediction error or single-step physical change, which can miss information that only emerges over extended interaction. HTC2 trains a horizon-conditioned contrastive critic in a learned Hamiltonian phase space, uses the representation distance between a state-action and its same-episode multi-step futures as the primary curiosity signal, and modulates it with bounded multi-step Hamiltonian variation. Combined with SAC and PPO on six DeepMind Control Suite tasks, HTC2 achieves the highest mean final return in eleven of twelve settings.

Exploration in physical control often relies on immediate prediction error or single-step physical change, which can miss information that only emerges over extended interaction. HTC2 trains a horizon-conditioned contrastive critic in a learned Hamiltonian phase space, uses the representation distance between a state-action and its same-episode multi-step futures as the primary curiosity signal, and modulates it with bounded multi-step Hamiltonian variation. Combined with SAC and PPO on six DeepMind Control Suite tasks, HTC2 achieves the highest mean final return in eleven of twelve settings.

Multi-objective multi-agent reinforcement learning has treated multiple objectives as an extra optimisation dimension stacked on the multi-agent problem. MIRL instead uses that objective structure to guide cooperation: a team preference fixes the system-level trade-off while each agent carries a learnable individual preference, adjusted by meta-gradient according to subsequent team performance. On four multi-objective particle-environment tasks, MIRL-S and MIRL-D improve Pareto-front quality over five baselines, with the dynamic variant MIRL-D strongest overall.

Multi-objective multi-agent reinforcement learning has treated multiple objectives as an extra optimisation dimension stacked on the multi-agent problem. MIRL instead uses that objective structure to guide cooperation: a team preference fixes the system-level trade-off while each agent carries a learnable individual preference, adjusted by meta-gradient according to subsequent team performance. On four multi-objective particle-environment tasks, MIRL-S and MIRL-D improve Pareto-front quality over five baselines, with the dynamic variant MIRL-D strongest overall.

Human vision is thought to use a straightened encoding of the natural visual world as it evolves in time, meaning frame-to-frame changes follow a more linear path in representation space. A study compared network architecture, different types of robustness, biologically inspired filtering and representational straightness, finding that adversarial training yields straighter representations in both CNNs and transformer-based architectures, but the effect is task-dependent and does not extend to segmentation or frame prediction, nor to other kinds of robustness. Straighter representations also give class predictions temporal stability, including on out-of-distribution data.

Human vision is thought to use a straightened encoding of the natural visual world as it evolves in time, meaning frame-to-frame changes follow a more linear path in representation space. A study compared network architecture, different types of robustness, biologically inspired filtering and representational straightness, finding that adversarial training yields straighter representations in both CNNs and transformer-based architectures, but the effect is task-dependent and does not extend to segmentation or frame prediction, nor to other kinds of robustness. Straighter representations also give class predictions temporal stability, including on out-of-distribution data.

Medical vision-language pretraining learns from paired images and radiology reports, but scanner and acquisition-protocol differences can impair transfer. Meta-CLIP uses structured acquisition information in DICOM headers as a training-time prior: a kernel over numerical and categorical fields defines a bounded pairwise margin that reweights cross-modal negatives during pretraining, while inference keeps the standard dual-encoder architecture and needs no metadata. On CT-RATE it improves macro-AUROC over CT-CLIP by 1.41 points, with a 95% paired bootstrap interval of 0.96 to 1.85 points, and reaches the highest macro-AUROC on four of five chest radiography benchmarks.

Medical vision-language pretraining learns from paired images and radiology reports, but scanner and acquisition-protocol differences can impair transfer. Meta-CLIP uses structured acquisition information in DICOM headers as a training-time prior: a kernel over numerical and categorical fields defines a bounded pairwise margin that reweights cross-modal negatives during pretraining, while inference keeps the standard dual-encoder architecture and needs no metadata. On CT-RATE it improves macro-AUROC over CT-CLIP by 1.41 points, with a 95% paired bootstrap interval of 0.96 to 1.85 points, and reaches the highest macro-AUROC on four of five chest radiography benchmarks.

NVIDIA Mission Control 2.5 extends an autonomous recovery engine that previously served only Slurm high-performance computing workloads to Kubernetes workloads managed by NVIDIA Run:ai. Such AI factories had relied on manual intervention, static checkpointing and reactive monitoring, which does not scale to dozens of concurrent multi-node training jobs. The release splits recovery into three coordinated layers: workload placement, job recovery that uses failure history and job topology to choose restart, retry or requeue, and hardware recovery that isolates suspect nodes and runs diagnostics, so recovery need not wait for root-cause analysis.

NVIDIA Mission Control 2.5 extends an autonomous recovery engine that previously served only Slurm high-performance computing workloads to Kubernetes workloads managed by NVIDIA Run:ai. Such AI factories had relied on manual intervention, static checkpointing and reactive monitoring, which does not scale to dozens of concurrent multi-node training jobs. The release splits recovery into three coordinated layers: workload placement, job recovery that uses failure history and job topology to choose restart, retry or requeue, and hardware recovery that isolates suspect nodes and runs diagnostics, so recovery need not wait for root-cause analysis.

NVIDIA has introduced Minerva, an AIOps platform for AI factories that feeds telemetry from GPUs, switches, storage and training jobs into a knowledge graph updated as jobs move, then applies supervised and unsupervised models to catch the gray failures static thresholds miss. In one NVIDIA case, 81 separate events triggered by a switch restart were merged into a single incident, saving about 54 hours of cluster uptime; in another, the platform found one quietly degrading NVLink connection among roughly 400,000 interfaces. These figures come from NVIDIA's internal testing and case descriptions, within its supported scenarios.

NVIDIA has introduced Minerva, an AIOps platform for AI factories that feeds telemetry from GPUs, switches, storage and training jobs into a knowledge graph updated as jobs move, then applies supervised and unsupervised models to catch the gray failures static thresholds miss. In one NVIDIA case, 81 separate events triggered by a switch restart were merged into a single incident, saving about 54 hours of cluster uptime; in another, the platform found one quietly degrading NVLink connection among roughly 400,000 interfaces. These figures come from NVIDIA's internal testing and case descriptions, within its supported scenarios.

Testing egocentric AI assistants has usually required real people, and the expensive, scarce interaction data left open whether today's multimodal models can actually help someone in the physical world. InteractiveWorld moves that loop into a verifiable environment: the assistant issues instructions, a human agent observes the scene, understands the instruction and acts independently, and the environment executes the action and updates what both see. InteractiveBench built on it holds 120 multi-turn tasks across six categories, each instruction shaping the outcome and scored by verifiable state-based rewards; across 12 frontier multimodal models, Claude Fable 5.1 reaches at most 20% and 25% on long-horizon Guidance and Planning tasks, respectively.

Testing egocentric AI assistants has usually required real people, and the expensive, scarce interaction data left open whether today's multimodal models can actually help someone in the physical world. InteractiveWorld moves that loop into a verifiable environment: the assistant issues instructions, a human agent observes the scene, understands the instruction and acts independently, and the environment executes the action and updates what both see. InteractiveBench built on it holds 120 multi-turn tasks across six categories, each instruction shaping the outcome and scored by verifiable state-based rewards; across 12 frontier multimodal models, Claude Fable 5.1 reaches at most 20% and 25% on long-horizon Guidance and Planning tasks, respectively.

Camera–LiDAR bird's-eye-view fusion can mistake corrupted visual texture for useful evidence when the camera stream degrades. DARFusion recasts fusion as modality trust estimation: learned reliability cues calibrate both feature streams, and calibrated features, task-driven contexts and a compact degradation descriptor derived from camera BEV features decide how much camera evidence to use. It holds 68.4 mAP and 71.5 NDS on the clean nuScenes validation set and beats the reproduced BEVFusion baseline by 1.41 and 1.31 mAP on real rain and night subsets.

Camera–LiDAR bird's-eye-view fusion can mistake corrupted visual texture for useful evidence when the camera stream degrades. DARFusion recasts fusion as modality trust estimation: learned reliability cues calibrate both feature streams, and calibrated features, task-driven contexts and a compact degradation descriptor derived from camera BEV features decide how much camera evidence to use. It holds 68.4 mAP and 71.5 NDS on the clean nuScenes validation set and beats the reproduced BEVFusion baseline by 1.41 and 1.31 mAP on real rain and night subsets.

For behavioral classes that arrive over time, TRACE stores no historical samples and never expands a shared classifier; it feeds a frozen temporal-spectral representation to a small autoencoder per class and labels by reconstruction error. Across five multivariate time-series class-incremental benchmarks it reaches 88.85% five-dataset average final accuracy with 9.17 MiB of retained memory, about 98.26% less memory than TS-ACL, the strongest analytical baseline by average accuracy, and an approximately 33.43% relative accuracy gain on GrabMyo.

For behavioral classes that arrive over time, TRACE stores no historical samples and never expands a shared classifier; it feeds a frozen temporal-spectral representation to a small autoencoder per class and labels by reconstruction error. Across five multivariate time-series class-incremental benchmarks it reaches 88.85% five-dataset average final accuracy with 9.17 MiB of retained memory, about 98.26% less memory than TS-ACL, the strongest analytical baseline by average accuracy, and an approximately 33.43% relative accuracy gain on GrabMyo.

Command-line AI agents such as Claude Code and Codex ship with search and web reading, yet they largely cannot reach platforms that require login, video subtitles and speech transcripts, Chinese platforms, or clean structured data. Agent Reach preinstalls and tests the upstream tools behind those platforms, so an agent calls them with one command in the form channel, action, arguments, and falls back to the next tool in an ordered list when the first fails. It only searches and reads, never logs in for you, and never asks for your password.

Command-line AI agents such as Claude Code and Codex ship with search and web reading, yet they largely cannot reach platforms that require login, video subtitles and speech transcripts, Chinese platforms, or clean structured data. Agent Reach preinstalls and tests the upstream tools behind those platforms, so an agent calls them with one command in the form channel, action, arguments, and falls back to the next tool in an ordered list when the first fails. It only searches and reads, never logs in for you, and never asks for your password.

VRATF moves the comparison in retrieval-augmented forecasting from raw time series to denoised trend line charts: FFT low-pass filtering extracts the low-frequency trend, which is rendered as a 224×224 image and encoded by a frozen CLIP ViT-B/32 to build the retrieval database, with top-K neighbors selected by cosine similarity and aggregated by Gaussian-kernel weights. The retrieval forecast is fused at the prediction level with a DLinear base forecast through one learnable scalar. Across six multivariate benchmarks, VRATF reaches the lowest average MSE on four, with 0.340 on ETTh2 against 0.358 for the next-best RAFT, a 5.0% improvement.

VRATF moves the comparison in retrieval-augmented forecasting from raw time series to denoised trend line charts: FFT low-pass filtering extracts the low-frequency trend, which is rendered as a 224×224 image and encoded by a frozen CLIP ViT-B/32 to build the retrieval database, with top-K neighbors selected by cosine similarity and aggregated by Gaussian-kernel weights. The retrieval forecast is fused at the prediction level with a DLinear base forecast through one learnable scalar. Across six multivariate benchmarks, VRATF reaches the lowest average MSE on four, with 0.340 on ETTh2 against 0.358 for the next-best RAFT, a 5.0% improvement.

Hearing aids must denoise and recognize speech within tight space and battery limits, and high-order finite impulse response filters cost area and power through many multiply-accumulate operations. This architecture places a memory-less, multiplier-less distributed-arithmetic FIR filter and a deep spectro neural network in one application-specific integrated circuit design, with inference using an approximate Wallace tree multiplier. Post-synthesis analysis reports 160,000 square micrometres, 1.4 milliwatts and about 2.1 milliseconds of hardware processing latency, while functional validation stays in MATLAB and prototype fabrication and real-time hardware validation remain future work.

Hearing aids must denoise and recognize speech within tight space and battery limits, and high-order finite impulse response filters cost area and power through many multiply-accumulate operations. This architecture places a memory-less, multiplier-less distributed-arithmetic FIR filter and a deep spectro neural network in one application-specific integrated circuit design, with inference using an approximate Wallace tree multiplier. Post-synthesis analysis reports 160,000 square micrometres, 1.4 milliwatts and about 2.1 milliseconds of hardware processing latency, while functional validation stays in MATLAB and prototype fabrication and real-time hardware validation remain future work.

India's Punjab Vidhan Sabha records from 1950 to 2024 survive as roughly 200,000 scanned pages mixing Hindi, Punjabi and English, where broken layouts and OCR errors propagate into retrieval and question answering. The system couples layout segmentation and multi-script OCR with an agentic retrieval-augmented generation pipeline, in which an evaluator model judges whether retrieved evidence suffices and rewrites the query when it does not. On a stratified benchmark of 1,000 pages and 200 gold-annotated question-answer pairs, the optimized pipeline reaches MRR@1 of 0.810 and Hit@3 of 0.925; raising the retrieval-rewrite iteration cap from 1 to 3 lifts answer F1 from about 0.18 to about 0.35.

India's Punjab Vidhan Sabha records from 1950 to 2024 survive as roughly 200,000 scanned pages mixing Hindi, Punjabi and English, where broken layouts and OCR errors propagate into retrieval and question answering. The system couples layout segmentation and multi-script OCR with an agentic retrieval-augmented generation pipeline, in which an evaluator model judges whether retrieved evidence suffices and rewrites the query when it does not. On a stratified benchmark of 1,000 pages and 200 gold-annotated question-answer pairs, the optimized pipeline reaches MRR@1 of 0.810 and Hit@3 of 0.925; raising the retrieval-rewrite iteration cap from 1 to 3 lifts answer F1 from about 0.18 to about 0.35.

Tool-using language model agents can refuse one phrasing of a harmful request and comply with another, and a missed refusal becomes an action. Llama-3.2-1B-Instruct was fine-tuned with LoRA on data not drawn from AgentHarm, comparing latent attribution regularization (LAR) against an augmentation-only control over three seeds. At 10% over-refusal on harmless tool tasks, LAR lowered the share of tasks with mixed refusal decisions from 25% to 19% and paraphrase-induced flips from 11.3% to 9.3%, while harmful-versus-harmless detection did not change significantly.

Tool-using language model agents can refuse one phrasing of a harmful request and comply with another, and a missed refusal becomes an action. Llama-3.2-1B-Instruct was fine-tuned with LoRA on data not drawn from AgentHarm, comparing latent attribution regularization (LAR) against an augmentation-only control over three seeds. At 10% over-refusal on harmless tool tasks, LAR lowered the share of tasks with mixed refusal decisions from 25% to 19% and paraphrase-induced flips from 11.3% to 9.3%, while harmful-versus-harmless detection did not change significantly.

OpenBot places coding agents such as Codex, Claude Code, Grok CLI, OpenCode, Gemini, Cursor CLI and Cline CLI in one local-first desktop workspace, where each agent keeps its own workspace and conversation while sharing local queues, file transfers, an embedded browser and agent-to-agent messaging. It also offers shared desktop channel chats with one task owner, explicit delegation, and Stop, Resume, Reassign, Archive and Restore controls. The release is a development preview: after first-launch consent, agents run with danger-full-access and approvalPolicy: never, so they can read and modify files, run commands, use the network and control the browser without per-action confirmations.

OpenBot places coding agents such as Codex, Claude Code, Grok CLI, OpenCode, Gemini, Cursor CLI and Cline CLI in one local-first desktop workspace, where each agent keeps its own workspace and conversation while sharing local queues, file transfers, an embedded browser and agent-to-agent messaging. It also offers shared desktop channel chats with one task owner, explicit delegation, and Stop, Resume, Reassign, Archive and Restore controls. The release is a development preview: after first-launch consent, agents run with danger-full-access and approvalPolicy: never, so they can read and modify files, run commands, use the network and control the browser without per-action confirmations.

JetBrains has released ThinkRail, an agentic IDE that places coding agents in separate workspaces cut from a git repository, each with its own branch and working directory, so agents can work in parallel without touching the developer's checkout. It wraps the pi coding agent in an editor, terminals and review surface, and lets the agent keep project intent through a graph of SPEC.md files beside the code. Merging the good branch and deleting the rest closes the loop.

JetBrains has released ThinkRail, an agentic IDE that places coding agents in separate workspaces cut from a git repository, each with its own branch and working directory, so agents can work in parallel without touching the developer's checkout. It wraps the pi coding agent in an editor, terminals and review surface, and lets the agent keep project intent through a graph of SPEC.md files beside the code. Merging the good branch and deleting the rest closes the loop.

AIPOCH Open-Science v1.0.3 connects seven scientific sources, spanning cancer imaging, pathways, lipids, structures, molecular regulation and cancer cell-line dependencies, into one local-first research workbench, and folds literature discovery, attachments and screening export into a single inbox. These sources previously differed in identifiers, query conventions and access controls, and connection was an assumed link; the release keeps discovery default-off and persists each source choice, making availability an explicit configuration. The exported screening CSV carries final verdicts, manual decisions, reasons and model probabilities, but those fields support spreadsheet review and do not establish study quality or scientific truth.

AIPOCH Open-Science v1.0.3 connects seven scientific sources, spanning cancer imaging, pathways, lipids, structures, molecular regulation and cancer cell-line dependencies, into one local-first research workbench, and folds literature discovery, attachments and screening export into a single inbox. These sources previously differed in identifiers, query conventions and access controls, and connection was an assumed link; the release keeps discovery default-off and persists each source choice, making availability an explicit configuration. The exported screening CSV carries final verdicts, manual decisions, reasons and model probabilities, but those fields support spreadsheet review and do not establish study quality or scientific truth.