Daily report
Only content delivered through the publication boundary on this date is included.
Sealed edition v1 · 1 withdrawn or replaced articles omitted
CRBugBench tests code review agents against 1,800 human-fixed bugs: the best configuration catches 79.6% of pre-merge bugs within its top-5 comments, but performance drops noticeably on proprietary industry repositories.
An audit layer added to VLMEvalKit reuses saved per-question outputs from completed evaluations plus one text-only rerun, and on MMStar measures a 20.2% zero-gain share that still exceeds an item-matched permutation null.
Linear injection probes could reach 0.91–0.98 AUROC on the standard evaluation using conversation position and text length alone; adding same-moment clean twins and harmless look-alikes to training cuts false alarms on injection-free conversations from 41–80% to 4–27% across six open models.
When a latent world model ranks candidate actions by terminal Euclidean distance to a goal latent, an acute bend at a mandatory junction can make it prefer a candidate that moves backward over one that makes real progress, even with exact rollouts.
A user simulator now writes a first-person evaluation of the assistant before producing its next line, including a 1-to-5 rating and a critique-or-praise judgment; on real user-ChatGPT logs it tracks human users better than prompted simulators of the same size, and it lets multi-turn RL train assistants to comparable quality with less than half the compute.
A backprop-free training path reaches FID 3.64 on unconditional CIFAR-10 against 3.54 for backpropagation, the first CIFAR-scale generative model trained with equilibrium propagation.
A world action model can regenerate complete rollouts of previously learned tasks from their instructions alone, and after truncation and physical-consistency filtering, replaying them cuts the forgetting of sequential fine-tuning by over 90% on four LIBERO suites and five real-world tasks without keeping any data from previous tasks.
The action-conditioned world model HF-WM trains on egocentric human manipulation video together with perturbed failure actions, reaching state-of-the-art action-conditioned video generation and predicted success rates that track real-world deployment.
Across vision models, language models and multitask recurrent networks, networks that show no unit-level modularity still yield selective ablation of content versus style, syntax versus semantics, and individual cognitive tasks once an invertible linear transformation is applied.
Functionality-preserving code transformations mined from real-world commits cut frontier vulnerability-analysis agents' mean success on CyberGym tasks from 63.3% to 36.8%, while comparable synthetic obfuscations only bring it to 49.7%.
A credit-assignment framework that redistributes delayed reward by explanation rather than temporal proximity traces sparse terminal reward back to the actions that actually mattered hundreds of steps earlier, across Crafter, WordCraft, KuaiRand, and an industry-scale generative recommender.
Instead of letting an LLM guess tool combinations from training memory, the red-teaming framework X-AgentRed samples cross-service paths from a tool-dependency graph, reaching a 19.2% average harmful path rate across four frontier agents in a simulated ecosystem of 30 services and 914 tools, above STAC's 10.5% and TRACE's 2.7%.
FlowCred splits the single terminal reward of joint text-to-video post-training into prompt-level credit and prompt-conditioned generator residuals, lifting a strictly matched joint-training control on LTX-2.5 from 61.47 to 64.46 on T2V-CompBench without increasing decoded-video or reward-model-call budgets.
VEPA writes an agent guardrail as executable symbolic rules, has an LLM propose patches from deployment feedback while a deterministic verifier admits only structurally valid, conflict-free policies, and reaches the highest F1 on events covered by its active policy across three agent safety benchmarks while expanding the scope of deterministic enforcement during deployment.
A pretrained vision-language model can act without robot-action pretraining once a tool-center-point action frame is projected into the wrist-camera view and actions are emitted through its native text interface, reaching 79.4% average zero-shot success on simple manipulation tasks, close to VLA models fine-tuned with about 30 demonstrations.
Long-context RLVR rollout generation now drafts tokens with sparse attention and verifies them with full attention, raising end-to-end training throughput to 1.39x and 1.35x of autoregressive on Qwen3-1.7B and Qwen3-8B while keeping comparable task performance.
When geo-distributed LLM serving schedulers account for an engine's batching and queueing state, they cut carbon emissions by up to 60% relative to round-robin under MLPerf per-token latency targets, with an average SLO violation of about 3%.
Encoding motion and fingertip contact forces from a single demonstration into an ordered prompt lets a frozen policy run wiping, rotation, and insertion on unseen objects at 78.7% overall success, 13.0 percentage points above the strongest baseline.
AgentCIScope turns privacy cases into 936 executable agent tasks across 11 application environments and finds single-run leakage rates from 32.5% to 94.4% across 12 agents, so strong task completion does not imply contextual integrity.
Variational quantum band-structure calculations for periodic solids have been limited by the input step: each material needs a second-quantized Hamiltonian built by density-functional theory plus Wannierization or by hand-fitted parameters. SlaKoNet-VQD replaces that step with a pretrained universal Slater-Koster tight-binding model, reproducing all eight silicon bands along the standard high-symmetry path with a mean absolute deviation of 1.78 meV from exact diagonalization of the same Hamiltonian. The same workflow, without retraining, covers Al, Ta, V, Nb and ZrN, and a representative aluminum k-point ground state runs on IBM Quantum hardware with an error of about 0.37 eV.
Time-series foundation models forecast well zero-shot, yet deploying them still means hand-building input and output pipelines. TimeRSI has a coding agent repeatedly rewrite preprocessing, features, context selection and post-processing around a frozen forecaster, using errors on historical forecast windows to guide each revision, then picks among the resulting pipelines at deployment by similar past conditions. Across 198 tasks on TIME and FEV-bench it beats all published submissions on six point and probabilistic metrics; against direct inference with the same frozen TimesFM 3.0, TIME MASE falls 1.52% and CRPS 1.61%.
Mistral released Timestral, its first family of time-series foundation models, whose largest version at 209M parameters ranks first on the public GIFT-Eval, TIME and fev-bench forecasting benchmarks while using two-thirds the parameters of the previous state of the art. It produces probabilistic multivariate forecasts with past-only and known-future covariates, spans four sizes from 5M to 209M parameters, and needs no GPU: on an 8-thread server CPU, a one-day-ahead forecast from three weeks of 15-minute history takes 10 ms at 5M and 140 ms at 209M.
Language model agents are increasingly handed the post-training stack, yet a single agent converges on supervised fine-tuning and spends its effort reweighting data categories. AutoSFT splits the loop into an executor, four analysis subagents that read each round from training data, failures, training signals and history, and a synthesis subagent that writes new examples for undersupplied, verifier-checkable capabilities. Every round retrains from the same base checkpoint, so each component is attributable: analysis reaches 63.33 on IFEval at round six, above the single agent's best of 62.35 across 50 rounds.
Collaboration among pretrained unimodal models can become less accurate than a surviving unimodal model when modalities are missing. Lluna keeps each model's direct prediction as a private path, adds a graph-based common branch as a residual correction, and lets richer availability states guide poorer ones along single-modality removal chains. Across eight datasets in affective computing, scene understanding and medical prediction, it improves Mean-Missing performance by 4.26% on average over the strongest evaluated baseline, with positive collaboration gains at minimal inputs on all eight tasks.
High-resolution Adriatic ocean reanalysis fields previously came from bilinear interpolation or costly nested regional models, and interpolation smooths away fronts and mesoscale structure. OcDiffSR applies a conditional denoising diffusion probabilistic model to jointly super-resolve sea-surface temperature, salinity, and horizontal velocity, using 1/12° GLORYS12V1 fields as conditioning input to reconstruct 1/24° Med MFC fields. On the independent test year 2009, scalar-field errors fall well below interpolation and the residual diffusion model CorrDiff, with salinity RMSE of 0.346 psu and temperature RMSE of 0.477°C, while velocity correlation stays near 0.64.
Switzerland's Apertus project released 1.5 versions of its 8B and 70B models, continuing pretraining from 1.0 to add image and audio understanding, an optional thinking mode, and a context window of 262,144 tokens. Multimodality arrives by early fusion: images and audio become discrete tokens placed in the same sequence as text, with no separate vision encoder, so the architecture is unchanged and training runs at text-only speed. The models read images and audio but generate only text, speech handling remains experimental, and the project says the release still trails frontier open-weight models.
In outcome-dependent sampling, covariates are collected for only part of the cohort, so standard survival models trained on the sampled data target a distribution that differs from the cohort. A gradient boosting machine carries the sampling weights into training and tuning, selecting iterations by a weighted integrated Brier score, and the weighted empirical risk is shown to converge to the corresponding integrated Brier risk. In simulations it outperformed the Cox model in high-dimensional and nonlinear settings, and on uranium miner cancer data the test C-index was 0.803 versus 0.801 for Cox.
For debris fragments too small to catalog, radar detection is coupled to a Bayesian population model that picks the next elevation after each observation. Across 300 simulated worlds with 24 observations each, service-weighted adaptive pointing left 41.2% less uncertainty than fixed pointing, 15.4% less than random pointing and 5.35% less than a preplanned multi-angle schedule; service weighting itself added only 1.55% over unweighted adaptation, with a 95% interval spanning zero.
On 783 de-identified free-text mammography, hip-knee-ankle and brain MR reports, an LLM Council of three general small models plus a chairman reached 83.56% overall accuracy, below the medically trained MedGemma 4B at 88.72%, with the gap driven mainly by hip-knee-ankle classification. A single council member, Qwen3-4B-2507, reached the highest accuracy at 91.69%, above both the council and MedGemma, while the council needed four model calls per report and about 2.44 times MedGemma's peak GPU memory.
Across two independent auditory event-related potential VR studies, deep models classifying cybersickness-related discomfort repeatedly flagged the same scalp regions, left frontal, central and right parietal, and an early window roughly 80 to 260 milliseconds after the stimulus. Interpretability maps attributed each model decision to specific electrodes and time steps, and related patterns appeared in data collected in different countries with different headsets, EEG systems and VR tasks. Cross-subject generalization stayed limited, while per-subject calibration raised accuracy sharply, so these features are best treated as testable candidate markers.
After porting the quantum-chemistry pipeline of a generative model writing circuits, a quantum sampler drawing bitstrings, classical repair expanding them and a classical solver picking the best, to maximum weighted independent set, a preregistered held-out test returns a negative answer: tabu search reaches repair-basin mass 0.82–0.94 and simulated annealing comes second, while every quantum sampler and a matched learned classical policy trails both at every size. An earlier positive-looking result came from picking the best of 50 configurations per instance, an upper bound rather than a deployable method; under random configurations, oracle QAOA reaches only 0.014 at n=28.
Age-associated gene expression signals in breast tumors narrow from 80 to four candidates, IL7R, SQSTM1, EGR1 and IGF1, after receptor and stage adjustment, cross-platform replication and estrogen-receptor stratification. The same molecular workflow is also encoded as a bounded symbolic controller that completes, blocks or fails safe against a prespecified reference policy across 760 scenarios, rather than only finishing. The output is research prioritization, not a clinical biomarker or treatment recommendation.
Distribution-network voltage transformers have been checked mainly by de-energizing and comparing them against standard units, which is slow and interrupts supply. A State Grid Wuhan team embedded switching-station Kirchhoff voltage constraints and a cable thermal-balance equation as soft loss terms in a neural network, added radial-basis multi-scale time features and residual attention weighting, and estimated transformer error online. On a laboratory physical simulation platform the method cut RMSE by 35.2% and MAE by 36.4% and raised R² by 2.63 points over the runner-up MRSA-DGCN; in a field pilot in a 10 kV Wuhan area it flagged one out-of-tolerance transformer, matching the de-energized check.
Arctic sea ice concentration forecasting has been limited by the difficulty of fusing multi-timescale information, with statistical models missing nonlinear relationships and numerical models depending on initial conditions. STICEnet replaces the U-Net encoder-decoder path with strictly unidirectional temporal causal and transposed convolutions and embeds channel and spatial attention in the skip connections, trained on NSIDC sea ice concentration and ERA5 atmospheric and oceanic variables. Across one- to six-month leads its RMSE, MAE and R2 all beat CNN, ConvLSTM, IceNet and MT-IceNet, reaching 7.51% RMSE and 0.925 R2 at month six.
A two-layer, 427k-parameter transformer shows grokking on the finite-field discrete logarithm map: after a long memorization phase it reaches 99.8% held-out accuracy at p = 113. A 12-model capacity sweep from 90k to 14.4M parameters finds grokking recurring at multiple scales, with a non-monotonic region whose outcome changes when training data increase; hidden states grow aligned with multiplicative-character structure, and prime-order and disjoint-generator controls also generalize.
AI systems for scientific discovery usually learn only from published positive results, leaving them structurally blind to failures and suppressed directions. A three-layer operational protocol turns that gap into executable steps: it scores theoretically possible but undocumented relationships in a knowledge graph, uses an adversarial classifier to surface features systematically suppressed in unpublished, rejected or failed data, and generates hypotheses that diverge from consensus yet pass viability filters. In a CPEB3/CPEB4 neurobiology case, it flagged four structural blind spots and produced a testable hypothesis about astrocytic translation during sleep and active forgetting.
Automatic dance quality assessment has largely followed action recognition, judging which dance is performed rather than how well. RSDQ-Net maps target and reference performances onto a 19-node canonical skeleton, learns soft phase correspondence, models spatial execution deviation and temporal phase deviation separately, and constrains one latent quality variable with ordinal expertise labels and human pairwise preferences. Under a choreography-disjoint five-fold protocol on ImperialDance it reaches a QWK of 0.814 and ordinal MAE of 0.207, against 0.755 and 0.261 for the strongest baseline, 2T-GCN.
Knee osteoarthritis diagnosis draws on multimodal data scattered across institutions, and centralizing it is both costly and privacy-risky. An asynchronous federated scheme first fuses structured records, semi-structured files and imaging into a queryable data table, then has ten simulated institutions compute gradients on vertically partitioned feature columns, with Bulyan aggregation and adaptive differential privacy filtering malicious updates on the server. On 5,580 knee osteoarthritis cases it reports 99.64% accuracy, 0.998 AUC, 285 MB communication overhead and a 1.8% attack success rate.
LangChain released a set of LangSmith features at Interrupt NYC and London so teams can own their intelligence in production. Previously, finding conversations meant relying on traditional filters; AI Search now retrieves traces by semantic understanding, so a team can ask whether a user got frustrated or whether PII leaked. The same update adds Managed Deep Agents v0.9, Engine v2 with red teaming, Trajectories, Custom Apps, and Fine-Tuning, covering how agents are built, understood, and improved.
Many variants found in monogenic diabetes genes remain variants of uncertain significance because evidence is insufficient or conflicting. A pipeline trained on 7,706 confidently classified single nucleotide variants across 22 such genes reached ROC-AUC 0.9989 and AUPRC 0.9970 on 771 held-out variants, with recall 1.00 on the pathogenic class. Feature ablation showed that part of this accuracy reflects overlap with the ACMG criteria used to assign ClinVar labels, yet on non-missense variants, where computational evidence could not have been directly invoked, the model still beat a majority-class baseline, 99.0% versus 87.0%; it then ranked 6,306 uncertain variants with per-variant explanations.
ai-memory moves a coding agent's long-term memory out of a single tool and machine and onto a server the user runs, so more than twenty coding harnesses read and write one shared store. Previously Claude Code, Cursor and others kept their own notes, which vanished when you switched tools or handed work to a teammate. The source of truth is a git-backed wiki of ordinary Markdown files, with the database as a rebuildable index, and the default path makes no LLM calls, so capture, search and handoffs need no API key.
For clinical diagnosis from multi-device wearable sensors, a cross-platform framework chains stream selection, class-imbalance correction and spatiotemporal feature learning into one pipeline. Previously, sensor modalities, sampling rates and transmission properties were handled separately, rare abnormal samples were scarce, and inference shifted with the device. On two public datasets the framework reaches AUC 0.966 and 0.947; at 24 streams it reports accuracy 0.9348, precision 0.9305, sensitivity 0.9615 and false-positive rate 0.0915.
In downlink NOMA, power allocation has relied on fixed, equal, or centralized learned rules that struggle to serve weak-channel users and throughput at once. A Kampala International University team models each user as an independent learning agent, coordinates them with a shared reward over local observations, and builds imperfect-SIC residual interference into the learning environment. In a three-user simulation the scheme reached about 18 bps/Hz and a fairness index of 0.96 at 30 dB SNR, with outage below 10% at 10 dB.
The open-source coding agent OpenCode ships two built-in agents: build, the default full-access agent for development work, and plan, a read-only agent that denies file edits by default and asks permission before running bash commands, aimed at exploring unfamiliar codebases or planning changes. A Tab key switches between them, and a general subagent invoked with @general handles complex searches and multistep tasks. Installation spans a shell script, npm, brew, scoop, pacman, mise and nix, with a desktop application also available.
DeepSeek V4.1 Flash 1017 is now available on AI Gateway as a minor update to DeepSeek V4.1 Flash, improving reasoning, coding, agentic tasks and image understanding. Coding gains appear in terminal work and in building and editing whole repositories, agentic gains in multi-step automation, and with images the model reads charts and handles visual reasoning more accurately when it can call tools. It also finishes tasks faster and with shorter outputs than the previous snapshot, so agent runs tend to take less time and use fewer output tokens.