Event History
Browse verified industry events and important research by month, area, or content type.
Filter by domain
Start with all, official, or research. Expand to filter by domain.
BenchMIRT: What are LLM benchmarks actually measuring?
Hugging Face published a blog post titled 'BenchMIRT: What are LLM benchmarks actually measuring?' on 2026-09-01.
NEW · Updated Sep 1, 2026 · DeepSoftwareAnalyticsEfficient SWE Agent Benchmarking via Trajectory-Aware Evaluation
PTA-IRT, a Privileged Trajectory-Aware Item Response Theory framework, fuses process and outcome signals from historical execution trajectories to improve calibration subset selection and ability estimation for software engineering agent benchmarks. Under low calibration budgets, PTA-IRT outperforms prior IRT baselines on score and ranking recovery across four SWE benchmarks. Code and data are publicly available at https://github.com/DeepSoftwareAnalytics/PTA-IRT.
NEW · Updated Sep 1, 2026 · ACToRAdaptive Critical Token-Aware Retrieval for Repository-Level Code Generation
The paper proposes ACToR, an adaptive critical token-aware retrieval framework for repository-level code generation. It identifies critical tokens during generation and triggers targeted retrieval on demand to provide repository context at these decisive positions.
NEW · Updated Sep 1, 2026 · CordisBenchCordisBench: Can Language Models Reason About Component Lifecycles in Dynamic Agent Harnesses?
CordisBench is a 1,200-question benchmark introduced to evaluate language models' reasoning about component lifecycles in dynamic agent harnesses. It combines a controlled formal setting with programs executed against Cordis, a runtime managing component dependencies and cleanup. The benchmark asks models to identify affected components, predict state after a specified teardown order, determine conditions under all or some orders, and choose reconfigurations that succeed. Three efficiency-oriented models were evaluated at low reasoning effort with 2, 4, 8, 16, 24, or 32 relevant interactions, using deterministic task-specific scoring. Models generally handle small systems well but become less reliable as more interactions become relevant, especially in predicting final state and reasoning across teardown orders. Additional inference effort recovers marked gains for some models. On the 16-interaction subset, GPT-5.6 Luna uses nearly 3,000 reasoning tokens per question.
NEW · Updated Sep 1, 2026 · arXivThe Rise of Verbal Reinforcement Learning
A paper titled 'The Rise of Verbal Reinforcement Learning' was published on arXiv on 2026-09-01. It proposes Verbal Reinforcement Learning (VRL) as a unified paradigm where natural language serves as a primary feedback channel for improving language agents. The paper organizes VRL into three pillars: Language as Grounding Signal, Language as Deliberative Feedback, and Language as Learning Signal.
NEW · Updated Sep 1, 2026 · arXivMechanism Design for Alignment and Control
A framework for mechanism design with AI agents whose alignment and capabilities are unknown is developed. The framework incentivizes honesty and obedience, uses a one-sided imitation structure, and yields a revelation principle, characterization of implementable policies via nested cyclical monotonicity, and conditions for disciplining multiple agents via higher-order beliefs. Applications include sandbagging, alignment-interpretability trade-off, peer scoring, competition-inducing rewards, and scalable oversight.
NEW · Updated Sep 1, 2026 · arXivDesigning Proactive Thought Partners for Writing
A study deployed a technology probe with 16 participants for one week to explore proactive AI writing partners. The probe allowed users to configure partner roles and proactivity, and partners offered suggestions at appropriate moments. Participants used suggestions for idea generation and self-monitoring, and valued lightweight visual representations and non-directive rhetorical framing.
NEW · Updated Sep 1, 2026 · arXivScaling Near-Optimal SFT-RL Annotation Budget Allocation from Small to Large LLMs
A paper on arXiv (2609.01573v1) frames SFT-RL annotation budget allocation in terms of near-optimality, showing the near-optimal region is wide, widens with model scale, and transfers from small proxy models to large target models. Results hold across tasks, model families, and both preference-based off-policy and reward-supervision on-policy RL methods.
NEW · Updated Sep 1, 2026 · SAGESelective Agent Guidance via Entropy: Learning Autonomous Policies from Imperfect VLM Teachers
SAGE (Selective Agent Guidance via Entropy) is a framework that queries a Vision-Language Model (VLM) only when the learner is uncertain, executes the suggested action during training, and distills guidance into a lightweight Reinforcement Learning (RL) policy. Across sparse-reward visual reasoning and navigation tasks, SAGE learns policies that act without VLM guidance at evaluation time and improves over unguided RL in several environments, including settings where the learned policy exceeds its VLM teacher.
NEW · Updated Sep 1, 2026 · arXivFrom Confusion to Clarity: Confusion-Aware Retrieval and Knowledge Injection for Text Classification
A framework identifies confusable label pairs, expands candidate sets, and generates targeted rules to differentiate similar labels for LLM text classification. It requires no fine-tuning and transfers to smaller models. On WOS, Flipkart, and LEDGAR benchmarks, it improves Macro F1 by up to 10.0pp over retrieval baselines, with smaller models (2B-20B) gaining up to 11.5pp via cross-model transfer.
NEW · Updated Sep 1, 2026 · MiniMax-H3H3-World: Turning Language Understanding into World Control
H3-World is a framework that turns the 33B MiniMax-H3 video generator into an interactive world model. It uses natural-language instructions for zero-shot control of character behavior and camera motion. Actions are represented as structured combinations of character and camera instructions aligned with temporal video latents. Temporal attention routing restricts each instruction to its intended time interval. The framework reuses semantic representations from large-scale video pretraining and requires only lightweight adaptation: 8,000 gameplay samples, 10,000 LoRA optimization steps, and 0.199% trainable parameters.
NEW · Updated Sep 1, 2026 · arXivRetrieved but not ranked: surface-form bias in structural retrieval, from mathematics to agent trajectories
A study evaluates embedding retrieval where surface form and meaning are separated, using competition mathematics (MathNet-Retrieve; 500 queries, 117,088-item corpus) and embodied-agent trajectories (ALFWorld-derived; 118 queries, 336 trajectories). In mathematics, strict Hit@1 at the heaviest disguise tier is 0.0% for both production embedders, while the correct item is in the top 10 nearly always; in 95.2-99.8% of misses the winner is more lexically similar to the query than the correct answer. In trajectories, models land at or near hypergeometric chance when gold involves a different object, and below chance when gold differs in object and receptacle. A lexical reranker control hurts in mathematics and helps in trajectories (closing 26-36% of the gap).
NEW · Updated Sep 1, 2026 · autoPET/CT V challengeBS: Take the Hint - Interactive Multitracer PET/CT Lesion Segmentation with a Scribble-Conditioned ResEnc U-Net
A scribble-conditioned residual encoder U-Net for interactive multitracer PET/CT lesion segmentation was submitted to the autoPET/CT V challenge. The model uses four input channels: CT, PET, and sparse scribble maps for foreground and background. It is initialized from autoPET-III winning weights, with scribble channels zero-initialized to preserve pretrained representation. PET intensities are normalized using a per-scan aorta blood-pool reference from CT segmentation. Five fold models are ensembled by averaging at inference.
NEW · Updated Sep 1, 2026 · SCILAWS-BENCHCan LLMs Discover Scientific Laws in Real and Parallel Worlds?
A benchmark called SCILAWS-BENCH is introduced for scientific law discovery, built from published research and real scientific data. It comprises 118 problems drawn from 381 scientific papers, covering 291 candidate laws and roughly 8M real data points across six scientific disciplines. Each problem is instantiated in two settings: SCILAWS-REAL asks models to propose laws from fixed real observations and evaluates held-out predictive fit and scientific validity; SCILAWS-PARALLEL asks models to actively query residual-calibrated worlds and recover synthesized hidden laws.
NEW · Updated Sep 1, 2026 · arXivA Mathematical Theory of Reusable Neural Bases for Network Compression
The paper introduces the Linear Reusable Neural Bases Architecture (LRNBA), a framework that represents each network block as a linear combination of shared neural bases to improve parameter efficiency and reduce memory cost. Experiments show comparable or faster convergence and lower loss than classical architectures while maintaining stable training dynamics.
NEW · Updated Sep 1, 2026 · Google DeepMindIntroducing agentic video understanding with Gemini
Google DeepMind published a blog post titled 'Introducing agentic video understanding with Gemini' on September 1, 2026.
NEW · Updated Sep 1, 2026 · Fraunhofer Versatile Video EncoderCan LLMs Design Video Coding Tools? A Case Study on Planar Mode
A paper on arXiv (cs.AI) explores whether large language models can design video coding tools, using a case study on Planar mode. Experiments in a generation-and-evaluation loop show an LLM-generated mode outperforms the conventional Planar mode in Fraunhofer Versatile Video Encoder (VVenC) faster preset, achieving 0.18% bitrate savings with 0.4% complexity overhead. The study also extends to Enhanced Compression Model (ECM) with directional Planar modes.
NEW · Updated Sep 1, 2026 · OpenAIHow AI-native companies turn workflows into operating capability
Basis, Clay, and Exa Labs use AI agents to improve onboarding, account management, and developer integrations.
NEW · Updated Sep 1, 2026 · EvoSCMEvoSCM: Scientific Belief Revision Through Causal Model Evolution and Experimentation
EvoSCM equips scientific agents with explicit structural causal models that evolve as new experimental evidence is collected. It maintains a population of competing SCM hypotheses, each encoding a candidate causal explanation of the environment, and evolves them through a closed discovery loop. In each round, the agent abduces latent mechanisms from accumulated evidence, designs discriminative interventions, and commits to falsifiable predictions that it tests through experimentation. Discrepancies between prediction and observation are inductively distilled into correction rules that revise the causal structures and mechanisms of each hypothesis, and the agent then deductively validates the revised population against accumulated evidence and structural consistency to guide the next round. EvoSCM is evaluated on DiscoverPhysics, a benchmark requiring agents to uncover the hidden dynamics of noncanonical physical worlds through experimentation.
NEW · Updated Sep 1, 2026 · ClickGraphRelational-Core Graph Analytics Querying graphs at SQL scale, and why the node/edge model is a performance tax, not a truer picture of connected data
A paper argues that a columnar relational engine fronted by a graph query language matches or exceeds native graph engines on analytical graph queries and scales past the point where in-memory graph engines fail. It presents ClickGraph and DeltaGraph, systems that translate Cypher directly onto the native relational schema and execute in place on ClickHouse, Databricks, or in-process on lakehouse files, with no import and no separate cluster.
NEW · Updated Sep 1, 2026 · arXivWhen Guardrails Look Effective: Construct Validity Failures in LLM Agent Commerce Evaluation
An arXiv paper (2609.01519v1) reports that interactive simulations using language-model agents in a buyer-seller hotel transaction testbed initially showed welfare gains from two marketplace guardrails of +87.4, +35.0, and +28.8 across a Qwen2.5 1.5B–14B ladder. After holding the offer schema and buyer chooser fixed, paired contrasts changed to +7.2, -13.9, and +23.8. The four largest 14B single-generation effects averaged +229, but after three generations per profile-condition they averaged +37.6 (95% bootstrap interval [-34.2, 109.3]), with generation residuals accounting for 49.9% of variation. A seller-incentive check was non-monotone: increasing profit pressure produced less profit than the default seller prompt. Scripted positive controls showed a profit-maximizing seller already attains first-best welfare, so guardrails mostly redistribute and reduce welfare.
NEW · Updated Sep 1, 2026 · TempClozeTempCloze: Can Video-LLMs Identify the Missing Middle?
TempCloze is a video cloze benchmark for evaluating visual temporal reasoning in Video-LLMs. It contains 1,521 filtered videos from seven sources, mainly long-take and egocentric videos. Models must identify the true missing middle clip from four candidates given beginning and ending clips. Distractors are constructed along three dimensions: Semantic, Alignment, and Progression. Evaluation of 10 proprietary and 21 open-source Video-LLMs reveals Alignment as the primary bottleneck.
NEW · Updated Sep 1, 2026 · LatentPressLatentPress: Context Compression Beyond Text and Vision
LatentPress introduces continuous memory tokens for context compression, achieving 4-16x compression with a small adapter (4.2M-26.2M parameters, ~0.1% of decoder). On LongMemEval, it reaches 0.504 accuracy at 7.70x compression versus 0.490 for uncompressed evidence, outperforming text summaries (0.184) and OCR-based compression (0.426 to 0.312). On LongBench-QA, in-domain writers match or exceed raw-context reading at 4-8x compression, while 16x trails raw. Writing takes 43ms per conversation, roughly an order of magnitude faster than text summarization or OCR reconstruction, and reading is 5-9x faster than raw context or cached OCR. Transfer settings include zero-shot from UltraChat to LongMemEval memory QA and from LongMemEval-derived data.
NEW · Updated Sep 1, 2026 · arXivOptimizing Byzantine Node Placement in Decentralized Federated Learning
A research paper introduces Byzantine Placement Influence (BPI), a set-level measure derived from gossip dynamics to quantify cumulative exposure of honest nodes to Byzantine sources. It formulates Byzantine placement as an adversarial decision under a fixed compromise budget and develops efficient algorithms for optimizing BPI, evaluated across six heterogeneous graph families with untargeted model poisoning.
NEW · Updated Sep 1, 2026 · arXivRethinking Learnability in Offline Data-driven Optimization
A paper proposes algorithm-dependent learnability, which requires accuracy only on the optimizer's trajectory, and proves its value-query form is sufficient for greedy and local search in submodular maximization, while its first-order analogue is sufficient for projected gradient descent on convex minimization.
NEW · Updated Sep 1, 2026 · GlossoGenGlossoGen: Emergent Language in Complex Multi-Agent LLM Interactions
Researchers introduced GlossoGen, a platform for studying multi-agent language evolution in complex scenarios. Within GlossoGen, they built the SaveVeyru scenario, which requires agents with partial information to communicate under pressure. They found that language evolution occurs between LLM agents, resulting languages are compositional and morphologically productive, and they deviate from the LLMs' English prior in ways that render them incomprehensible to humans. They identified several qualities essential to this evolution: pressure towards efficiency, strength of the models backing the agents, and access to a 'postmortem' stage in which agents can agree on linguistic conventions. They observed that agents learn new languages from usage alone and take an active role in this learning.
NEW · Updated Sep 1, 2026 · SkillSonarDefense-as-Skill: Evolving Runtime Guard Skill for Skill-Augmented Agents
Skill-augmented agents load reusable skills as persistent runtime context, improving task performance but also giving malicious skills a durable channel for steering future actions. Such skills may leak secrets, corrupt code, bypass approvals, or stage data for exfiltration only after a concrete user task and workspace state make the unsafe action appear useful. This makes pre-install vetting insufficient and calls for runtime, task-conditioned protection. The paper proposes Defense-as-Skill, a defense paradigm that implements the runtime guard itself as an installable, inspectable, and editable skill. The guard, SkillSonar, runs alongside untrusted task skills and checks sensitive actions against the user's task boundary, routing each action to an allow, replan, or confirmation decision without modifying the underlying agent runtime. The study constructs SCOPE-R, a task-conditioned dataset covering 6 risk families and 21 sub-categories, with 206 attack-confirmed malicious instances and 43 benign tasks. SkillSonar is improved on the SCOPE-R training subset using runtime guard-skill evolution, a Monte-Carlo Tree Search procedure that evolves the on-disk guard skill.
NEW · Updated Sep 1, 2026 · Harness-of-HarnessHarness-of-Harness: Multi-Day Autonomous Software Development with Continual Improvement
A paper introduces Harness-of-Harness (HoH), a framework that enables LLM-based coding agents to continually improve software during autonomous development. HoH organizes existing coding-agent harness executions into iterative planning-coding-testing loops, balances repair with capability growth, scopes development into small verifiable increments, separates implementation-time testing from independent evaluation, and constrains verifiable outputs. On GameCraft-Bench, FrontierSWE, and ProgramBench, three harness-model pairs (Codex with GPT-5.5, OpenCode with DeepSeek-V4-Pro, and Pi with MiniMax-M3), HoH consistently outperforms standalone harnesses, achieving an average relative gain of 52.25 percent.
NEW · Updated Sep 1, 2026 · arXivParsing the Stream: A Live Trace Model for Long-Horizon Agents and Their Observers
A live trace model is presented as an append-only event ledger folded incrementally into typed run state and compiled into per-consumer views. For the observer side, evaluated with an LLM reader as proxy, the compiled view answers monitoring questions using approximately 14x and 15x fewer input tokens (by reader) and at 5-7x lower cost than a budget-capped single-call reading of the raw trace, with higher accuracy (0.85-0.87 versus 0.48). For the agent, on 120-link sequential-dependency tasks, mechanisms that maintain the task's running statistic in per-step state succeed where full-context prompting fails (30/30 versus 8/30 under a clean protocol, n=30).
NEW · Updated Sep 1, 2026 · arXivWhen Safety Routing Breaks: Understanding Alignment Fragility under Benign Fine-Tuning
Benign fine-tuning severely weakens the safety alignment of large language models (LLMs). The paper proposes a Fisher-geometric explanation: safety Fisher is low-rank, and alignment makes the safety geometry flatter while preserving an output-routing pathway. After 100 benign fine-tuning examples, this pathway is selectively re-sharpened in output-side MLP modules, explaining asymmetric fragility: safety can collapse to high attack success rates, while general utility degrades mildly. The routing view explains why few safety examples can restore refusal behavior, indicating internal safety-relevant representations are preserved. LoRA and ASAM mitigate early collapse by suppressing output-side sharpness, but their protection weakens at larger fine-tuning scales.
NEW · Updated Sep 1, 2026 · Power-Law Entropy SearchEfficiently Estimating Optimal Hyperparameter Scaling Laws through Power-Law Entropy Search
Power-Law Entropy Search (PLES) is introduced as a computational cost-aware acquisition function built on multi-fidelity Bayesian optimization to efficiently estimate optimal hyperparameter scaling laws for large language model training. PLES selects candidate configurations that maximally reduce uncertainty of scaling law estimates per unit computational cost, favoring informative small-scale experiments. The method was evaluated on synthetic benchmarks, surrogate models fitted to real LLM training data, and actual LLM pre-training runs.
NEW · Updated Sep 1, 2026 · TREVISLearning Sparse Decision Trees via Transformer Variational Auto-Encoders
TREVIS, an approach for learning decision trees with complex objectives, is proposed. It uses a Tree Transformer Variational Auto-Encoder (TTVAE) to map decision trees to latent representations, enabling gradient-based optimization via a differentiable surrogate model. Experiments show TREVIS discovers decision trees matching predictive performance of existing near-optimal algorithms while improving structural sparsity.
NEW · Updated Sep 1, 2026 · arXivSemantic-Guided Multimodal Preprocessing for Vision Transformer-Based Clear Cell Renal Cell Carcinoma Grading
A study proposes a semantic-guided multimodal preprocessing method that integrates nuclei classification maps with RGB histopathology images for Vision Transformer-based clear cell renal cell carcinoma grading. The method achieves 0.916 balanced accuracy, outperforming RGB-only baseline (0.707) and max-voting aggregation (0.427). Sensitivity analysis shows the 21 percentage point improvement persists under simulated perturbation matching current nuclei classification error thresholds.
NEW · Updated Sep 1, 2026 · arXivProvably Safe Sim-to-Real Transfer
A paper on arXiv proposes a reward-free safe reinforcement learning algorithm for sim-to-real transfer that provably reduces real-world interaction while ensuring safe exploration and enabling near-optimal feasible policies.
NEW · Updated Sep 1, 2026 · EdiTikZEdiTikZ: Scientific Figure Editing from Revision Trajectories
A paper titled 'EdiTikZ: Scientific Figure Editing from Revision Trajectories' was published on arXiv on 2026-09-01. It introduces DaEdiTikZ, a dataset of 391K TikZ edit pairs mined from arXiv, GitHub, and TeX SE, with 781K directed edit instructions inferred by a VLM. It also introduces DaEdiTikZ-Bench, a human-refined benchmark with 790 instances, and trains two Qwen3.5-based EdiTikZ models (4B and 9B) using reconstruction and editing learning followed by reinforcement learning with rewards for rendered fidelity and edit application.
NEW · Updated Sep 1, 2026 · NeuSOGANeuro-Symbolic Geometric Abstraction (NeuSOGA): From Observations to Symbolic Mathematical Representations
A paper titled 'Neuro-Symbolic Geometric Abstraction (NeuSOGA): From Observations to Symbolic Mathematical Representations' was published on arXiv on 2026-09-01. It proposes a framework that transforms observations into topological abstractions, geometric abstractions, and symbolic mathematical representations, combining topology-guided structural discovery, foundation-model perception, adaptive multi-scale geometric abstraction, and symbolic synthesis.
NEW · Updated Sep 1, 2026 · DroneCATS-AgentEvaluating Multimodal LLMs as Generalist Vision-Language-Action Agents for Drone Control: Commanding, Approaching, Tracking and Searching
A research paper introduces DroneCATS-Agent, an architecture where a multimodal large language model (MLLM) is a swappable component in a drone control loop, and DroneCATS, a benchmark treating the model as the independent variable. The agent entrusts the model to yaw, search, deliberate when unsure, and self-declare arrival without fine-tuning or function-calling schemas. Evaluation across four capabilities—approaching a visible target, tracking a moving one, searching outside the initial view, and commanding a multi-drone fleet—reveals that even the simplest embodied settings are far from solved. The roster scales down to 2B parameters, and findings indicate that small open models often navigate into the success radius, suggesting flying is not the primary failure point.
NEW · Updated Sep 1, 2026 · arXivMeasuring consistency via ensemble margin and local prediction variability: Auditing decision systems in the presence of predictive multiplicity
A paper on arXiv (cs.AI) proposes a consistency criterion for auditing ensemble predictions under predictive multiplicity. The criterion combines ensemble margin with local prediction variability of constituent models. The authors show that consistency scores of finite ensembles converge to the expected model's score as ensemble size and sample count increase. The framework is evaluated on transformer models for natural language understanding and parameter-efficient fine-tuning of large language models for tabular data classification.
NEW · Updated Sep 1, 2026 · EDGEEDGE: Error Dependency Graph-Guided Multi-Error Attribution in Multi-Agent LLM Systems
EDGE constructs an error dependency graph from observed error events and validates a reliable causal subset through counterfactual rollout. The inference graph guides a two-stage LLM-as-judge detector for error attribution. Experiments on TRAIL and MAST show that EDGE improves category-level multi-error attribution across most evaluated models and settings. Experiments with adapted Who&When-style prompts show that the graph helps across prompting strategies.
NEW · Updated Sep 1, 2026 · PopPertPopPert: Population-level Joint-Distribution Modeling for Single-Cell Perturbation Prediction
PopPert is a framework that parameterizes population-level joint gene expression distributions for collective transcriptional state modeling. It predicts perturbation-induced changes in distribution parameters from a control population distribution and a perturbation condition, eliminating the need for cell-level correspondence. PopPert uses a low-rank Gaussian Copula to model cross-gene statistical dependencies and allows sampling of synthetic perturbed single-cell profiles.
NEW · Updated Sep 1, 2026 · SymFoldSymFold: Synergizing Evolutionary and Structural Priors for Accurate Protein Inverse Folding
A paper titled 'SymFold: Synergizing Evolutionary and Structural Priors for Accurate Protein Inverse Folding' was published on arXiv (cs.AI) on 2026-09-01. It introduces a symmetric dual-path architecture that combines protein language models (PLMs) for sequence evolution knowledge and multimodal protein language models (MPLMs) for structural knowledge to iteratively guide protein sequence generation. The method achieves state-of-the-art performance on standard protein inverse folding benchmarks, surpassing prior approaches.
NEW · Updated Sep 1, 2026 · CHARMCHARM: Character Hallucination for Multicultural Role Play Benchmark
CHARM is a multicultural benchmark of 40 real and fictional characters from five cultural-linguistic regions, validated by native reviewers. It probes Temporal and Cross-Universe boundary types using abstention-enabled multiple-choice questions. A two-stage evaluation separates Boundary-Awareness from Boundary-Compliance. Evaluations across six LLMs show hallucination is driven predominantly by compliance failures.
NEW · Updated Sep 1, 2026 · arXivScalable Rao-Blackwellized Online Planning for High-Dimensional POMDPs
A paper extends the Rao-Blackwellized online POMDP framework to high-dimensional settings using hybrid continuous-discrete belief representations. It integrates with FastSLAM 2.0 in a robotic search-and-rescue task and achieves higher cumulative rewards with fewer particles and simulations than purely sampling-based methods.
NEW · Updated Sep 1, 2026 · arXivCheap Verifiers, Large Blind Spots: Measuring the Reliability Cost of Cost-Saving Cascades
Inference cascades answer most queries with a cheap model and escalate a hard tail to a frontier verifier. The verifier's blind spot (fraction of wrong student answers accepted) grows with student capability (beta from 0.12 to 0.55 as student scales 0.5B to 32B) and shrinks with verifier capability. A frontier verifier drives beta to about 0.05 but escalates on 46% of hard-MATH queries against a 39% true error rate. Naive corrective fine-tuning on verifier-rejected tail degrades and ultimately collapses the small student across every teacher tried.
NEW · Updated Sep 1, 2026 · SIFTProbing Factual Knowledge Transfer with Training Data Interventions
A study proposes an intervention-based framework to test whether multilingual language models transfer factual knowledge across languages during continued pretraining. Starting from an English-pretrained model, researchers continued pretraining on Persian data with specific facts removed at varying granularity. They constructed SIFT, a resource of 500 triples across 20 relations, stratified by cultural origin of the fact's subject into general and Persian-related entities, with natively written Persian cloze templates. Results show fact transfer is very limited: under the strictest removal condition, a large majority of English-acquired facts fail to transfer into Persian. Sentence-level co-occurrence removal is insufficient to eliminate fact signal. Easier (randomly selected) negative candidate sets substantially inflate apparent transfer by rewarding shallow associative heuristics, while performance on a harder candidate set is lower.
NEW · Updated Sep 1, 2026 · LEAPLEAP: Likelihood Elicitation and Aggregation for LLM-based Probabilistic Forecasting
LEAP (Likelihood Elicitation and Aggregation for Probabilistic forecasting) is proposed as a method that reorganizes how collected evidence is used in the prediction stage of LLM-based forecasting systems. It examines each evidence item separately, elicits likelihood parameters, and combines them with an explicit prior and deterministic probabilistic model into a posterior distribution. LEAP supports continuous, single-choice, and multi-choice forecasts. A benchmark covering forecasting, information-seeking, and browsing tasks was built, and LEAP was evaluated on the authors' own agent loop and several agent CLI frameworks. Given the same evidence, LEAP improves most prediction and calibration metrics.
NEW · Updated Sep 1, 2026 · IMABOBandits in Prod: Hyperparameter Optimization at Inference Time
A paper titled 'Bandits in Prod: Hyperparameter Optimization at Inference Time' was published on arXiv on 2026-09-01. It formalizes Online Hyperparameter Optimization (OHPO) as an infinitely many-armed bandit over mixed and conditional search spaces, and introduces IMABO, a framework combining any bandit policy with any oracle for proposing new configurations. The paper also introduces IMOSS, a restart-free anytime policy with active set growing as t^β, and proves an expected cumulative quantile-regret bound of O(p_ρ^{-1/β} + T^{(1+β)/2}).
NEW · Updated Sep 1, 2026 · arXivAutomated Event Log Generation from Unstructured Text Using Finetuned LLMs
A research paper proposes a scalable framework using finetuned large language models (LLMs) to automatically generate structured event logs from unstructured text for process mining. The authors created a text-to-log dataset and finetuned LLMs, showing that finetuning outperforms few-shot and zero-shot prompting by a large margin.
NEW · Updated Sep 1, 2026 · MIDRMIDR: Enrichment-Augmented Indexing for Multimodal Document Retrieval
MIDR is a training-free framework for enrichment-augmented indexing that shifts multimodal reasoning to index time. During ingestion, a multimodal LLM converts rendered pages into verified textual fields indexed with BM25F and optionally fused with dense retrieval. On ViDoRe V3, MIDR Hybrid achieves 0.6219 average nDCG across five English domains, a 23.0% relative gain over BM25, remaining competitive with ColQwen2.5. On two French-document domains, enrichment bridges English queries and French page text, lifting BM25 from 0.1532 to 0.5448 nDCG and outperforming ColQwen2.5. Across all seven domains, MIDR leads ColQwen2.5 on four while using approximately 9x smaller index memory.
NEW · Updated Sep 1, 2026 · OmniEvaluatorA Composable Evaluation System for Reproducible Omni-Modal Foundation Model Evaluation
OmniEvaluator connects existing inference engines and curated evaluation libraries, exposing four inference backends, four evaluation frameworks, and over a thousand benchmarks through a single interface. Every run is recorded as an artifact capturing the full configuration for exact reproduction, and results flow into a shared dashboard for cross-model comparison. A federated mode shares GPU inference servers across concurrent evaluations, and a built-in verifier, small enough to run on CPU, keeps its score stable across engines and prompts where rule-based scoring fluctuates under configuration mismatch, matching cost-efficient commercial LLM judges without their recurring API cost.
NEW · Updated Sep 1, 2026 · OpenAIPath to Astra: critical capabilities and frontier safeguards
Astra is the first OpenAI model to meet the Critical cybersecurity capability threshold under the Preparedness Framework, with stronger safeguards for release.
NEW · Updated Sep 1, 2026 · StainPresetNetStainPresetNet: Stain Preset Network for Fast Multi-to-Multi Stain Normalization
StainPresetNet is a novel framework for stain normalization that combines structural preservation with dataset-level color mapping while maintaining computational efficiency. It implements pixel-wise normalization guided by preset reference images, enabling multi-directional adaptability without retraining. Evaluations on cytopathology and histopathology datasets demonstrate superior color mapping accuracy compared to conventional methods.
NEW · Updated Sep 1, 2026 · Celeb Twins Test SetRevisiting Face Recognition for Monozygotic Twins: The Celeb Twins Test Set
The Celeb Twins Test Set (CTTS) contains web-scraped image pairs for 80 sets of celebrity twins. It is the only twins test set with meta-data for twins with distinguishing skin marks and possible mirror asymmetry. Current deep CNN matchers can achieve over 76% accuracy in classifying CTTS same-person / different-person image pairs. The paper discusses the feasibility of using generative AI tools such as Grok, ChatGPT and Gemini to create images of imagined monozygotic twins.
NEW · Updated Sep 1, 2026 · OpenAIHealthcare organizations can now connect EHR and additional industry data to ChatGPT
OpenAI announced that ChatGPT can now connect to trusted healthcare data, helping clinicians securely access patient context, medical research, and more.
NEW · Updated Sep 1, 2026 · DNC-IMMDNC-IMM: Early Lane-Change Intention Recognition via Neural Calibration Based on Driving Context Information
A paper proposes a Dual Neural-Calibrated Interacting Multiple Model (DNC-IMM) for early lane-change intention recognition. It encodes driving-context information with a neural network to calibrate transition probabilities and measurement likelihoods. Experiments on the highD dataset show reliable recognition before lane crossing, with strong performance at 2-3 second prediction horizons.
NEW · Updated Sep 1, 2026 · Latent Recurrent ThoughtsLatent Recurrent Thoughts: Recurrent Refinement of Proposed Latents for Reasoning with Frozen LLMs
A paper titled 'Latent Recurrent Thoughts: Recurrent Refinement of Proposed Latents for Reasoning with Frozen LLMs' was published on arXiv (cs.AI) on 2026-09-01. The paper proposes a method where a frozen LLM is used for modeling and decoding sequences, while a small auxiliary network supplies continuous latent thoughts as input. These latents are produced by recurrence: a tiny recurrent reasoner refines them over many steps. The method is instantiated as Latent Recurrent Thoughts (LRT), with a task-dedicated proposer supplying base latents, a recurrent reasoner refining them through bounded residual corrections, and the frozen LLM decoding the answer. The paper reports results on symbolic reasoning tasks (Countdown-4, Sudoku) with answer supervision but no reasoning traces, and on natural language tasks (the summary is truncated).
NEW · Updated Sep 1, 2026 · EDRACEDRAC: Benchmarking Arabic Dialect Reading Comprehension
EDRAC is introduced as the first large-scale benchmark for dialectal Arabic machine reading comprehension and generative QA, covering Egyptian, Moroccan, Emirati, Syrian, and Saudi Arabic. It contains 499 passages from naturally occurring spoken interactions and 4,977 QA pairs generated through a human-LLM collaborative pipeline. Benchmarking of Arabic-centric and multilingual LLMs reveals substantial gaps between semantic answer quality and dialectal fidelity.
NEW · Updated Sep 1, 2026 · QwenHints Help But Do They Teach? Evaluating Skills Transfer in Code Generation
A study on arXiv evaluates whether hints that turn failing generated programs into passing ones provide missing information or merely steer the model toward a solution it could already produce. Tests on HumanEval+ and MBPP+ with Qwen2.5-3B-Instruct and Phi-3.5-mini show that relevant hints rescue some failures, but many of those rescues are also achieved by unhinted sampling, suggesting hints may not teach new skills. Mechanistic tests find a stable activation direction shared by relevant and unrelated hints, with no detectable net accuracy gain from adding this direction.
NEW · Updated Sep 1, 2026 · StateSwapStateSwap: Probing Support-Elimination Hidden States in Multiple-Choice Questions
Large language models often answer the same multiple-choice question inconsistently when posed under support-oriented and elimination-oriented framings. A dual-framing protocol with minimally varied prompts was introduced. An untrained special token, [STATE], was appended and its residual-stream activation used as an intervention interface. Across both models, the two framings induce separable [STATE] activations concentrated in intermediate layers. Swapping these activations between paired prompts systematically changes predictions and improves cross-framing agreement. Mean-difference steering directions derived from the dual-framing contrast exhibit more bounded layer-wise responses than matched contrastive activation addition directions under the evaluated protocol.
NEW · Updated Sep 1, 2026 · TFMatText-guided flow matching enables sample-efficient crystal structure generation
TFMat, a text-conditioned flow-matching framework, uses structured materials language as a semantic prior for a CrystalFlow generator. Across Perov-5, Carbon-24 and MP-20 crystal structure prediction benchmarks, TFMat improves one-candidate match rates over CrystalFlow and reaches a 92.04% MP-20 match rate with 20 candidates. In de novo generation, it improves element-count and density distribution alignment while retaining coarse property consistency in composition-selected outputs.
NEW · Updated Sep 1, 2026 · arXivSpace Generative AI with Solar Energy Harvesting
A research paper proposes a framework for solar-powered space generative AI where a satellite receives a prompt, runs a diffusion-based image-generation model, and downlinks the compressed result within a strict time window. It identifies computation-communication trade-offs governed by shared harvested-energy budgets and develops a joint resource-optimization framework using predictable solar-energy harvesting dynamics.
NEW · Updated Sep 1, 2026 · ARISE-RLARISE-RL: Agentic Rubric-Grounded Iterative Self-Evolution with Reinforcement Learning
ARISE-RL is a full-cycle self-evolution framework that couples a task/rubric Generator and a reasoning Solver through rubric-mediated co-evolution. The Generator grounds tool-related rubric criteria in real tool observations and is rewarded for producing valid, intermediate-difficulty tasks aligned with the Solver's evolving capability boundary. The Solver learns from fine-grained rubric satisfaction signals through multi-step reasoning and tool use. The framework introduces Reward-Gated Self-Evolution Distillation (RG-SED), which selectively distills a memory-augmented variant of the same policy back into itself only when the memory yields empirical reward improvement.
NEW · Updated Sep 1, 2026 · CM-PTMUser Representation via Cross Multi-source Behavior Pre-training for Mobile Games
A research paper titled 'User Representation via Cross Multi-source Behavior Pre-training for Mobile Games' was published on arXiv (cs.AI) on 2026-09-01. The paper proposes CM-PTM, a Cross Multi-source Behavior Pre-Training Model for mobile game user representation learning on device-level behavioral logs. CM-PTM uses hierarchical cascaded mask-then-predict proxy tasks to infer the source of the next behavior and refine predictions at the app-action level. Experiments on large-scale real-world mobile datasets demonstrate that CM-PTM effectively captures users' endogenous interests.
NEW · Updated Sep 1, 2026 · WorldBenchWorldBench: Culturally Grounded Benchmark for Multilingual Agents
WorldBench is a multilingual benchmark of persona-grounded everyday workflows with 1,600 tasks across seven languages and eight cultures. It introduces Constrained Task Success (CTS) combining natural language instructions and testbeds. Frontier models reach only 49.2% CTS, showing gaps between correctness and environment preservation.
NEW · Updated Sep 1, 2026 · PythiaLagged Coupling: Internal Representations Become Readable Before They Become Causal
Across the full Pythia suite (160M-12B, eight checkpoints, four task families), a linear probe can read a target variable from the residual stream as early as step 1,000 at every scale, yet steering along that same reading direction remains null-equivalent in 43 of 48 model-checkpoint cells. Internal readability systematically outruns causal efficacy, and the lag does not shrink with scale.
NEW · Updated Sep 1, 2026 · HiveTraceGuard-ProHiveTraceGuard-Pro: A Compact Generative Guardrail for Prompt Injection, Jailbreaks, and Adversarial Obfuscation
HiveTraceGuard-Pro is a 0.6B generative guardrail LoRA-tuned from Qwen3-0.6B, trained on Russian and English, using a binary safe/unsafe scoring rule for the final target turn. Its training corpus pairs harmful examples with benign examples from the same domain and applies eight obfuscation transforms to both labels. In a comparison with thirty-four other guards on nineteen benchmark groups (sixteen public), its aggregate key is 0.7432, behind 0.7641 and 0.7552 for the two higher-scoring guards. Over the sixteen public groups alone, its key is 0.7153 and four of the thirty-four other suite guards score higher. In a fifteen-model comparison, HiveTraceGuard-Pro has the highest clean Russian robustness combined-F1 (0.88) and Russian prompt-injection recall (0.999).
NEW · Updated Sep 1, 2026 · AgentFactoryAgentFactory: Towards Automated Agentic System Design and Optimization
A paper titled 'AgentFactory: Towards Automated Agentic System Design and Optimization' was published on arXiv (cs.AI) on 2026-09-01. The paper presents AgentFactory, a framework that jointly optimizes foundation models and workflow structures in agentic systems, considering multiple objectives including performance, cost, and efficiency. It uses LLMs as optimizers and a three-stage optimization pipeline.
NEW · Updated Sep 1, 2026 · CoBRACoBRA: Learning Tool-Use Boundaries via Counterfactual Margins
CoBRA is a counterfactual boundary-learning framework for tool-augmented language models. It constructs internal and external experts from the same base model, collects paired trajectories, and estimates the reward margin between answering with and without tools. The margin partitions data into internal-favored, external-favored, and ambiguous cases. CoBRA uses clear-margin samples for Boundary-Aware Cold-Start SFT, followed by MARS-RL with reference-split rollouts and counterfactual marginal advantages to optimize boundary decisions. Experiments with retrieval as the main tool on Qwen3-4B are reported.
NEW · Updated Sep 1, 2026 · arXivFew-Shot Out of Domain Intent Detection with Covariance Corrected Mahalanobis Distance
A paper proposes a covariance corrected Mahalanobis distance for detecting out-of-domain intents in few-shot settings, addressing limitations of prior work by Podolskiy et al. (2021).
NEW · Updated Sep 1, 2026 · arXivCalibration is the Bottleneck: An Action-Class Diagnostic of Multi-Turn Tool-Calling
A research paper proposes an action-class-oriented diagnostic framework for multi-turn tool-calling in LLM agents. It decomposes failures into action-class miscalibration and action-execution failure, using a four-class action space (TOOL_CALL/ASK/REFUSE/CONFIRM) and a self-revealing upper bound Acc <= GAR (Gold Action Recall). The framework is validated on a panel of tool-calling models across multiple multi-turn benchmarks. The diagnostic reveals action-class miscalibration as a substantial failure mode that state graders cannot see, inflating standing for heavily tool-trained families.
NEW · Updated Sep 1, 2026 · SciGramFrom Terminology to Diagrams: Visual-Instruction Generation for Scientific Diagram Understanding
A framework generates diagram-grounded instruction data using terminology from scientific curricula. It extracts domain concepts, synthesizes atomic facts, retrieves diagrams from the web, and generates captions and multiple-choice questions. The resulting SciGram dataset contains over 194K diagrams and 1.4M visual instructions across life, earth, and physical sciences. Models fine-tuned on SciGram achieve substantial improvements on diagram-centric benchmarks including TQA, ScienceQA, and AI2D, outperforming or matching state-of-the-art VLMs while using fewer training instances. Augmenting LLaVA OneVision with SciGram establishes new state-of-the-art performance on diagram understanding.
NEW · Updated Sep 1, 2026 · arXivEmbedded Conditional Independence Tests for Large Language Model Generated Text with an Application to German Parliament Speeches
A paper proposes embedded conditional independence tests (eCITs) for testing whether LLM output carries information about an attribute beyond the source text. The method embeds text and applies existing CITs to representations. The paper includes a semi-synthetic simulation design to assess type I error control and power, and applies the method to German parliament speeches.
NEW · Updated Sep 1, 2026 · DualStakeDualStake: Dual-Path Confidence Calibration in Deep Research Agents
DualStake is a dual-path calibration method for Deep Research agents that applies margin-clipped, confidence-dependent stake rewards to jointly align Evidence Confidence (E-Conf) and Answer Confidence (A-Conf) with answer correctness. Experiments on Qwen2.5-7B, Qwen2.5-7B-Instruct, and Qwen3-4B across 8 QA benchmarks show improved calibration without sacrificing answer accuracy. Code is available at https://github.com/FloXXXt/DualStake.
NEW · Updated Sep 1, 2026 · arXivContext-Grounding Gains Are Mediated by Pre-existing Machinery: Auditing GRPO, SFT, and DPO
A study compared nine post-training arms (GRPO, SFT, DPO) from one starting checkpoint, with key comparisons extended across scales and families. Across five tested GRPO variants, grounding gains are small. Conflict-SFT improves grounding moderately, while DPO drives grounding near ceiling on its matched distribution. Conflict-SFT and DPO largely use the same causal attention-head set as the starting model. Subtracting the starting-model direction suppresses both gains, while adding it to the starting model recovers 35% of DPO's gain.
NEW · Updated Sep 1, 2026 · PLANETBeyond the Image Plane: World-Grounded Queries for Multi-Object Tracking
PLANET is an end-to-end multi-object tracker that lifts existing 2D tracking datasets into 3D, forms world-grounded queries by embedding reconstructed 3D scene geometry into features and positional encodings, uses an auxiliary 3D location prediction task, and employs a dual-resolution temporal memory. It achieves state-of-the-art performance across three diverse benchmarks.
NEW · Updated Sep 1, 2026 · Hugging FaceIntroducing @huggingface/kernels: 200+ WebGPU Kernels for Local AI
Hugging Face introduced @huggingface/kernels, a library of over 200 WebGPU kernels for local AI, as announced in a blog post on September 1, 2026.
SUN: Persistent Programs For Language-Grounded Control-to-Learning-to-Real Policies
Kuafu, driven by large vision language systems, automatically synthesizes Semantically UNified (SUN) Programs from language and scene semantics, screens feasibility via MPC, and retains semantics while training stage-conditioned policies. Across nine tasks, Kuafu achieves 82.03% macro-success, outperforming sparse-reward (35.67%) and Stage-BC (24.75%) baselines. At 8192-way scale, it generates 10.57x the successful trajectory time per hour of human teleoperation. With 500 trajectories per task, Kuafu data trains DP3 policies to 46.0% simulation success (vs. 22.4% for alternatives) and 34.7% on physical Franka.
NEW · Updated Aug 31, 2026 · GLM-5.3Auditing Anonymous AI Models: A Four-Stage Protocol for Black-Box Identity Verification
A 2026 arXiv paper proposes a four-stage forensic audit protocol for black-box identity verification of API-served anonymous AI models. Stage 0 reconstructs launch-time configuration from archived platform snapshots; Stage 1 fingerprints configuration against the platform catalog; Stage 2 tests tokenizer identity with a cross-length differential; Stage 3 corroborates with behavioral probes. The protocol was tested on 10 known-identity releases, achieving 7 exact matches, 2 precision differences, 1 partial, and 0 counter-directional results. A prospective validation on a flagship case pointed to the GLM-5.3 version.
NEW · Updated Aug 31, 2026 · OntoAlignerOntoAligner-Ensemble: Voting-Based Fusion across Heterogeneous Ontology Alignment Techniques
OntoAligner-Ensemble is a modular, aligner-agnostic framework that combines candidate correspondences through a configurable two-stage process comprising voting-based fusion strategies followed by post-fusion selection policies. It supports any aligner implemented within OntoAligner that produces candidate correspondences. The framework was instantiated using representative lightweight string-aligner, KGE-based, and Retrieval-Augmented Generation aligners powered by both open-weight and API-based LLMs. Individual aligners and ensemble configurations were evaluated across eight benchmarks.
NEW · Updated Aug 31, 2026 · QwenWhen Does Bigger Help? A Controlled Study of LLM Scale for Ontology Learning
A controlled study evaluated 13 large language models from Qwen3.5, Qwen3.6, and proprietary GPT release variants on ontology learning tasks using the OntoLearner retrieval-augmented generation pipeline. All models used the same embedding model, retrieval configuration, prompt templates, decoding settings, datasets, and metrics across term typing, taxonomy discovery, and non-taxonomic relationship extraction on four biomedical and materials science and engineering ontologies. Within the dense Qwen3.5 lineage, increasing parameter count primarily improved precision rather than recall, with the largest gains between 9B and 27B parameters. Dense 27B models outperformed larger sparse models on term typing, while larger Mixture-of-Experts models achieved the strongest open-weight results on taxonomy discovery. Non-taxonomic relationship extraction remained difficult across model scales.
NEW · Updated Aug 31, 2026 · BLOOM-WILTBLOOM-WILT: Logit Tilting for Behaviour Elicitation in Automated LLM Auditing
BLOOM-WILT is a full auditing pipeline that elicits natural multi-turn instances of rare behaviours from deployed language models without training cost or access beyond the target's next-token distribution. It uses an auditor model that revises its conversational strategy across rounds and adaptively reweights the target's decoding using the model's own distribution conditioned on an elicitation prompt. Evaluated across 4 target models and 8 behaviours, it beats the baseline auditor in 30 of 32 settings and overturns previous model safety rankings.
NEW · Updated Aug 31, 2026 · arXivLLM Post-Training as Brownfield Maintenance: An Industrial Perspective on Dataware Engineering
Industrial post-training is described as a brownfield regime where teams inherit a deployed checkpoint and must land targeted improvements under fixed compute and mixture budgets without regressing the rest. The maintained artifact is increasingly dataware: behavior governed by a curated post-training mixture, updated via bounded mixture patches rather than clean-slate retraining. In a case study, interventions that raised the conversion of teacher distillation into usable training data increased accepted supervision by 2.84 times while using the same solution teacher and four solution attempts per candidate problem. In the primary evaluation, the yield-engineered patch improved CodeForces pass@1 by +2.59 points (+3.11 pass@3) and held-out LiveCodeBench v6 pass@1 by +6.11 (+8.05 pass@3), all statistically significant.
NEW · Updated Aug 31, 2026 · arXivCross-Regional Grapevine Cold Hardiness Prediction via Learned Multimodal Latent Representations
A research paper proposes a cold hardiness prediction framework that learns transferable latent representations using region-specific embeddings. It supports zero-shot and few-shot transfer to unseen regions by inferring embeddings from text descriptions and limited historical observations. Experiments on datasets from six North American regions show the approach outperforms state-of-the-art cold hardiness prediction methods.
NEW · Updated Aug 31, 2026 · arXivToken-Efficient Data Reasoning Agents via Adaptive Structuring of Unstructured Data
A research paper proposes agentic data cracking, a method that structures unstructured data adaptively and speculatively as a byproduct of reasoning. On the FanOutQA benchmark, reasoning over an ideal pre-structured store is 28X cheaper than over unstructured data, with the gap growing as questions fan out over more documents.
NEW · Updated Aug 31, 2026 · TASPOReconciling Process Supervision with Outcome-Based Credit in Agentic Policy Optimization
A paper titled 'Reconciling Process Supervision with Outcome-Based Credit in Agentic Policy Optimization' was published on arXiv on 2026-08-31. It introduces TASPO, a method that converts privileged supervision into outcome-grounded action credit for language-model agents.
NEW · Updated Aug 31, 2026 · AutoSciRubLearning to Evaluate Before Improving: Automatic Rubric Induction for Automatic Research Agents
AutoSciRub is an evaluation-first framework that induces a task-specific executable rubric before research execution, decomposing underspecified instructions into atomic scientific goals, grounding them in literature and data, and synthesizing verifiable criteria to guide execution, verification, and iterative revision.
NEW · Updated Aug 31, 2026 · arXivScaling Large Reasoning Models beyond Human Supervision: A Path toward Superintelligence
A paper published on arXiv on 2026-08-31 studies how large reasoning models can improve as human supervision recedes, examining reward and experience axes through a five-level ladder from L0 to L4.
NEW · Updated Aug 31, 2026 · Chulalongkorn UniversityReal-Time Video Anomaly Detection Using YOLO Pose Estimation and CLIP-Based Semantic Scoring
A lightweight two-stage framework for real-time video anomaly detection is proposed. The first stage uses YOLO v11n-pose to detect persons and extract seventeen skeletal keypoints in a single forward pass. The second stage encodes each cropped person region through CLIP ViT-B/32 and computes cosine similarity against predefined textual descriptions of anomalous behaviors. The architecture eliminates the need for optical flow, standalone pose estimators, and density-based scoring modules. Experiments on CUHK Avenue, ShanghaiTech Campus, and a custom indoor dataset collected at Chulalongkorn University demonstrate an end-to-end throughput of approximately 51 FPS on an NVIDIA Titan XP GPU, a 3.36x speedup over the multi-feature baseline, while maintaining frame-level AUROC values of 89.26%, 70.26%, and 84.13%, respectively.
NEW · Updated Aug 31, 2026 · Qwen3.5Wrong Prediction, Right Answer: Recovering Evidence from Collapsed LLM Sequence Scores
A study on arXiv (cs.AI) reports that hidden-state probes can decode correct answers even when native sequence scoring collapses in large language models. A diagnostic protocol using a minimal, target-label-free additive correction, fitting two parameters on as few as 25 unlabeled examples, recovers 9-34 accuracy points for Qwen3.5 models and transfers to OLMo-2-1B and Llama-3.1-8B. Recovered decisions persist on hard instances and exceed count-preserving permutation baselines.
NEW · Updated Aug 31, 2026 · arXivMeasure Before You Manage: Evaluating Agent Working Memory in Coding Agents
A study of 55 archived coding-agent trajectories finds that semantically different working-memory objects (instructions, artifacts, tool outputs, agent-generated state) exhibit distinct retention and compression behavior. Two semantically informed strategies—an object-aware compression policy and a retrieval-based policy—show that calibration gains may not transfer to held-out tasks, and equal token budgets do not imply equal delivered context or management cost. A real-system replay exposes serving limits not captured by nominal budgets.
NEW · Updated Aug 31, 2026 · MNIST-PROMNIST-PRO: MNIST is Back as a Partially Observable World for AI Agents
MNIST-PRO is a benchmark that isolates agentic perception by converting MNIST digit recognition into a sequential, glimpse-based search task with lookback constraints. Ten multimodal models were evaluated across four memory representations: raw visual history, textual states, structured metric grid maps, and a consolidated visual canvas. Models excel under full observability, but partial observability exposes a clear performance gap. Three bottlenecks were identified: perceptual-state construction and interpretation, premature stopping of exploration, and failure to revise early incorrect beliefs despite contradictory evidence.
NEW · Updated Aug 31, 2026 · arXivOne note in three: a verified census of three deployed AI scribes, and the instrument that counted it
An audit of three commercial ambient AI scribes on 142 consultations (565 notes) found 31.3% of notes contained verified failures, concentrated in allergy and medication information, invented patient identity, and history written as examination on telephone consultations. Excluding errors a patient record would have prefilled, the rate was 24.8%. Two clinicians adjudicated blind samples, upholding 20 of 21 and 12 of 12 findings.
NEW · Updated Aug 31, 2026 · arXivLLM Judges Verify Presence, Not Absence: Omission Blindness in AI Clinical Notes and What Recovers It
A benchmark of 500 single-error note pairs (298 with a named fact certainly absent, 202 added-or-altered controls) found that across eight LLM judge designs, paired discrimination on omissions was 0.50-0.63 (near coin flip), while on added or altered content it was 0.79-0.94. On single notes, no design flagged omissions reliably more often than perfect notes. Restructuring the task to list transcript facts then check the note recovered detection via a per-fact pipeline and a GEPA-evolved prompt.
NEW · Updated Aug 31, 2026 · QwenStick to What You Know: A Study of Knowledge-Aligned Supervised Fine-Tuning
A research paper proposes knowledge-aligned supervised fine-tuning (SFT) to reduce hallucinations by constraining training targets to the base model's parametric knowledge. It compares existing generation-based and estimation-based methods and introduces Evidence Rewrite and Recall Rewrite. Experiments with Qwen 3 4B and OLMo 3 7B show reduced factual hallucinations on WildHalu and Biography while largely preserving general capabilities. Recall Rewrite yields the strongest factuality gains and improves refusal behavior on UnknownBench.
NEW · Updated Aug 31, 2026 · arXivEvaluating and Improving LLM Self-Modeling
A study introduces a benchmark for self-modeling, defined as an LLM's ability to answer verifiable questions about its own behavior, such as whether a prompt edit would change its final answer. Current models show non-trivial but limited self-modeling skill and make systematic mistakes on simple counterfactual questions. A scalable synthetic-data pipeline and reinforcement learning improved aggregate self-modeling skill across three open-source model families with some transfer to held-out tasks, but gains did not consistently arise from privileged access to internal decision processes.
NEW · Updated Aug 31, 2026 · MR-JEPAMR-JEPA: A General Purpose Video Foundation Model for Cardiac MRI
MR-JEPA is a self-supervised video foundation model for cardiac MRI that extends LeJEPA to 3D spatiotemporal inputs using tubelet tokenization, spatiotemporal masking augmentation, and initialization from a 2D CMR foundation model. It is pretrained on multi-sequence data (cine, LGE, mapping) from 10,505 patients across two centers without annotations. Evaluated on six downstream tasks using a unified multi-view gated attention architecture, it outperforms compared methods on all five regression tasks, achieving LV EF MAE of 4.79% (r=0.764) and GLS MAE of 1.87 (r=0.805), with 21-27% MAE reductions over baselines on strain tasks.
NEW · Updated Aug 31, 2026 · CoJEPACoJEPA: Combining Contrastive Learning and JEPA for Global-Local Music Representations
CoJEPA combines contrastive learning and Joint-Embedding Predictive Architecture (JEPA) in a single shared backbone, jointly trained with a JEPA objective on masked sequence tokens and a contrastive objective on the class token. The contrastive gradient removes the need for an EMA teacher. No extra parameters are added to the backbone. CoJEPA outperforms or matches both individual methods across global and local music information retrieval (MIR) tasks, with a particularly strong advantage on tonal and harmonic tasks.
NEW · Updated Aug 31, 2026 · CogEvolCogEvol: Towards Efficient and Reliable Learning Environment Generation
CogEvol is a family of models trained for Learning Environment Generation, turning a course brief into structured-JSON slides or self-contained interactive HTML pages in a single pass. Across 220k production requests, CogEvol completes a slide in a median of 17 seconds and an interactive page in 59 seconds. A production-grounded data pipeline turns real failures into 53,687 verified SFT samples, and a hybrid rule-plus-VLM reward drives GRPO-based RL. CogEvol-27B scores 83.7 on slide quality and 63.7 on a 500-case interactive-HTML benchmark with 26.9x fewer parameters than flagship coding models. CogEvol-4B is released openly under the Apache 2.0 license at https://github.com/CogEvol/CogEvol-4B.
NEW · Updated Aug 31, 2026 · QwenA Universal Context-Reuse Layer for Cross-Model KV Sharing
A study on cross-model KV sharing translates KV states from a source model to a target model, including models differing in scale, architecture, attention configuration, tokenizer, and model family. For Qwen2.5-7B to Qwen2.5-1.5B, translated KV states improve LongBench2 accuracy from 27.59% to 34.48% over the native 1.5B baseline. For Qwen2.5-1.5B to Gemma-2-2B, KV handoff reduces target-side prefill cost by up to 67.05% at 4K context length while maintaining decoding perplexity close to native.
NEW · Updated Aug 31, 2026 · LOCILOCI: A Locator-Critic with Refinement Loop
LOCI is a training-free framework that decouples visual search from evidence verification using a Locator agent and a Critic agent in an iterative refinement loop. It improves accuracy on complex visual benchmarks for open-weight models like Qwen3-VL (+12.1 on V*, +5.8 on HR-Bench, +11.2 on VisualProbe-Hard) and proprietary models like Gemini 2.5 Pro (+8.9 on V*, +4.3 on HR-Bench, +4.8 on VisualProbe-Hard).