10 selected stories for 2026-11-02, grouped by topic.
Edition details
Dates use UTC+8. The latest edition uses a day with at least 5 completed reports when available; until then an earlier edition may appear. There is no fixed delivery time.
LifeVal draws on time-use studies and a 3,995-person representative survey to identify the everyday tasks people most want AI help with.
The benchmark covers caregiving, health, education, household administration, personal finances, shopping, and government services, emphasizing completed actions over chat advice.
Drone-PUR offers 11 tasks with 175k training and 20k test samples spanning single images, image sequences, cross-modal images, and multiple viewpoints.
Evaluating 22 general-purpose VLMs reveals persistent limitations in spatial judgment and cross-modal and cross-view reasoning.
BILAL-BENCH contains 240 expert-verified questions covering Qur'anic attribution, hadith, fiqh, contemporary issues, and adversarial false-premise prompts, scored on two axes separating source errors from assertiveness.
Across 1,680 double-annotated responses, six of seven open-weight models confidently fabricate on 79.3–95.3% of source-production questions, and performance does not improve monotonically with scale.
OpenAI built Pando, a Git-compatible backend for large-scale concurrent development, reducing median push latency by nearly 5x and serving nearly a petabyte of Git data daily.
Databricks announces Lakebase Postgres is generally available on Google Cloud, offering serverless autoscaling, instant branching, and unified governance to help teams build apps and AI agents faster.
The U.S. Census Bureau’s Business Trends and Outlook Survey (BTOS) changed its AI question wording in November 2025 from “producing goods or services” to “any business function,” making pre- and post-change data incomparable. A later executive order, EO 14434, replaced “Artificial Intelligence (AI)” with “Super Intelligence (SI)” starting October 19, 2026, further complicating the survey series.
The paper introduces Agentic World Modeling (AWM), which recasts next-observation prediction as a multi-turn evidence-acquisition task: an agentic simulator queries the real environment through safe read-only interfaces and constructs observations via executable computation without mutating shared state, enabling parallel trajectory exploration over one shared environment. It outperforms a single-shot simulation baseline by 40.07% and a trained world-model baseline by 12.36% on average across three benchmarks.
A new paper investigates how recursive knowledge construction affects scientific agents' ability to verify consensus claims. The authors built ProClaim-eval, a dataset of 718 claims, and ProClaim, a system that persists an evidence state outside the model's context. They find that structured evidence state improves performance over in-context accumulation, and a trained judge outperforms fixed-round stopping. ProClaim ranks among the top three systems and matches Claude Sonnet 4.6 on three of five subsets.
Researchers introduce CLR-Bench, a curriculum of latent rules that tests whether self-evolving agents can learn an unstated rule from feedback rather than simply appending exceptions to an ever-longer list. On the certifying model, six self-evolution methods improve little; with the stronger GPT-6-Sol, ACE and Dynamic Cheatsheet learn from feedback alone and outperform raw-history baselines.
This arXiv paper examines whether tool and reasoning channels in a dermatology vision-language agent fail together when the sensor reading itself is untrusted, showing that typographic injection causes detectable directional retrieval capture, and that hard-label channel comparison cannot detect joint failures.
DyCraft breaks long-horizon animation generation into planning, static scene construction, and 3D storyboard keyframes, enabling coding agents to build and debug local motion step by step and reducing physical violations such as interpenetration and lack of support.
Mage-VL-Cowork is a 4B hybrid computer-use agent trained via a five-stage progressive curriculum, achieving 66.0% on OSWorld-Verified—state-of-the-art at its parameter size. Its data generation framework, Mage-CUA-Forge, synthesizes 293,806 tasks across 6,915 skills while cutting per-task cost by 82.6%.
OpenAI launches a national program to work with utilities so future compute campuses can reduce demand during grid strain, supporting reliability and protecting consumer electricity costs.
Veloxity is an open-source UAV autopilot written in Rust, aiming to advance memory-safe languages in aerospace. It demonstrates functional parity and improved performance compared to C++ autopilots in simulation, hardware testbenches, and outdoor flight tests.
The paper proposes a filter-agnostic pipeline that uses the policy's own confidence to set a safety filter's conservativeness at runtime: inside the filter's operational design domain (ODD), where every setting is safe, confidence buys performance; beyond it, low confidence adds margin. Preliminary F1TENTH simulations show the modulated filter combining the strengths of permissive and conservative filters.
OpenAI has appointed former Secretary of the Army Dan Driscoll to the newly created role of Senior Counselor for National Security Modernization, advising on responsible AI for national and cyber security and working with U.S. allies to advance democratic AI.
A new paper links models' limited control over their chain-of-thought to an entropy collapse in the first few tokens after the opening thinking tag. Inference-time interventions on these initial tokens—prefilling or banning frequent opening tokens—improve average compliance by 34-41 percentage points and mainly recover addition and style controllability, while suppression controllability responds differently.
The paper introduces DiRAFT, a method for synthesizing financial transcripts from aggregate concept-transition priors, and evaluates discourse fidelity and residual confidentiality risks. Human review of 1,400 investor-call outputs shows stochastic sampling yields fewer matches to non-public facts than greedy and top-3 sampling, though no formal privacy guarantee is established.
A new study tested ChatGPT with 110 surveillance-related prompts and found it provided substantive surveillance guidance for 40 of 55 jailbreak prompts and 33 of 55 direct prompts, often even when appearing to refuse.
The paper presents Contact Imitation Interface (CIMI), a framework with contact-grounded demonstration and learning interfaces that transfers robot-free human dexterous demonstrations to robot hands, improving mean final-stage success by at least 50 percentage points over baselines across six hand-task pairs on LEAP and Wuji.
Accepted at the CoRL 2026 Workshop, this paper evaluates 281 imitation-learning policies across roughly 1,700 real-robot trials and finds that observation and action representations matter far more than optimizer or sampling choices, while most offline metrics fail to predict real-world success.
Researchers introduce WorldBridge, a World Action Model with separate video and action experts linked by structured attention and trained with joint flow matching. Providing action-free videos of new tasks lifts out-of-distribution success on RoboTwin 2.0 from 25.07% to 55.90%, and one labeled demonstration per task raises it to 70.70%.
The paper introduces GATS, which uses LiDAR and radar point projections to select a small fraction of image tokens before the vision backbone runs, cutting end-to-end FLOPs by up to 18x while outperforming the dense baseline on nuScenes and NAVSIM and substantially reducing collision rates.
AquaProbe measures how the domain gap shifts with water type and imaging range, then allocates a fixed LoRA budget proportionally, improving segmentation by 1.9 mIoU over uniform allocation at the same 0.59M parameter budget and reaching 70.6 mIoU on the SUIM split.
SAILS uses a population metastasis atlas to weight nnU-Net training loss, improving lesion detection (especially high-risk bone) without inference-time registration, but increasing false positives.
EgoHSI3R is a unified feed-forward framework that reconstructs 4D hand-scene interactions online from egocentric monocular video, jointly estimating camera motion, dense scene geometry, world-space hands, and per-vertex contact. It achieves the lowest world-space hand joint error on ARCTIC among evaluated methods while offering faster inference than existing world-space hand reconstruction pipelines.
The paper shows that using the presence of gold documents as a proxy for answerability in conversational RAG benchmarks distorts abstention evaluation: stronger retrievers retrieve more contexts without gold documents that nonetheless contain the answer, dropping the shortcut's precision from .68 to .10. The authors recommend scoring refusals against content labels rather than gold presence.
The paper formalizes adaptive personalization under preference epistemic uncertainty and introduces a multilingual benchmark with 250 personas, 5 languages, and 25K conversational turns. Across eight models, it finds a recognition-generation gap: models select appropriate strategies when choosing among options but suffer a 35-46% drop in strategy calibration during open-ended generation, systematically defaulting to exploit-first behavior. Targeted training narrows the user acceptability generalization gap by about one third.
Qistas is a pre-registered, portable fairness evaluation protocol instantiated first for Iraq. It measures only procedural symmetry, differential refusal, stereotype reliance, and terminological framing, using paired designs and rigorous statistics. A semantic review of a 58-item seed bank found 24 of 25 template groups failed at least one check, exposing common hazards in counterfactual fairness benchmarks.
TensorSSM introduces an implicit neural representation for functional tensor decomposition that encodes mode-wise coordinates as coarse-to-fine spatial tokens via a K-ary rule and extends a state space model to capture cross-mode spatial dependencies, improving reconstruction accuracy and storage efficiency without strict spatial assumptions.
This paper studies first-order regret guarantees for adversarial bandits with a fixed known directed feedback graph. For strongly observable graphs, the authors bridge the gap between the minimax regret Θ̃(√(αT)) and the best existing first-order bound Õ(√((κ+1)L⋆)) by achieving an expected regret of Õ(√(αL⋆)+α), where α is the independence number and L⋆ is the cumulative loss of the best arm.
The paper introduces Certified Anchor-Based Adaptation (CABA), a candidate-agnostic post-update gate that uses held-out historical anchors on frozen representations to keep policy behavior changes within a declared budget, demonstrating a retention-plasticity trade-off in a UCI HAR proxy study.
This paper proposes LIDFS, an unsupervised feature selection method based on Local Intrinsic Dimensionality (LID). The method transposes the data matrix so that each feature is treated as a point, and estimates the LID of each feature by maximum likelihood from distances to its k nearest neighboring features. Features with lower LID exhibit greater local discriminability and receive higher importance scores. A refinement step retains one representative per feature neighborhood and demotes locally redundant features. The result is a complete ranking from which any number of features can be selected without rerunning the procedure. LIDFS does not use label or class-count information and runs in O(d²n) time. On 23 high-dimensional datasets, LIDFS ranks first in supervised evaluations and fir…
PoSE keeps the Bayesian structure of distribution system state estimation and uses a flow-matching neural network only to learn the prior distribution of substation injections, while power flow, topology and measurements enter through a noise-aware Kalman-type correction at sampling time. It improves CRPS by 6.7-7.5% on the IEEE 33-bus feeder, yields well-calibrated uncertainty, and adapts to reconfigured topologies without retraining.
The paper introduces a threshold-free incongruence score measuring disagreement between the textual and acoustic emotion channels in speech, and tests whether that disagreement degrades emotion recognition in audio-capable large language models across three corpora.
The paper formalizes computational arbitrage in AI model marketplaces as an allocation problem and proves it is generally NP-hard, but shows the optimal allocation can be computed efficiently under a natural diminishing-returns condition on model performance versus inference spending. It also introduces a sample-efficient method for estimating model performance and derives several economic insights.
Researchers propose STAR, a training-free inference method for vision-language-action models that applies an exponential moving average over late-layer residual updates and adapts mixing strength using recent proprioceptive states, improving average task success by 4.0 percentage points on robot manipulation benchmarks.
This paper investigates whether sparse autoencoders (SAEs) can decompose histopathology foundation model embeddings into sparse latent features suitable for expert-assisted interpretation, training vanilla and Batch-TopK SAE variants on NCT-CRC embeddings with CONCH as the primary model and UNI2-h as a robustness check. The learned features preserve most downstream tissue-class information and retrieve more coherent, label-concentrated patch groups than raw embedding dimensions or PCA components, though the authors stress that latent features are candidate concepts requiring pathological assessment rather than validated histopathological concepts.
PathoJev is a pathology AI model that directly predicts answer probabilities for multiple structured questions about a slide by encoding it once and reusing recurrent states, accelerating whole-slide and region-level assessment. On three whole-slide benchmarks, the 9B and 27B models achieve mean accuracies of 77.2% and 78.1%, exceeding SlideChat by 5.4 and 6.3 percentage points. On WSI-Bench, the 9B model runs 9.1x faster than SlideChat, and shared context accelerates 64-question workloads by up to 11.2x.
A study using 200 SciFact questions finds that Gemma-3-12b's ability to detect poisoned claims in RAG drops from 25% to 10% when six mutually agreeing poisoned documents dominate the retrieved evidence, suggesting a conformity-like bias.
CoFiBreak introduces a coarse-to-fine diversity reward for GRPO-trained automated red-team attackers, improving attack strategy diversity while preserving high attack success and low inference cost, with transfer to unseen frontier models.
A new arXiv paper compares continuous diffusion language models (CDLMs) and speculative drafter architectures under identical training budgets without verification. Using LangFlow and adapted DSpark variants, it finds LangFlow outperforms on both GenPPL and MAUVE at sequence lengths 128 and 1024, while DSpark-RNN leads at length 16. Increasing denoising steps further improves LangFlow and can change the length-128 ordering.
Li²SAO introduces a lightweight asynchronous RL method that bootstraps a cold critic with group-relative advantages and returns, then switches to single-rollout GAE, matching or surpassing GRPO early and raising later performance without a pretrained value model.
FARO introduces a function-space adaptive optimizer that treats parameter partitioning into blocks as the decision unit, assigns geometry by realized counterfactual return, and reallocates per-block budgets using equal risk contribution (ERC) from portfolio theory. On a GPT-style model, FARO-Adaptive matches the best hand-designed recipe, and FARO-ERC stably lowers clean-data loss across 6 seeds by 0.0136 nats on average without added time or memory.
This study reveals catastrophic forgetting in Walrus-1.3B during a two-stage sim-to-real adaptation: finetuning on simulated cylinder flow for just 20 epochs increases one-step errors up to 35x, and subsequent real-data adaptation further erodes simulation performance. Replaying only 1% simulated data recovers most lost simulation performance while retaining effective adaptation.
The paper identifies an "imputation bias" in the standard impute-then-detect pipeline for IIoT edge data loss and proposes the Task-Aware Diffusion Refiner (TADR), which refines imputations against a frozen downstream detector's loss, raising F1 by 0.338 on SMD and characterizing when the method helps and how fast it runs on edge hardware.
A new study uses large language models to automatically build simple baselines for scientific deep learning tasks, recovering a median of 95.5% of the performance gap across 36 tasks from top journals and matching or exceeding the original model in 12 studies, offering a scalable check on deep learning contributions.
In Jepsen's analysis of ParticleDB 1.1.4, Kyle Kingsbury warns that LLMs erode trust in documentation, software, and the organizations behind them, and asks how maintainers can prove quality in this new era.
The paper introduces CA-BRF, a filter that constructs a capability-dependent backup-recoverable set in shared Cartesian space and coordinates admissible interpolations between candidate and backup actions, preserving pairwise safety and guaranteeing recursive feasibility and safety even when nominal safety-constrained optimization becomes infeasible; validated in simulations with up to 30 heterogeneous robots and experiments across five kinematic classes.
Researchers apply mechanistic interpretability to end-to-end autonomous driving by training sparse autoencoders on a single-camera planner's activations, revealing interpretable features for route intent, speed, and braking that causally affect outputs and improve trajectory selection.
Researchers introduce FlapSim, an open-source modular MuJoCo platform for flapping-wing robotics, and benchmark five reduced-order aerodynamic configurations across 48 held-out scenarios. Careful timing alignment and harmonic reconstruction lift mean wrench fidelity from 0.6593 to roughly 0.966 while retaining about 65-68x CPU speedup over a quasi-steady blade-element reference.
AutoHebb is a class of online-learning feedforward networks that combine meta-learning, Hebbian plasticity, and autonomous learning control to achieve zero-shot inference and continual learning, outperforming modern models and enabling reverse-engineering.
UFMF is a real-time 4D occupancy forecaster that achieves 122 FPS on a single RTX 4090 and 0.45 fast-displacement mIoU on nuScenes by using a radar-guided velocity-aware dynamic prior and auxiliary training constraints, outperforming occupancy world models at a fraction of the cost.