10 selected stories for 2026-10-28, 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.
ServiceNow will bring OpenAI models and agents into workflows from development to security, and expand ChatGPT Work and Codex to employees.
A new ServiceNow integration will bring business context and workflow actions directly into ChatGPT, enabling teams to investigate issues and take next steps in the same conversation.
MedJev achieves 72.5% on MedQA, 75.4% on USMLE, and 80.1% on NEJM, approaching published GPT-4 performance with well-calibrated confidence (ECE 0.028–0.063).
On patient-level tasks, MedJev reaches AUROC 0.926 for in-hospital mortality, exceeding XGBoost (0.902), but underperforms supervised models on ED triage admission (0.745 vs. 0.809).
HIDRA employs a hierarchical dual-attention mechanism that separates instruction-level expert retrieval from context-dependent refinement to mitigate routing interference.
A key-level geometric regularization enforces within-task alignment and cross-task separation, stabilizing expert selection under representation drift.
Lipi-Ghor is the first Bengali speech corpus to combine large-scale coverage with speaker-diarized, long-form audio annotations.
The corpus comprises 885 hours of audio from 998 YouTube videos across 596 unique channels, spanning 17 domains such as audiobooks, drama, sermons, news, and regional dialects.
NVIDIA expands its RTX Spark laptop lineup with new models from major OEMs, introduces Nemotron 3.5 Transcribe and GPU resource controls for agents, and adds ray tracing and DLSS support for new games like Phantom Blade Zero and Minecraft via RTX Remix.
Simon Willison argues that October 2026 marks an inflection point for vibe-coded games. Claude Opus 5.5's improvements in 3D modeling, music generation, and game design mean simple prompts can now produce genuinely fun games. He shares four browser-based games built entirely with AI and discusses the gaming community's reaction.
SG^2SLAM couples streaming spatial geometry with a ray-conditioned 3DGS kernel, replacing per-frame 2D-3D correspondence with a chunk-level metric prior and jointly optimizing geometry and poses, outperforming per-frame baselines on KITTI and Waymo.
ASPEN introduces an event-triggered scheduler that uses a lightweight multimodal predictor to estimate whether a frozen navigation VLA policy would reissue its previous command, reusing commands when persistence is likely and invoking costly semantic reasoning only when needed. Evaluated with frozen NaVILA on RxR-CE val-unseen, it cuts VLA calls by 9.5% and estimated inference time by 8.1% while preserving identical baseline action sequences on 92.1% of episodes.
This paper introduces a framework for program-level adversarial co-adaptive self-play where coding agents generate attacks and defenses from tested opponent code and mission feedback without model fine-tuning. Two eight-generation case studies on autonomous driving and humanoid locomotion show defenders adopting direct-actuation recovery and revised state validation, respectively, but these responses only provide local task recovery, not sustained robustness.
E0.5 is a geometry-native world-action model that uses a 3D foundation model as its primary backbone, treating 3D geometry as the native representation for both world modeling and action prediction while explicitly predicting future 3D states. Without embodied pretraining it reaches 89.90% and 50.09% success on RoboTwin 2.0 clean and randomized settings; after pretraining on over 160k trajectories from AgiBotWorld-Beta, real-world manipulation success improves by roughly 15 and 7 percentage points on AgileX Piper and Zerith H1.
Any3R is a feedforward geometry foundation model that accepts images from any central camera without per-lens retraining. It replaces the pinhole assumption by conditioning the positional encoding and the training loss on the dense ray field, enabling robust reconstruction from mixed camera types and delivering the most consistently accurate results across thirteen central cameras on three datasets.
A study accepted at EMNLP 2026 main conference shows that vision-language models often comply with requests to transform faces into specific racial or sexual orientation identities, producing statistically distinguishable images that amplify stereotypes and propagate bias into textual reasoning and criminality profiling.
This paper argues that LLMs often answer ambiguous inputs as if they had one meaning, and existing input uncertainty methods may disambiguate before measuring ambiguity. It proposes a label-preserving perturbation framework with black-box rule-based and white-box importance-guided detectors, outperforming output-side uncertainty and LLM-paraphrase perturbation baselines across four datasets.
Researchers present LOTUS, a multilinear low-rank decomposition framework that extracts compact representations from LLM hidden states and trains a lightweight classifier to detect hallucinations, outperforming recent state-of-the-art baselines across multiple models and QA benchmarks.
A new study measures language models' number preferences without sampling error, revealing that instruction tuning makes distributions less uniform and installs round-number avoidance while sharpening endpoint avoidance. The heuristics partially align with human number-choice theories.
FinLLM-KG is a model-agnostic framework that combines causal LLMs with a financial knowledge graph to inject temporally grounded, company-level knowledge from SEC 10-K filings at inference time without updating model weights. It supplies context via natural-language company profiles or relevance-ranked factual triples, enabling sentiment predictions that account for broader firm-level information and supports explainability, with simulations showing improved portfolio performance over the state-of-the-art baseline.
A new arXiv paper demonstrates that Physics Informed Neural Networks (PINNs) can solve the semiconductor drift-diffusion equations with accuracy comparable to conventional TCAD software, handling non-equilibrium bias conditions and inverse problems for doping profile retrieval.
A registered study with 256 synthetic worlds and 90,084 answers from 11 LLM endpoints finds that lessons written from a chance failure of a better tool increase the rate at which readers choose the worse tool (11.6% vs. 0.4% with an irrelevant item), with the effect size depending on whether the log states that failures are independent.
A new study systematically measures dynamic evaluation across hybrid linear-attention, diffusion language models, and long-context benchmarks, finding that gradient updates adapt style and vocabulary rather than retrieval, and that writing facts into weights harms reasoning while distilling conclusions as behavior into low-rank adapters reaches in-context ceilings without GSM8K degradation.
Researchers present MARQ, a modular, LLM-agnostic prototype that boosts LLM-based requirement quality assessment by adding a criterion-specific knowledge stack operationalizing INCOSE criteria. On an expert-labeled NASA requirement set, median F1 rose to 0.80 for Claude Opus 5, 0.62 for GPT-5.6 Sol, and 0.53 for Gemma 3 12B, but gains are model-dependent and still require criterion-specific interventions and human-in-the-loop review.
A new arXiv paper introduces The Trinity, a recursive self-improvement framework that couples world generation with Harness evolution through Creator, Explorer, and Evolver roles, reorganizing tasks into a capability tree. Evolved Harnesses improve mean performance by 34 points across 28 training benchmarks and 26 points across 18 unseen benchmarks, including two real-world embodied tasks.
This article reviews the benchmarking industry, highlighting that while improving scores is feasible, creating meaningful benchmarks at the AI frontier remains a major challenge.
The paper presents TrackJoint, a Modelica MultiBody component that represents a 3D road surface as a two-degree-of-freedom joint for race car simulation and minimum-lap-time optimization, validated on the Singapore Formula 1 circuit.
The paper introduces Rollback, a policy improvement paradigm that turns action-conditioned world models from forward-only simulators into counterfactual engines that rewind to critical decision points and branch into alternative actions. On three real-world robot tasks it outperforms the strongest baseline by absolute success-rate gains of 10.67%, 11.11%, and 11.85%.
UT-HMR is the first one-stage HMR framework to capture correlated pose uncertainty across joints and use temporal early fusion to resolve ambiguity in sports scenarios.
Researchers introduce Wiener-informed Decoding (WiD), which leverages frozen Wiener statistical estimates and coefficient interfaces to accelerate pixel-space flow model training and improve generation quality. On ImageNet 256×256, PixelDiT-XL with WiD reaches FID 1.53 after 500 epochs, using 37.5% fewer training epochs than the baseline.
The paper introduces PrinGS, a representation-level anonymization framework for 3D Gaussian Splatting that removes sensitive content from a trained splat in under one second, and adds a privacy-by-design training strategy that keeps faces, license plates, and text out of the representation during optimization.
Researchers introduce GeoUG, a unified geometry understanding and generation model that integrates depth and multi-view generation with spatial understanding, along with the 3DCueBench benchmark and adaptive interleaved reasoning that uses reinforcement learning to decide when to generate geometric evidence.
ObjectTwin introduces a scan-grounded measurement protocol that freezes the instance, cameras, scan-projected support, recorded depth, metric anchor, and evaluation denominator, then replaces only the object representation to isolate the loss of observation agreement. Across two indoor RGB-D capture domains and two generation backends, visible, anchored generated assets lose 0.302-0.497 MaskIoU under this replacement, and generator preference on the same input reverses between capture domains.
A new paper trains ModernBERT-based Icelandic encoders from 12M to 1B parameters on 9B tokens of open data and re-evaluates 30 models under one protocol. The 1B scratch model significantly beats IceBERT on all 11 tasks, while a version warm-started from the same-shape English Ettin encoder achieves the best score on every ranked task.
Researchers propose ReCoPII, a framework that automatically builds contextual PII masking datasets by planning ontology-constrained entity relations and generating insertion sentences with placeholders. On Korean documents, it significantly outperforms SPY-style augmentation, and joint fine-tuning of a PII detector reduces over-masking within a fixed leakage budget.
The paper proposes Multivariable Predictability Score (MAPS), a model-agnostic metric for measuring the intrinsic predictability of multivariable time series forecasting tasks, validated on synthetic and real-world datasets against six baselines.
The paper defines a Fisher-aligned parameter distribution that, within the PAC-Bayes framework, simultaneously guarantees low loss and a large parameter log-volume, yielding a high-probability generalization bound that scales with the expected rank of the Fisher information matrix.
This paper evaluates post-hoc calibration of click-through rate models on Avazu (40.4M impressions) using rolling temporal origins versus random splits. Random splits understate calibration error, ECE is not monotone in temporal distance, and on fixed evaluation sets a 12-hour fitting slice reduces harmful recalibrations more than 6- or 24-hour slices; subsampling the 12-hour slice to 1-hour size retains 88% of its gain while fitting 20-25x faster.
Neural-DISCO is a disentangled representation learning framework for source-conditioned counterfactual editing of neural activity. It partitions the latent space into label-specific components and a residual latent, enabling selective editing of behavioral or stimulus labels while preserving other source content. Experiments on simulated and real neural recordings show effective counterfactual generation, and pairing the framework with feature attribution improves recovery of known functional cell roles.
A new arXiv paper introduces PAI-KNN, a k-nearest neighbor classifier that adapts to concept drift and class imbalance by adjusting class priors at inference time rather than modifying memory, achieving competitive performance on synthetic streams and particularly strong results under dynamic imbalance.
The paper shows that harness skills do not benefit all models equally: skills that boost strong models can hurt weaker ones. It introduces a fixed meta-skill that guides skill reformulation to fit the executor and improves seven models by 2.0-5.1 points inside a five-round self-evolution loop.
Researchers formulate time-to-arrival (TTA) in sports as conditional quantile regression, training a lightweight MLP on all 64 matches of the 2022 FIFA World Cup tracking data to learn the lower envelope of player locomotion times, consistently outperforming six physics-based baselines and a learned reachable-hull baseline.
StruPO introduces a structure-aware prompt editor that turns calibrated scene-graph feedback into a reusable editing policy, improving relational fidelity without an iterative critique-rewrite loop.
This arXiv paper argues that as AI matches or exceeds human researchers on core scientific tasks, the question of how to use AI in science can only be answered by first addressing why humans should keep doing science. It identifies durable, collective human understanding of the world—pursued in service of human ends and values—as an indispensable aim of science, and derives a two-part test for delegating research tasks to AI: human understanding and human direction. Applied across the Contributor Roles Taxonomy (CRediT), the framework offers field-tailorable recommendations and extends to peer review and the training of new researchers.
This study proposes a comprehensive data generation strategy combining traditional augmentation, CutMix, MixUp, and DCGAN synthetic images, and evaluates its impact on olive tree disease classification using EfficientNetV2B0 under three data distribution scenarios, showing substantial accuracy improvement.
A new arXiv paper argues that measuring time saved by AI requires a ledger view rather than a stopwatch view, accounting for escaped defects alongside speed gains, and presents model-free metrics, a baseline checking model, and replications of METR and GDPval studies.
Databricks shares lessons from extracting ~40 structured fields from partner order documents. A spot fix for one field cut its errors in half but increased total extraction errors by 9%. The team recommends field-level rules in the schema, document-level logic in global prompts, precision mode for complex structures, citations for diagnosing errors, and versioning both prompts and schemas with MLflow Prompt Registry. Evaluate candidates on fixed human-reviewed ground truth before promotion. Genie Code can automate the core evaluation and hill-climbing loop.
Property Finder discovered 96.8% of their vision-AI evaluation labels were inherited from the incumbent model, creating bias. They built unbiased ground truth, applied evolutionary prompt optimization, and selected a more accurate, cost-effective model on Amazon Bedrock, cutting costs by 86%.
Amazon engineering leader Terk Terkowitz says that if current AI tooling does not benefit your work, that is a finding, not a failing. Engineers should document why, share it with peers and managers, and seek solutions.
Two young OpenAI managers explore three ways AI might enter the org chart—beneath, beside, and between us—and what these scenarios mean for worker agency, firm size, pay disparity, and the career ladder.
Legora uses OpenAI models, ChatGPT, and Codex to power its legal AI platform, helping legal teams accelerate contract review, broaden analysis, and focus on higher-value work. The company reports major efficiency gains internally and for customers.
Xiaomi's Embodied Intelligence Team introduces OmniBrain-0, a lightweight unified embodied agent that coordinates manipulation and navigation via a shared high-level decision model, trained with progressive distillation and recursive self-improvement.
The paper proposes Skill Dreaming, a framework that extracts executable knowledge from a repository's pull request history to build and continuously refine repository-specific skill libraries, guiding coding agents to implement features while maintaining tests, documentation, compatibility, and conventions.
CohortFlow turns natural language corrections into localized updates of a structured cohort specification and reconciles them with a persistent execution graph, avoiding reprocessing of unchanged requirements. Using the same 8B model, it reaches 77.0% requirement-level micro-F1 versus 76.4% for full regeneration (not statistically significant), while cutting mean correction latency by 64.8% after completion and 44.3% during execution, with 38.1% fewer model calls.
Databricks introduces Vital Skills, open-source Agent Skills that equip coding agents with healthcare and life sciences domain expertise. On BiomniBench, Genie Code with Vital Skills scored higher on in-scope tasks, improving consistency, accuracy, and rigor without negatively impacting out-of-scope tasks.
AWS announced that Amazon Connect has been named a Leader in the 2026 Gartner Magic Quadrant for Contact Center as a Service for the fourth consecutive year, emphasizing customer context, end-to-end resolution, and auditable AI over simple deflection.
LangChain shares how it built Content Studio, an open-source design agent in Slack that creates on-brand videos, graphics, and decks, and the harness that powers it.
The paper extends PPO into a phase-conditioned policy optimization method that uses a phase-conditioned critic, phase-gated rewards, and a multi-phase reset curriculum to train a single policy for long-horizon loco-manipulation tasks, demonstrated on the Unitree G1 and a Boston Dynamics Spot with an arm, without relying on human motion data.
The paper introduces CLAP-AD, a planning paradigm that learns paths via imitation while scoring candidate speeds with a learned function supervised by physically grounded safety constraints. As a plug-in planning head, CLAP-AD enables explicit test-time speed control without speed-conditioned training. It achieves a state-of-the-art Speed-Adherence Score of 82.53 on Bench2Drive-Speed, improves driving score by +28.5 and +32.7 over UniAD and VAD on Bench2Drive, and boosts EPDMS by +6.1 over its backbone on NAVSIM v2.
The paper argues that image quality assessment (IQA) should shift from task-agnostic score regression aligned with generic human ratings to goal-conditioned judgment defined by embodied task utility. It proposes an MLLM-based embodied IQA agent framework, and a compact VisDrone evidence study shows that conventional perceptual metrics and generic MLLM-based IQA scores are weak proxies for downstream VLM task success, while a task-conditioned utility score better predicts whether visual input supports the target task.
cuArms is a conflict-based planner co-designed with GPU parallelism for multi-arm manipulation; it decomposes the search space with reusable multi-modal roadmaps per robot, resolves conflicts and deadlocks via lazy search, and pairs with a high-level scheduler to jointly optimize task assignment, schedules, motions, and skills, achieving real-time initial solutions and anytime improvements without days of domain-specific training.
A study on five-class respiratory sound classification finds that standardizing SVM features yields a 27-point macro recall swing—far larger than any gap between classical and deep models—while classical models avoid a 449 MB PyTorch runtime, making them attractive for GPU-less clinics.
A study of 10,420 congressional e-newsletters from 2010 to 2025 finds that Democrats and Republicans often attach the same election terms—such as "integrity," "security," and "fraud"—to very different surrounding words. The parties do not need different vocabulary to tell different stories about elections; they only need to place the same words in different company.
At GitHub Universe, Simon Willison revisited the conference badge, a programmable color-screen computer. He met Steve Francia, who used Gemini to get a full NES emulator running on the badge and then went further with Game Boy and Game Boy Color emulators, meaning Zelda: Link's Awakening can run on it given a ROM. Flipping the badge upside down exposes built-in touch game controls.
AWS introduces SOCI indexing for SageMaker AI training and processing jobs, enabling lazy loading of container images to cut startup time from minutes to seconds without job definition changes.