10 selected stories for 2026-11-09, 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.
HFX3D contains about 492 million LiDAR points from 26 Halifax buildings across four seasons, day and night, and varied weather.
Each facade is annotated at semantic, instance, and per-instance functional attribute tiers, claimed as the first facade dataset pairing geometry and semantics with functional labels.
ActPrune moves the pruning decision ahead of the vision encoder by grounding it in the action chunk predicted at the previous control step.
It applies asymmetric pruning to the third-person and wrist views: retaining end-effector trajectory patches and semantic anchors in the fixed view, while pruning unchanged patches after camera-motion compensation in the wrist view.
NVIDIA unveiled GR00T 2 world action model, Isaac Lab 3.0, Newton physics integration, DexBench and VABAR benchmarks, and agentic robotics workflows at CoRL, advancing robots from locomotion to precise manipulation with partners like Ambi Robotics, Bosch, and Siemens.
A new arXiv paper, STRIDE, proposes a reinforcement learning reward that relies only on the onboard IMU by tracking a commanded acceleration profile instead of base velocity, enabling Unitree G1 and Booster T1 humanoids to learn to walk within about 14 minutes of samples while sharply reducing training falls.
Researchers propose ForPC, a force-aware whole-body control framework that regulates cart velocity via two-stage reinforcement learning. Without external force sensors or human demonstrations, it significantly outperforms a locomotion-only baseline in simulation and transfers zero-shot to a physical Unitree G1, pushing a 30 kg cart at 0.5 m/s.
BEVCast introduces a two-stage bird's-eye-view representation whose resolution adapts to the predicted driving path: a coarse 4.096 m grid encodes the whole scene, and only the highest-heat cells are refined to 1.024 m under a fixed budget, cutting BEV encoder time to 66% while improving accuracy on nuScenes and NAVSIM.
This issue covers an AI startup proving a quantum conjecture, Meta's Tactus gloves for dexterous manipulation data, a NASA paper on affordable lunar industrialization, and Core Automation's distributed training efforts.
This paper presents an object-centric reinforcement learning framework that treats each part as an agent and learns a shared policy for geometric assembly. Using a graph neural network to encode parts, target poses, and spatial relationships, combined with per-agent PPO and a centralized graph critic, the method achieves concurrent assembly in simulation on planar Tangram tasks and transfers to sequential execution on a real robot.
This work presents an automated 3D segmentation pipeline that uses classical medical image-processing techniques to encode known phantom anatomy as explicit geometric and intensity priors, eliminating the need for training data. Across 57 scans, it achieved 0.83 median Dice on 11 contrast-suppressed scans, detected all 32 annotated focal lesions without false positives, and completed scans in 2-8 minutes on CPU versus 1-2 hours manually.
This paper proposes using procedural knowledge encoded in Agent Skills as supervision to precisely determine tool and workspace permissions for subagents. Experiments show significant improvements in permission precision and task success on ClawArena-Team.
FinEvolveKG introduces a knowledge graph construction harness that separates reusable general rules from persona policies and mechanically confines prompt optimization to designated blocks, improving financial KG quality while keeping changes reviewable and attributable.
OpenAI introduces Control Room, a centralized system to track model workloads across research infrastructure and rapidly halt model output within 35 seconds when safety or security concerns arise. The system is designed to fail closed and is resilient to buggy workload code.
CS-Pi introduces a sampling method that enforces novel constraints at inference time for flow-matching policies by combining projection efficiency with rejection-sampling plausibility. It achieves 81% safe success on a 7-DoF shelf task and 55% on a zero-shot Franka deployment, outperforming projection baselines.
The paper shows that common human-intervention modeling assumptions do not match real behavior, introduces a progress-based intervention model, and proposes Q-OIL, which adds optimistic bonuses to intervention transitions with decoupled critics, achieving 10-50% higher success in simulation and on real robots.
A new paper shows that consolidating web-scraped retail product catalogs yields accuracy gains primarily from changing the evaluation label space, not from improved model learning.
VoluCompose is a training-free, model-agnostic framework that decouples what visual content to retain from how many tokens to allocate, using hyperbolic token composition and attention-entropy-based capacity control to compress medical volumetric images. Across three datasets and four backbones, it retains 22.78% of visual tokens on average while reaching 121.68% mean relative quality versus uncompressed references.
The paper introduces HaNDF, an object-conditioned neural distance field used as a data-driven prior for plausible hand-object configurations. Leveraging Riemannian geometry, it represents plausible hand poses as the zero-level set of an object-induced conditional field and can also optimize for the interacting object, supporting pose refinement, reconstruction under occlusion, and physically plausible manipulation synthesis. It was accepted at NeurIPS 2026.
The paper introduces Semantic Conditional Similarity (SCS), a conditional form of STS that computes similarity between two documents after eliminating semantic elements shared with a third, extending pairwise STS to ordered triples with directionality and role-dependent asymmetry.
An arXiv study uses qualitative coding and computational analysis of lexical change to examine how English Wikipedia editors and students identify AI-generated text. It finds that such identifications draw on diverse and often contradictory evidence, are context-dependent, and are frequently motivated by 'pragmatic infelicity'—a perceived mismatch between textual elements and contextual expectations—rather than any specific feature. AI identification thus serves as a discursive resource for maintaining sociolinguistic norms, and detection tools that treat AI-likeness as context-free may inherit and enforce the norms of their training contexts at scale.
OpenAI introduces Control Room, a centralized system to track model workloads across research infrastructure and stop them quickly when safety or security concerns arise.
The paper analyzes how U.S. states regulate SAE Level 4-5 automated driving systems, proposes a three-model certification typology, compares practices in the EU, China, Japan, and Australia, and finds North Carolina currently follows a baseline compliance model.
Korea's Ministry of Science and ICT and NVIDIA hosted the AI Startup Partnership Forum at COEX, where NVIDIA announced an investment in Point2 and planned investments in Motif Technologies, RLWRLD, and Upstage, covering the full AI stack from infrastructure to models to robotics applications.
NVIDIA is helping advance Korea's physical AI ecosystem by tapping into the country's manufacturing and engineering strengths. Industry leaders like Samsung, Kakao Mobility, and LG Electronics are applying accelerated computing, AI, and simulation to semiconductor manufacturing, robotics, and autonomous vehicles, bringing physical AI from development into real-world production.
InstaGen recasts block selection in permuted block-sparse attention (PBSA) as importance sampling: top-scoring keys are computed exactly, the rest are sampled in proportion to estimated attention and reweighted by inverse sampling probability, with one sample shared across each query cluster. Parameterized by SVG2, it improves PSNR by 3.4 dB and 2 dB over SVG2 on Wan2.1-T2V-1.3B and 14B at matched attention density, and by 1.6 dB and 2.2 dB at matched wall-clock time; on the 14B model it also beats SVG-EAR by 1.1 dB at similar density while running faster.
NVIDIA releases Dynamo-optimized deployment recipes for Korea’s top open models. Optimized for B200 GPUs, the recipes deliver up to 2x performance out of the box, helping developers deploy models faster.
Hyundai Motor Group is exploring NVIDIA NemoClaw blueprints and the NVIDIA OpenShell runtime, while AIM Intelligence integrates its Starfort platform with OpenShell to support secure AI agents. LG AI Research is opening API access for K-EXAONE-2.5, developed using NVIDIA Nemotron datasets.
VICTR introduces progress-informed context retrieval for vision-language-action models via an In-Context Value Function Estimator (IC-VFE). By estimating task-relative progress from a reference demonstration, a language instruction, and the current observation, VICTR selects demonstration chunks at matching execution stages for a π0.5 policy post-trained for in-context action generation. On unseen out-of-domain real-world manipulation tasks, it raises success from 6.2% to 23.8% and task progress from 28.9% to 55.5%.
Meituan Robotics introduces MR0, a vision-language-action foundation model and embodied-agent system for stationary-base desktop manipulation. It combines 147 million vision-language samples with 15,060 hours of robot trajectories across 12 embodiments, uses a three-stage training recipe, and proposes Rollback for CFG-DAgger. MR0 achieves 93.6% success on RoboTwin 2.0 and ranks fifth in the RoboDojo VLA track.
This paper introduces TeViD, the first multimodal video fact-checking dataset for Turkish, and compares two verification pipelines built on the same open-source base model: a zero-shot agentic pipeline (3MFact-TR) and a LoRA fine-tuned closed-book model, reporting accuracy, macro F1, and ablation studies.
A study accepted to WACV 2027 finds that vision-language models can transfer hidden preferences—such as a bias toward a specific animal or facial attribute—from a teacher model to a student model through an unrelated visual task involving geometric shapes and generated numbers. The transfer reaches up to about 10 percentage points in the animal preference task and also shifts how students explain their choices.
This paper studies recursive self-training of language models and finds that as training temperature rises, different Renyi entropies (min-entropy, collision entropy, Shannon entropy, and observed richness) cross from loss to gain at distinct temperatures in a consistent order, which persists across various settings.
A paired audit of 25 biomedical fine-tuned models against 11 base checkpoints finds that appropriate knowledge-centric abstention on MedAbstain drops from a median of 8.3% in base models to 4.2% after fine-tuning, decreasing in 21 of 25 pairs with heterogeneous magnitude and direction.
A controlled study finds that organizing a respondent's survey history into a fixed Background–Decision procedure–Evaluation (BDE) profile improves LLM digital-twin prediction accuracy over raw survey transcripts, and that the gain stems mainly from profile composition rather than compression or deeper inference.
Researchers propose RL-ATTACK, a trainable reinforcement learning framework for black-box adversarial attacks that learns from the target model to generate tailored adversarial transformations. It drives post-attack accuracy to 2% within 20 queries on natural language inference and outperforms an untrained attacker and a strong optimization baseline in generative Q&A. The paper is accepted at IEEE EdgeCom 2026.
A study of seven LLMs on the GoEmotions dataset finds that hidden representations overcompress fine-grained emotion categories into coarser sentiment and emotion-family groupings, yet linear probes retain substantial fine-grained predictive information.
A study accepted to BlackboxNLP 2026 provides the first mechanistic explanation for why multiple-choice question format reduces refusal rates in LLMs: models commit quickly to answer letters, activations lie farther from the refusal direction, and answer selection competes with refusal.
Simon Willison, inspired by a Hacker News comment on Paul Ford's essay, used Claude Opus 5.5 to port a Three.js Nudibranch-themed Faberge Egg to vanilla JavaScript and WebGL without Three.js. The original is about 77KB and the port is about 98KB; the two are not identical. Willison finds the Three.js code slightly easier to read but says he rarely looks at the code for such 3D projects.