Weekly notable papers by field
October 5, 2026 ~ October 11, 2026. AI papers are ranked by Hugging Face Daily Papers upvotes; healthcare papers are picked from open-access papers on Europe PMC (PubMed Central, bioRxiv, medRxiv).
Jump to a field · fields with more picks first
Language models
-
On-policy parameter update directions govern generalization in LLM post-training
A study that shows SFT updates parameters in a consistent direction while on-policy training keeps changing direction, and tries to use this as a principle for improving SFT generalization
arXiv 2609.36659 · Upvotes 87 · 2026-09-29 · PDF
-
Latent-MOPD: Multi-teacher on-policy distillation at the hidden representation level
The first representation-level multi-teacher OPD method for LLMs, which transfers to the student not only the output distributions of multiple specialist teacher models but also their hidden states
arXiv 2610.02381 · Upvotes 70 · 2026-10-01 · PDF
-
Doubt the question: How to sustain self-evolution in reasoning models
Identifies the cause of performance collapse during repeated training on self-generated problems as the accumulation of flawed problems and mathematically equivalent variants, and explores ways to sustain self-evolution
arXiv 2610.04299 · Upvotes 70 · 2026-10-03 · PDF
Generative media
-
UltraText Bench: A Korean–English and Chinese–English bilingual benchmark for text rendering in image generation
An English–Chinese benchmark that uses 432 prompts across 24 real-world scenes and three difficulty levels to evaluate whether models accurately render long text in multiple regions of an image
-
FrameMorrow: Long video generation that selects past frames based on future information
A study that addresses the memory cost and redundancy of ever-longer video generation, based on the idea of selecting generation history by the information that will be needed later rather than what is needed now
-
GRACE: Generation-aware latent compression for efficient video generation
A method that addresses the retraining cost caused by highly compressed video autoencoders being misaligned with existing DiTs, and accelerates video diffusion models by compressing the latent space with generation quality in mind
Computer vision
-
EditHero: A benchmark for long-horizon part-level 3D editing
The first benchmark for evaluating multi-turn part-level 3D editing driven by natural-language instructions and target images, comparing agentic and non-agentic editing approaches
-
World Embedding Benchmark: Evaluating whether video representations capture physical information
Measures, through three tasks, how much physical information video representations capture, using 8,000 simulated videos from 80 families—including fluids, solid mechanics and optics—with physics annotations
-
Tetris3D: Generating 3D scenes with interlocking objects
Reconstructs physically interlocking 3D scenes from a single image by generating each object conditioned on the shapes and physical relations of surrounding objects, and also releases ComOb, a dataset of 1.2 million scenes
Multimodal
-
VepAgent: A video event prediction agent that bridges causal transitions with tool-use reinforcement learning
An agent framework that strengthens video event prediction by explicitly modeling the logical flow from the last observed state to future events and combining it with tool-use reinforcement learning
arXiv 2610.06293 · Upvotes 55 · 2026-10-05 · PDF
-
OmniReasoning: Pushing the limits of joint audio-visual reasoning
Evaluates and improves joint audio-visual reasoning by presenting a 1,150-question benchmark that requires both audio and visual evidence, an automated data engine and a training method
-
OneSearch-VL: A unified multimodal deep research agent covering images and video
Unifies single-image, multi-image and video deep research in one agent, centered on an evidence graph (VGEG) that links visual grounding, entity relations and source facts
Speech and audio
-
DEFINE: Zero-shot TTS that controls accent using reference speech
Zero-shot TTS that conditions speaker timbre and accent separately on different reference utterances, and continuously adjusts accent strength with a single weight without retraining
-
WorldSonus: Adding sound to world models
An interactive video-audio generation framework for world models, aimed at real-time streaming generation, responding to sound instructions mid-generation, and spatial stereo matched to the scene and camera
-
HelixWorld: A real-time interactive audio-visual world model
A real-time world model in which video and camera-based spatial stereo sound change together in response to user actions, targeting the limitation of existing silent world models
Agents
-
Scaling trajectories for complex tasks with recursive self-rewriting
Proposes RSR, in which a single model reconstructs successful solutions found across multiple harnesses into training trajectories for a general harness, solving 759 of about 3,000 terminal tasks
arXiv 2610.02826 · Upvotes 103 · 2026-10-02 · PDF
-
VeriHarness: Scaling agentic verification for long-horizon tasks
A method that turns an LLM into an agentic verifier by giving it a workspace and evidence tools, even without reference answers or grading criteria, and uses it to identify trustworthy claims across multiple rollouts
-
HyperBrowseComp: A multilingual, multimodal stress test for web browsing agents
A web browsing agent benchmark of 423 hard questions written by native speakers of 13 languages, requiring searches across heterogeneous sources such as videos, scanned documents and maps
arXiv 2610.03574 · Upvotes 60 · 2026-10-02 · PDF
Physical AI and robotics
-
MotorMind: Zero-shot robot manipulation with general-purpose VLMs
A study that attempts zero-shot robot manipulation by having a general-purpose VLM reason directly from observations and output actions like a teleoperator, without relying on external tools
arXiv 2609.38078 · Upvotes 114 · 2026-09-29 · PDF
-
ProWAM: A world action model with progressive visual planning
A world action model that predicts ordered, sparse visual subgoals along with actions to guide action generation throughout long-horizon tasks
arXiv 2610.02508 · Upvotes 97 · 2026-10-01 · PDF
-
NAVA-WAM: A world action model that learns action priors directly from video
A training method for world action models that pretrains action policies directly from observation videos without action labels, reducing reliance on expensive robot action annotation data
arXiv 2610.03391 · Upvotes 88 · 2026-10-02 · PDF
Reinforcement learning
-
Self-Retrospection Distillation: Turning hindsight experience into predictive ability
Targets the RLVR limitation in which the learning signal vanishes when all rollouts receive the same reward, proposing prospective learning, which uses hindsight experience to supervise predictions made before acting, and SRD
-
MiMo-V2.6: Scaling reinforcement learning toward self-improvement
A technical report on the omni-modal MiMo-V2.6 model series, describing how reinforcement learning compute was scaled in three directions, including asynchronous training with up to 1M context
arXiv 2610.11959 · Upvotes 74 · 2026-10-08 · PDF
-
MetaRubric: Reward learning for rubric-based reinforcement learning
Identifies the 'vacuous score' problem, in which points are awarded even for information absent from a response, and improves the reliability of rubric rewards by alternating counterfactual prompts and rubric adaptation
Efficiency and infrastructure
-
LoGRA: Scaling LLM reinforcement learning with low-rank gradient sketches
Reduces memory usage in LLM reinforcement learning post-training by up to 45.7% by compressing gradients to low rank and adjusting update size with KL prediction
arXiv 2610.06647 · Upvotes 147 · 2026-10-05 · PDF
-
TRIAGE: Stabilizing mismatch in native NVFP4 reinforcement learning
Analyzes how the mismatch between trainer and sampler under low-precision NVFP4 execution interacts with policy update directions to destabilize training, and proposes a method to stabilize it
-
MC-Sparse: Closing the dense-sparse attention gap in Diffusion Transformers
Breaks down the causes of quality loss at high sparsity into three factors and reduces video and 3D generation latency with training-free sparse attention that selects individual KV tokens
AI security
-
LTBD: Learnable trust boundary delimiters for prompt injection defense
A lightweight defense that leaves LLM parameters unchanged and uses a small number of learnable delimiters to mark the trust source of inputs, blocking malicious instructions in external data
arXiv 2610.11634 · 2026-10-08 · PDF
-
Package hallucination attacks on coding agents via rule-file prompt injection
Presents an attack that plants malicious prompts in shared rule files such as AGENTS.md to make coding agents install attacker packages, along with PackHallu, which optimizes this attack
arXiv 2610.09264 · 2026-10-07 · PDF
-
AdvSim2Real: Training web agents against adaptive prompt injection inside a web world model
A defense training method for web agents that co-evolves a task curriculum, an injection attacker and the agent inside a web world model, overcoming the limits of training against fixed attacks
arXiv 2610.08773 · 2026-10-06 · PDF
Safety and alignment
-
Source preference in the wild: LLM agents choose items by source
Shows that 12 agent models favor items by source in shopping, lodging and paper recommendations even when conditions are identical, and that this preference can outweigh how well items meet requirements, and presents mitigation methods
arXiv 2610.03195 · Upvotes 47 · 2026-10-02 · PDF
-
Making LLMs say what they think: Measuring and improving alignment between CoT and internal reasoning
Proposes CIA, a metric that measures agreement between CoT narratives and the model's internal reasoning strategies detected with interpretability tools, measures it on three tasks, and explores ways to improve it
-
U-Space: Revealing when and why uncertainty arises in language models
An approach that shows where uncertainty arises during reasoning and how it changes, without repeated generation or a separate model, to help judge the reliability of individual answers
Science & Bio
-
PDE-JEPA: Representation learning that predicts latent dynamics of parametric partial differential equations
Shows that prediction-based Representation pretraining, rather than reconstruction, preserves rich physical information, and builds on this to propose PDE-JEPA, a model of PDE dynamics
-
Optimizing the optimizer: language models find faster algorithms for molecular structure optimization
Has an agent rewrite the quantum chemistry structure optimizer itself to reduce the number of force evaluations, starting from Sella, the fastest open-source optimizer, and arriving at AutoSella
-
Does learning protein folding generalize to broader reasoning ability?
Uses FoldingCorpus, a question-answer dataset built from protein structures with exactly verifiable answers, and Fold2Reason, a post-training method, to test whether learning protein folding helps general reasoning
Data & Evaluation
-
RealCompanion: a benchmark that evaluates AI companions' understanding of people using real long-term conversations
Releases 27,218 messages from real conversations between people and AI companions over up to 120 days, and links every label to its supporting messages to evaluate long-term user understanding
arXiv 2610.01780 · Upvotes 282 · 2026-10-01 · PDF
-
Science or slop? A benchmark and mitigation for 'scientific slop' in AI-generated papers
Defines, on six scales, the problem of AI-generated papers whose individual parts look plausible but whose scientific logic breaks down, and builds SciSlopBench, 390 papers paired with human-written papers
-
LESSER: selecting post-training data with output-layer gradients
Shows that output-layer gradients, obtained from a forward pass alone instead of full-parameter gradients, are enough for effective data selection, and implements this as a wrapper that plugs directly into existing methods
arXiv 2610.03702 · 2026-10-02 · PDF
Data Engineering
-
Zephon: elastic determinism for online, stateful foundation model data loading
A data loader that guarantees the same training batch order even when the GPU configuration changes, training resumes from a checkpoint or the execution backend differs, making comparisons between expensive experiments more reliable
arXiv 2610.03087 · 2026-10-02 · PDF
-
ARCS: accurate Text-to-SQL through structured disambiguation
Proposes 'structured disambiguation', which resolves ambiguity in questions through explicit, constrained interaction instead of free-form dialogue, and builds ARCS, the first Text-to-SQL benchmark for it
arXiv 2610.09396 · 2026-10-07 · PDF
-
Argo-Bench: evaluating data agents with enterprise-scale workflows
Simulates a New York food delivery platform at real scale and uses 210 data analysis tasks spanning dozens of tables to evaluate data agent abilities beyond query generation
arXiv 2610.02122 · 2026-10-01 · PDF
MLOps·LLMOps
-
AGO AI Quality Gate: an evidence-first quality gate for RAG deployment decisions
A quality gate used in real industrial RAG evaluation that treats missing data and LLM judge errors as explicit states and assesses regression risk probabilistically to decide whether to deploy, fix or block
arXiv 2610.01218 · Upvotes 27 · 2026-10-01 · PDF
-
MOLT: fine-grained GPU memory sharing for LLM serving and opportunistic fine-tuning
A system that uses idle GPU memory inside serving replicas for parameter-efficient fine-tuning, while letting inference reclaim the memory within seconds to meet latency SLOs
arXiv 2610.05748 · 2026-10-05 · PDF
-
Lachesis: lifetime-aware agent KV cache placement across HBM and high-bandwidth flash
A method that splits the KV cache in agent serving between HBM and high-bandwidth flash according to its lifetime, increasing capacity while avoiding flash's write endurance limits
arXiv 2610.08378 · 2026-10-06 · PDF
Traditional ML, Recommendation & Search
-
Test-time compute for tabular foundation models: mechanisms, gains and limits
Analyzes test-time compute for tabular foundation models along three axes (adaptation, aggregation and context construction) and proposes DiagScale, which reaches full fine-tuning-level performance while training only 0.003–0.03% of parameters
arXiv 2610.12005 · 2026-10-08 · PDF
-
Small, efficient indexes for Learned Sparse Retrieval
Replaces SEISMIC's inverted-index block summaries with representative document IDs and compresses the forward index, greatly reducing Learned Sparse retrieval index memory while maintaining search efficiency
arXiv 2610.12300 · 2026-10-08 · PDF
-
Autoregressive Retriever: universal multimodal retrieval that improves query understanding with feedback from retrieval results
A multimodal retrieval model that alternates between retrieving items and updating the query embedding, using retrieved content to refine the information need; trained with SFT and reinforcement learning
arXiv 2610.11666 · 2026-10-08 · PDF
Medicine & Bio
-
A cancer genotype foundation model that enables precise prediction of treatment response
MutationProjector, pretrained on genomic mutations from over 30,000 tumors, achieves state-of-the-art performance in predicting resistance to immunotherapy and chemotherapy across multiple cancer types and also finds new biomarkers
Cancer discovery · Europe PMC · 2026-10-01 · PDF
-
Extracting tumor data from unstructured medical reports: applying an AI tool at the Eastern Switzerland cancer registry
A real deployment in which a tool built by fine-tuning a German BERT extracted tumor information from 82,384 medical reports from about 200 institutions and fed it directly into cancer registry software
ESMO real world data and digital oncology · Europe PMC · 2026-09-23 · PDF
-
Guideline-based LLM chatbots supporting self-management for rheumatology patients: a nationwide multicenter evaluation
Deploys 10 disease-specific chatbots grounded in German guidelines to 13 centers and 6 patient organizations, analyzes 6,291 questions, and compares LLM judge ratings with assessments by rheumatologists
Journal of medical systems · Europe PMC · 2026-10-05 · PDF
Finance & Insurance
-
From benchmark to production: a Text-to-SQL system for complex financial data
FLINT, a domain-specific system that uses components such as a lookup agent to address how general-purpose methods drop below 50% on real financial databases where concepts are stored as opaque integer keys
arXiv 2610.03524 · 2026-10-02 · PDF
-
Ranking prior alignment for credit risk models: when does external prior knowledge matter?
A framework that, for cold-start credit scoring of new loan products with scarce labels, uses a KL loss to Distill ranking priors from experts, Teacher models and LLMs into any model
arXiv 2610.11146 · 2026-10-08 · PDF
-
MintEval: do LLMs correctly implement the trading strategies they are asked for?
A benchmark that evaluates behavioral equivalence between natural-language strategies and generated code, to catch silent failures where the code runs but behaves differently from the risk logic the trader described
arXiv 2610.03080 · 2026-10-02 · PDF
Manufacturing
-
Is in-domain training alone enough for precise industrial anomaly understanding?
Shows that even trained specialist models reach only 75.5% accuracy in locating defects, short of human experts (92.3%), and proposes SiGMA, a multi-agent system based on spatial grounding
arXiv 2610.12310 · 2026-10-08 · PDF
-
Adaptive sampling based on conditional Flow Matching for generalization of PINNs in metal additive manufacturing
Shows theoretically that adaptive collocation sampling tailored to process conditions is better than fixed sampling, and improves the generalization of PINNs for thermal modeling in metal 3D printing
arXiv 2610.09126 · 2026-10-06 · PDF
-
Narrow and deep: ontology towers as knowledge for LLM agents on industrial equipment
Proposes an ontology that captures, in depth, physically derived quantities and lessons from operating logs for a single piece of equipment, and applies it to an LLM agent that operates a real low-humidity HVAC test facility every day
arXiv 2610.11768 · 2026-10-08 · PDF
Legal & Professional Services
-
LexReward: a taxonomy-based reward framework for legal language models
A reward model framework that evaluates the quality of legal answers along three dimensions (style, legal elements and reasoning chain), providing interpretable reward signals beyond whether an answer is correct
arXiv 2609.39071 · Upvotes 66 · 2026-09-30 · PDF
-
When citations mislead: a benchmark for claim-level legal hallucination detection
Labels 3,396 claims built from New York Court of Appeals decisions as supported, contradicted or unsupported, to evaluate whether LLMs can tell if cited cases actually back the claim
arXiv 2610.10971 · 2026-10-07 · PDF
-
Legal Research Bench: measuring end-to-end reliability of long-horizon legal research agents
Attaches answers, supporting cases and grading criteria to 413 expert-written US legal research questions to measure agent reliability that collapses with a single missed case or outdated citation
arXiv 2610.00609 · 2026-09-30 · PDF
Autonomous driving
-
Digital twin-based Real2Sim2Real: Generating simulator conditions through driving scene reconstruction
Tests how far a pipeline that reconstructs real driving clips inside a simulator to produce generated data without annotation costs can replace real data from new regions
arXiv 2610.08339 · 2026-10-06 · PDF
Past weekly lists
One-line summaries were written by AI from the abstracts. Please check the original papers for details.