Hugging Face Daily Papers
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
EmbodiedLab: robots tested on experiments, and the best one solves less than half
The EmbodiedLab benchmark drops an AI agent into a lab simulator: 30 experiments, 14 setups, and even the strongest agent with native control succeeds less than half the time.
S3AC brings the critic back to RL for LLMs: gains from 1.5B to 397B parameters
Shanghai AI Laboratory researchers trained a value critic to work with long chains of thought. S3AC beat GRPO-X on models from 1.5B to 397B parameters and held up under single-rollout training on a 35B model.
Skill∞: an agent improves its own skills and its own meta-skill, 60.7% on SkillsBench
NUS researchers trained a meta-agent to optimize skills and rewrite its own optimization procedure. Across 83 SkillsBench tasks it scored 60.7%, versus 56.4% for human-written skills.
A 12-Author Review Charts How to Trust AI That Improves Itself
Researchers broke down recursive self-improvement in AI: four trust themes, three lenses of verification, and S1-S4 stages where quality gains can easily turn out to be an illusion.
Work Agents: 16 Researchers Build a Unified Coordinate System for AI Agents
The Work Agents survey defines an agent as a pairing of a model and a runtime harness, introduces two axes of working conditions, five forms of digital labor, and five families of capabilities.
MLE-World predicts the outcome of an ML experiment without running it: 77.9% accuracy
MLE-World predicts the logs, errors and score of an ML program from code alone, with no training. It matches real output 77.9% of the time vs 41.1% for larger LLMs; agent gain up to 22.1%.
VibeVideoBench: Agents Write Videos as Code, and the Best One Handles 76.3% of Tasks
A new benchmark tests whether AI agents can build a finished video with sound using only code in one hour. GPT-6 Astra and Claude Opus 5.5 lead, but no one avoids errors.
Thinking-in-Bets: Agents Lock In Claims Before Criticism and Sycophancy Drops
A fine-tuning-free method makes an LLM agent write down claims and confidence levels before reading a user's complaint, and false rollbacks during patch review almost disappear.
VMem-Bench Shows: Memory in Video Generators Drops Accuracy to 0.845
Researchers tested visual memory in long-form video generation: entity recall rose to 0.693, but avoidance of forbidden objects fell from 0.968 to 0.845.
Constrained Decoding Doesn't Break LLM Reasoning: Tests on 11 Open Models
Researchers tested 4 model families and 11 open models, showing that performance drops under constrained decoding are more often caused by mismatched prompts and formats than by the constraint itself.
OBB-PBD teaches LocateAnything-3B to find rotated objects: rotIoU jumps from 0.125 to 0.303
Korean researchers extended parallel box decoding to oriented boxes on aerial imagery: rotIoU climbed from 0.1253 to 0.303, with code and 9 checkpoints released openly.
iPrefBench: iOS agents get 91.6% of actions right but solve only 47.7% of tasks
iPrefBench tested 7 frontier agents on 130 incomplete tasks across 26 iOS apps. Action scores reach 91.6%, but weak personalization caps full task success at 47.7%.
ThinkBit shrinks Qwen3-14B to 3.74 GB and still scores 94.8 on MATH-500
ThinkBit's 2-bit scalar quantization preserves reasoning: Qwen3-14B shrinks from 29.5 GB to 3.74 GB and scores 94.8 on MATH-500.
ProAS speeds up LLM agent training by 2.44x, pushes ALFWorld to 94.3%
The LINs-lab team has introduced ProAS, a plug-in for critic-free GRPO that distributes reward across agent steps more precisely. It reaches 94.3% success on ALFWorld and 70.8% on WebShop.
AgentProofBench: CMU researchers test whether web agent verifiers can be reused
The AgentProofBench benchmark, spanning 97 tasks across 4 web services, found verifiers from DeepSeek Flash and Pro survive requirement changes in 82/96 and 79/96 cases, while test-panel errors predict delivery failure.
Δτ-Bench: Best AI Agent Solves Only 42.3% of Tasks in a Shifting World
Researchers from UCL and co-authors have released Δτ-Bench: 1,588 tasks across 42 domains where tools vanish and get replaced. The best agent managed only 42.3%.
DynaFlow teaches a lead agent to write its own workflows, beats the baseline by 10.9 points
NUS researchers trained a lead agent on Qwen3.5-27B to write its own programs for sub-agents. Result: 88.08% accuracy across six benchmarks, plus 10.90 points over the strongest baseline.
Alibaba DAMO Trains Robots on 30,000 Hours of Video Without Action Labels
RynnVLA-Latent from Alibaba DAMO Academy pretrains on 30,100 hours of video with no action annotations. The 4-billion-parameter model scored 98.8% on LIBERO and 88.6% on RoboTwin 2.0.
AMBIT trains one model to generate and edit 3D sound from text
NoizAI's team has unveiled AMBIT: a text command turns into a ScenePlan with sources and 3D trajectories, and a single Transfusion model renders Ambisonics and edits finished scenes.
SATE-Bench: Multimodal Models Give Up on Scene Exploration Too Soon
A new benchmark, SATE-Bench, covers 316 real-world scenes and 2,525 QA pairs. Top MLLMs stop their scene review prematurely while staying confident in their answers, and GRPO fine-tuning only confuses them further.
DIAYN Agents With Intrinsic Motivation Match Drosophila Behavior Across 20 Algorithms
NeuroAgents Lab researchers trained agents in a digital fly model with 20 algorithms and no external reward. DIAYN did best, matching the fly in both behavior and neural activity.
Independent Critic Lifts Coding-Agent Accuracy by 5.9–9.5 pp
Researchers compared 17 schemes for verifying agent work across seven models. A separate critic added up to +9.5 pp on Terminal-Bench at only 1.13–1.25x the cost.
Qwen-Image-PE Teaches Neural Networks Design: Prompt Enhancer Adds +7.81 Points
The Qwen team trained a prompt enhancer on design principles: the 397B version added 7.81 points to original queries, and the 9B model is already on Hugging Face.
PatEx-Gym Teaches AI Agents to Hunt for Patent Prior Art: 1 Million Patents and a 4B Model
Researchers have released PatEx-Gym, an environment for patent-examiner agents: 1,000 tasks, 1 million patents, 52,000 papers, and a compact 4-billion-parameter agent.
LLM Agents Spent 20 Days Building an AI for Hearthstone Battlegrounds That Beats Top-10 Players
The AutoTavern system didn't play the game itself — it built the tools to play it: a simulator and an RL agent called TavernAgent. Rating climbed from 6,000 to 15,000, with 80% first-place finishes versus 59% for the bes
CMU Researchers Teach RL Models to Stop Padding Answers: FORK Lifts Win Rate from 22% to 74%
The FORK reward adds a precision objective to rubric-based RL and cures bloated, inaccurate answers: win rate against the base model on HealthBench rose from 22.1% to 74.1%.
Mycelium: Researchers Break Down Hallucination Benchmarks Along Three Axes
Julia Belikova and Elisei Rykov analyzed LLM hallucination detection benchmarks along three axes and proposed a new metric, Span Coverage F1, for evaluation at the text-span level.
AV-FABLE: Researchers Assemble 998 Hours of Video to Hunt for Pinpoint Deepfakes
The AV-FABLE benchmark, built from 16,650 original-fake pairs, shows that AV-LLMs don't notice audio and video edits — but their hidden features do. The NFSR method capitalizes on that.
Raspberry Pi 5 off the grid: a signed poisoned knowledge package sailed through every check
The authors ran offline RAG on Qwen2.5 1.5B/3B/7B on a Raspberry Pi 5: signature checks missed 6 of 9 update attacks, while layered verification caught all 9 in 4.5ms.
RoboJev-27B judges robot success from 10 frames: 0.993 AUROC on SO-101
An open 27-billion-parameter VLM watches 10 frames and tells you whether a robot nailed its task: 95.2% balanced accuracy, 0.26 seconds per episode, and about $0.19 per thousand evaluations on an H100.