Hugging Face Daily Papers
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
Researchers Annotate 2,122 WildChat Conversations: 71.1% of Legal Chats Seek Personal Advice
A taxonomy of legal queries to LLMs, 2122 WildChat-1M conversations and open models: 71.1% of chats with legal content ask for personalized or practical advice.
StopLoss Cuts Off Doomed Coding-Agent Runs and Saves Up to 48% of Costs
StopLoss, a small external monitor, predicts agent failure from the latest steps and halts the run: 4.5% lower cost on Verified and up to 31% on SWE-bench Pro.
Changing the definition of "unsafe" moves a guardrail score from 0 to 100 with 0.03% of weights
The authors labeled ~53 thousand prompts twice under different policies and trained 20 guardrail models. Benchmark spread reached 23x, and some test sets were already in the training data.
Researchers Propose Anti-Persistence: LLM Agent Memory Without Stored Rules
A new memory design for LLM agents stores trajectories and outcomes, then rebuilds guidance from scratch for each task. It uses the fewest LLM calls of the designs compared.
3D Scene Generation: DiT Works in a 3D Latent and Encodes a Scene in 1K Tokens
Researchers move scene diffusion from 2D views into an implicit 3D latent space built on DINOv2, SigLIP2 and DA3. Per the paper's title, a scene fits in roughly 1K tokens.
POPE audit: the choice of "yes/no" tokens shifts LLaVA-1.5 F1 by 6.13 points
Researchers tried to reproduce a VLM hallucination-mitigation method and found a flaw in the evaluation itself: wrong token IDs for "yes" and "no" shift POPE F1 by 6.13 points.
Parallel Refinement Instead of Diffusion: DRLMs Hit 93.9% on Sudoku-Extreme
Microsoft has introduced DRLMs — a parallel reasoning method that generates a full draft and iteratively refines it instead of running reverse diffusion. On three hard tasks, a 6M-parameter model reached 44.4% (Countdown
TokenWorth Teaches AI Agents to Forget: Best Results on Three Environments With the Fewest Tokens
Researchers introduce TokenWorth, a lightweight agent memory system that scores the "worth" of memories and outperforms more complex baselines on three LifelongAgentBench environments.
Ropedia W0 trained on 10,000 hours of human video and 3,000 hours of robot data
Ropedia W0 is a world model for walking and manipulation in both humans and robots. Once distilled, it makes minute-long predictions in real time and runs in VR.
TokenGate speeds up LLM prefill by up to 11.3x with learnable token pruning
Liana Patel, Rishabh Ranjan, Ion Stoica, Matei Zaharia, and Carlos Guestrin of UC Berkeley have introduced TokenGate to accelerate prefill in LLMs. The mechanism cuts memory usage by up to 3.4x and speeds up TTFT by up t
T2Quant Reallocates Precision Across Diffusion LLM Steps: Up to +6.7 pp
Researchers propose T2Quant, which quantizes diffusion language models by varying activation bit-width per denoising step, adding up to 6.7 pp accuracy at the same bit budget.
RIDAR catches integrity violations by AI scientists: 3964 cases across 300 runs
The new RIDAR framework detects research integrity violations by autonomous agents with 85.2% F1. Flaws appeared in 99.7% of trajectories, and runtime mode removes 36.8% of violations.
FilmFlow: Four Agents Build 3D Camera Trajectories from Text and Video
The FilmFlow agent system turns text or a reference video into an executable 3D camera trajectory. On CinemaBench, a 750-task benchmark, it beat baseline methods and won in human evaluations.
BrickWright Teaches Coding Agents to Build with LEGO: 100% Valid Builds vs. 9–69%
The BrickWright framework and its Python toolkit BrickLayer lifted physically valid LEGO builds to 100% on 512 prompts, while baseline agents managed between 9% and 69%.
A Tiny 856K-Parameter Policy Speeds Up the Diffusion LLM LLaDA-8B
A controlled study shows how to train unmasking policies for diffusion LLMs: an 856K-parameter policy gave the best accuracy–NFE balance on GSM8K for LLaDA-8B-Instruct.
IQuest-Q1: An Open 320-Billion-Parameter MoE Model for Agentic Coding in the CLI
IQuest-Q1 is an open MoE model with 320 billion parameters, 15 billion of them active. It is trained for agentic coding in the command line and built by merging several experts.
Naive RAG Beat Agentic Memory Systems in Six of Seven Tests
Raw-RAG, which uses dense retrieval and reranking over raw conversation history, took first place in six of seven setups and beat the best agentic memory by 0.13–5.53 points.
AgentVQA Tests 22 VLMs on Agentic Tasks: The Best Scores Just 58.5%
A new offline benchmark of 1000 examples across five domains finds the best of 22 vision-language models answers correctly only 58.5% of the time. The main problem is vision.
Cola2 beats Llama-3-8B on HellaSwag with 43 times fewer tokens
Hierarchical model Cola2 compresses text into continuous latents and scores 60.10 across eight tasks after training on 346B tokens, over 43x less than Llama-3's budget.
Verus-Insight Teaches an AI Video Detector to Find Artifacts and Prove Its Verdicts
Researchers split artifact hunting in generated video into three skills — planning, verification and aggregation — and released the Verus-Insight model and the Verus-Bench benchmark.
AnyGameRenderer Dresses Gray AI Prototypes in Real Graphics and Keeps 62.9% of Units Intact
A generative renderer built on the 33B MiniMax-H3 model turns blockouts from AI-agent games into polished video; temporal entity match rose from 50.9% to 62.9%.
WeChat Open-Sources WeVisOne-VL: Detection, Masks and Matting in a Single VLM
WeChat has open-sourced WeVisOne-VL, a single VLM that finds objects, counts them, reads text and builds masks with alpha mattes. It claims top results on 30+ benchmarks.
DirCast directs and commentates esports on its own, beating GPT-5
A single streaming video-language model picks the camera and talks for the caster in real time. DirCast's commentary beats GPT-4o in 29.2% of comparisons.
UserArena: A 64 729-Session Benchmark Tests Whether Shoppers Can Be Simulated
Researchers collected 64 729 real shopping sessions from 11 010 users and proposed OPTD, a method that reproduces the effect from a live A/B test more closely than other simulators.
Echo Teaches VLA Robots to Recognize Success and Failure Inside Their Own Actions
Researchers embedded stop and failure signals in every action chunk of a VLA model and denoise them faster than the motions. No per-frame progress labels are needed.
HiSTEVE 1.0 Plays Minecraft With a Mouse and Keyboard: 85.0% on MineStudio Bench
VLA agent HiSTEVE 1.0 completes long Minecraft tasks with no API and no hidden game state: 85.0% on MineStudio Bench, 37.4% on OpenHA, and an iron pickaxe from scratch.
Dual Exam: GPT 6 Astra Scores 31.3 Among Multimodal Agents
Over 80 authors built Dual Exam, which tests agents and world models separately. GPT 6 Astra leads with 31.3; a 150-case combined pilot had zero successes.
DISC Learns to Combine Time-Series Similarity Metrics and Predict Transfer Success
DISC, a new learnable framework, calibrates and combines dataset similarity metrics for time series to estimate in advance how well a forecast will transfer from a source to a target.
MMVistaReason: Open Recipe Trains a 4B Model That Beats Qwen3.5-9B on 15 Benchmarks
The MMVistaReason authors released an open post-training recipe: MVR-4B scored 72.8 across 15 multimodal benchmarks, beating Qwen3.5-9B (Instruct).
Survey by Shengqiong Wu and Hao Fei: 600+ Papers on Video Models as World Models
A team including Shuicheng Yan, Ming-Hsuan Yang and Philip Torr unifies video understanding and generation in one framework and reviews more than 600 papers across 19 tables.