Hugging Face Daily Papers
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
SpatialBrain Teaches a Frozen VLM to Accumulate Experience, Winning on 4 of 5 Benchmarks
SpatialBrain keeps VLM weights frozen and adds 3D memory, a skill library and accumulated experience. It beats direct answers on four of five benchmarks and ties on the fifth.
DONIC teaches one policy to control two hands: 86.79% success without fine-tuning
Researchers unveiled DONIC, the first general-purpose tracker for two dexterous hands. A single checkpoint hits 86.79% success on new trajectories and 57.32% on unseen objects.
TraceForge trains models on "their own thoughts": +7.1 points on MBPP+
TraceForge writes reasoning for a closed model's answers in the student family's style. HumanEval+ gained 3.7 points, MBPP+ 7.1, and a judge scored the traces 0.68.
Researchers Propose PAPAV Framework: Only 3 of 64 Benchmarks Test Verification
A new paper compares computer-use agents and robots through five PAPAV functions. Of 64 benchmarks, 5 assess outcome prediction and only 3 assess result verification.
NUS Teaches a Model to Find What It Described: gIoU Rises From 57.5% to 68.2%
Generalized CycleGRPO links describing a region and finding it from text in a single RL loop, with no reference captions and no LLM judge. On GroundingSuite, gIoU rose from 57.5% to 68.2%.
TenetBench: 11 of 14 LLM Configurations Violate Professional Ethics Codes
A new Russian-language benchmark tests models against specific clauses of professional ethics codes. 11 of 14 configurations produce conflicting content on most prompts, usually without a warning.
DetectorScope Turns the Black Box of AI-Text Detectors into a Transparent One
Google DeepMind researchers taught AI-text detectors to explain their decisions: sparse autoencoders pull human-readable features out of the model with no loss of accuracy and a gain on new data.
LLM2Cert verifies 81 of 105 stability and safety certificates
Researchers introduce LLM2Cert and the DynaCertBench benchmark: an LLM proposes Lyapunov functions and barrier certificates, and SMT solvers and Lean check them. Result: 77.1% of tasks.
Peking University speeds up video loading for AI by 1.45x
Academician Wen Gao's team releases codec4AI: storing video for random access cuts data by 58.7% and boosts loading speed 1.45x versus LeRobot v3.
CausalAlign cures video models' "forgetfulness": geometry errors drop 20%
Researchers pinpointed where diffusion video models lose track of camera motion and aligned them with a 3D model. Wan2.2's geometric errors fell by more than 20%, and robots got 10 points more successful.
LEGO-SCD learns to detect changes in 3D scenes at 17.2 FPS without manual labeling
LEGO-SCD trains on automatic edits to 3D Gaussian Splatting scenes and posts top results: 67.64% mIoU on PASLCD, 75.82% on CCS3D, and 17.2 FPS.
HyperDrone Renders Millions of Frames per Second and Outpaces Rivals by 20x
HyperDrone, a new GPU ray-traced quadcopter simulator, runs more than 20 times faster than comparable tools and transfers a policy from simulation to a real Crazyflie.
RLSplat teaches 3D Gaussian Splatting to economize: up to 4.1× faster and 9.6× less memory
RLSplat generates exactly as many Gaussians as the budget specifies, and spends proportionally less time and memory doing it: up to 4.1× faster and up to 9.6× lower peak memory.
Stratified Sparse Attention speeds up decoding 2.96x at 1M context
The SSA sparse attention method combines deterministic top-k with importance-weighted sampling of the remainder. At 50x sparsity on 128K, it retains 85% of dense-attention quality.
StyleStream 2.0 Cuts Voice Conversion Latency in Half, to 520 ms
The second version of the streaming voice conversion system runs with 520 ms latency on a consumer GPU. It also learns to create and edit voices from text descriptions.
Diffusion LLM Agents Learn to Whisper in Vectors: +34 Points on QASC
Researchers had diffusion language models exchange four hidden vectors instead of text. Accuracy on QASC rose to 80.8%, and a single late handoff reached 87.8%.
RoboCasa-Shift: a benchmark shows where VLA robots break in the kitchen
The new RoboCasa-Shift benchmark ran GR00T N1.5 through 1 040 rollouts: two failures with the same ΔSR=0.188 break the robot at the first stage in 19.0% and 36.4% of cases.
WorldStream Builds an Endless 3D World and Remembers Every Player Edit
WorldStream extends a playable 3D world beyond the map, executes language commands, and restores visited places with their edits. It plans on two levels; deterministic code builds the geometry.
Qwen-Audio-3.1-ASR: 30 Languages, 16 Dialects, and Diarization in a Single Model
Qwen releases Qwen-Audio-3.1-ASR: 30 languages, 16 Chinese dialects, diarization with speaker labels, and audio understanding, from music to time-stamped event detection.
A Team of AI Researchers Found a Linear Attention Design That Beat GDN2 at 1.3B
A team of persistent AI agents with different scientific backgrounds searched for linear attention architectures. At 1.3B parameters, it beat GDN2 on language, reasoning and retrieval.
DiffResearch: A Diffusion LLM Makes Deep Research About 6x Faster
The open-source DiffResearch framework swaps an autoregressive model for a diffusion one and speeds up deep research roughly 6x. Its dFast mode adds another 1.99x on query decomposition.
Of 1 598 RL Papers, 37 Made It Into Frameworks: An Audit of Seven Open-Source Stacks
An audit of commits in seven open-source RL post-training frameworks: of 1 598 methods, only 37 reached code, while paper volume grew fivefold against the same 5–7 inclusions.
Researchers Ran 120 CPT Configurations for Game Agents, From 0.8B to 27B
A controlled study tests whether scaling works for game agents across different games: 120 runs, 50 to 1 600 hours of data, and models up to 27B parameters.
Researchers teach LLMs to adapt to dialect: the CAP method and 10 model families
An audit of 10 open model families found DPO rewards British and Australian spelling more than AAE and Singlish. The CAP method evens this out without hurting standard English.
Agent4Sci: Claude Opus Beats Records on 20 Scientific Benchmarks by 10.4%
The Agent4Sci testbed shows frontier agents beat published records on all 20 scientific ML benchmarks, but they win through engineering tricks, not scientific ideas.
MultiAgent-OPD Trains a Whole Team of Agents End to End and Scores 38.68% on Seven QA Benchmarks
TaoLive AIGC researchers combine on-policy distillation with RL for a Leader–Searcher–Answerer hierarchy. Result: 38.68% macro avg@4, 4.06 pp above Prompted Critic.
DuoLM Speeds Up LLM Recommendations 60x: 26,400 Scores per Second on a Single H100
Researchers split a recommendation LLM into an encoder and a ranker, cached item descriptions, and reached 26,400 candidate scores per second on one H100 at teacher-level quality.
RoboJudge-9B Judges Robot Videos More Accurately Than Gemini: 0.719 Correlation With Humans
CUHK researchers built an 800-video annotated benchmark and trained a 9B judge for robot videos: 0.719 correlation with humans, 0.089 higher than Gemini-3.7-Flash.
DeCert Teaches AI Judges to Explain Their Verdicts: An 8B Model Scores 76.2 on RM-Bench
The DeCert framework trains generative reward models not just to pick the right answer but to justify it. Qwen3-8B scores 76.2 on RM-Bench, with up to +14.8 in position consistency.
0.8B-Parameter OvisOCR3 Outperforms OCR Models: 96.74 on OmniDocBench v1.6
The ATH-MaaS team has released OvisOCR3, a 0.8B model that handles five OCR tasks at once. It scored 96.74 on OmniDocBench v1.6 and took first place on OCRBench v2.