Hugging Face Daily Papers
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
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.
LeWorldModel Tested for Action Sensitivity: Axes Diverge by 36–42°
Preprint authors compared how LeWorldModel and the simulator respond to small action changes. The principal sensitivity axes diverge by 36° in Reacher and 42° in OGBench Cube.
FocusVLA Teaches a Robot to Pick a Camera Based on Task Stage: +13.3 pp in Real-World Tests
Chinese researchers at ISCAS taught a VLA model to switch between global and wrist views based on predicted intent. On RoboTwin 2.0, it gained 10.10 pp over π0.5.
UCLA Team Pushes the Speed–Quality Frontier in Speculative Diffusion With Two Knobs
UCLA researchers tune two parameters at once: diffusion effort and how leniently the verifier accepts tokens. Together they extend the Pareto frontier beyond single-step diffusion drafting.
PaperFormalizer Turns LaTeX Papers into Verified Lean Code, with the Author in the Loop
Stanford researchers released PaperFormalizer, an agentic pipeline for Claude Code or Codex that finds gaps in proofs and formalizes a paper in Lean.
NVIDIA open-sources Flora: robot data labeling 125 times cheaper than GPT-6 Astra
NVIDIA has open-sourced Flora, the labeling engine for GR00T. It labels 200 000 hours of data a day on 512 GB300 and costs about 125 times less than GPT-6 Astra.
G-Retriever's Graph Token Turns Out to Be a Dud: Cosine Similarity of 0.985–0.998
Researchers at Sapienza show that G-Retriever's "graph token" is nearly identical for any input, and a constant with no access to the graph matches the method's accuracy.
OPTIC compresses Gemma 4 E4B's per-layer embeddings 100x while keeping 97.8% of quality
Researchers at VILA-Lab have shrunk Gemma 4 E4B's per-layer embeddings roughly 100-fold: the model retains 97.8% of its quality across seven benchmarks while GPU memory usage drops 29.7%.
HiP3R assembles 3D scenes from thousands of video frames without retraining models
Researchers at NTU Singapore proposed a hierarchical pose graph: video chunks are first aligned in Sim(3), then each frame gets a single shared SE(3) camera.
Nestor: nested sparsity in attention beats MoE at equal compute
Tianzhe Chu, Druv Pai, and Yi Ma propose the Nestor transformer: attention sparsity nests inside MoE, delivering gains at equal compute and in continual pretraining.
A 17-million-parameter encoder picks the right tool in 14ms on CPU
Tardelli Stekel has released six open encoders ranging from 17M to 1B parameters for closed-set selection. The 68M and 150M models beat the 435M DeBERTa-v3 by 4.9 and 5.9 points.
DecisionPivot Tests 40 Solver Models: Up to a 26-Point Gap at Equal Accuracy
DecisionPivot's 40,033-item benchmark shows two 27B models 2.6 points apart on the reference track diverge by 25.9 points on contrastive groups.
S-Agent 2.0 teaches a coding agent to understand 3D and beats SpatialClaw by 7.8 points
A ready-made coding agent with a set of 3D tools lifted Qwen3.8-27B to 66.5 on MMSI-Bench and 82.3 on ReVSI. Gemini-3.8-Flash reached 74.1 and 90.5.
RiseHarness Teaches Terminal Agents to Fix Their Own Harness Without Fine-Tuning
RiseHarness iteratively improves the harness of terminal agents with frozen weights. It compares successful and failed runs of the same task and validates each change with repeated runs.
Lafa, an 8B-parameter model, speeds up text generation by 2.54x without sacrificing quality
Researchers have introduced Lafa, an 8-billion-parameter streaming generation model that generates text 2.54 times faster than conventional models without losing quality. A blockwise architecture with bidirectional async
EvaOmni Benchmark Tests 24 Omni Models: Leader Scores 79.4% Accuracy, but Only 35.1% on Grounding
EvaOmni requires models to give an answer plus the time intervals of their evidence. Leader Gemini3.7-Flash hit 79.4% accuracy, but its B.G. fell to 12.3% on the hardest tasks.
Video-Point Model Predicts Multi-Camera Video and 3D Point Trajectories
Video-Point Model predicts future video from any viewpoint, plus 3D trajectories of scene points. Point supervision improves video quality and robot manipulation success.
TARSphere Tests MLLM-Powered Robots Across 360 Episodes in 30 Scenes
The TARSphere framework runs multimodal LLM agents in a physics simulator with NPC humans: 360 episodes, 30 scenes, 8 models, and frequent collisions and falls.
Agent Mosaic: LLM Agents Assemble Secrets From Harmless Documents in 7% of Responses
Researchers tested 12 240 agent responses: even GPT-5.6-Sol gives forbidden conclusions in 7.08% of cases, and rules plus classification labels still leave 32.37% unsafe.
Layer Looping Cuts Diffusion LLM Training Costs by Up to 3.3x
Looping early transformer layers at every denoising step cut training costs for uniform diffusion language models by up to 2.8x on OpenWebText and 3.3x on FineWeb-Edu.
FrontierCS 2 tested 12 AI agents on R&D tasks: the best passed 58%
A 48-task benchmark cuts the contribution out of real papers and asks agents to rebuild it. Claude Opus 5.5 leads with 58%, while no run passed 16 of the tasks.
SplatLab writes its own 3D Gaussian Splatting papers and beats baselines by up to +3.78 dB
An autonomous system reproduces a 3DGS paper, proposes an extension, trains models, and writes a manuscript. PSNR gains over baselines range from +0.18 to +3.78 dB.
Survey: Robots Are Learning to Improve Their Own Learning, but Evidence Is Still Missing
A new survey carries recursive self-improvement from code into the physical world and names four conditions under which robot gains accumulate instead of vanishing.
ToM2Vec Teaches LLM Agents to Model Their Partners in Latent Tokens
ToM2Vec builds a partner's theory of mind in continuous tokens and wins in Hanabi, Overcooked-AI and negotiation, including games with an unfamiliar GPT-5.4.
Quant-Arena: None of 9 LLMs Beat the A-Share Market in 2024
A new benchmark tested nine frontier models on the full quant research cycle. GPT-6 Astra ranked first, but no strategy beat the equal-weighted market on the sealed 2024 year.
LASAGNA Tests 450 RAG Configurations: The Promised +47% Gain Turns Into Just +17%
The open-source LASAGNA framework ran 450 RAG configurations on eight benchmarks. Three strong retrieval techniques promised +47% individually but delivered only +17% combined.
LoopRec Picks Compute Depth for Each Item and Beats Baselines on 4 Backbones
LoopRec makes a recommender model linger only on the hard items in a user's history. It uses one shared block and an exit gate on every loop, and it improves results across four backbone families.
RealMe Builds 2K Volumetric Video from a Handful of Cameras Using Diffusion
RealMe turns sparse low-resolution captures into 4D-consistent 2K volumetric video, using a video diffusion model and a latent-space frequency loss.
PACE compresses linear LLM state to 4 bits and saves a third of the memory
PACE quantizes the recurrent state of hybrid LLMs to four bits without retraining: on 4B–35B models it matches 8-bit DAMP while using about 2/3 of the memory.
SERA teaches a VLM planner to self-evolve: robot scores 51.38% on RoboMME
Researchers turned a VLM planner into a robotics agent with an editable harness. With self-evolution, SERA hits 51.38% success on RoboMME versus 44.38% without it.