Hugging Face Daily Papers
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
GAFU pulls knowledge out of generation in UMMs and lifts VStar-Bench by 20%
The GAFU post-training method compresses a unified multimodal model's generative representations into G-tokens and gains up to +20.00% over SFT on VStar-Bench.
RTAformer Teaches a Model to Notice When Relationships Between Objects Change in Video
Researchers propose RTAformer, a one-stage transformer for video scene graphs. It models object persistence and relationship changes separately, and improves recall on Action Genome.
MolLean Makes LLMs Check Their Chemistry: Wins on 12 of 12 Tasks
MolLean, a new Lean-based language, verifies every step of a model's chemical reasoning. On Qwen3.6-27B it beat plain chain-of-thought on all 12 tasks across 8,800 examples.
PrismaCoder: Open 9B and 35B Models Close In on Proprietary Ones in Visual Code Generation
Researchers release PrismaCoder-9B and 35B, trained on PrismaCoder-2M, the largest open dataset, where a synthesized rubric of criteria scores code quality.
FLOMSS: flow matching splits music into 4 stems and takes 2nd place in listening tests
Researchers applied flow matching to split music into vocals, bass, drums and the rest. FLOMSS-M placed second in a blind listening test, though it trails discriminative systems on SDR.
SliderEdit++ Smooths Out Sliders in AI Image Editors: +18.5% and Up to 20x Faster
Researchers present SliderEdit++ and the SlideBench benchmark. The method smooths the edit trajectory, improves the overall score by 18.5%, and runs 3–20 times faster than alternatives.
JockeyBench: 28 LLM Agent Configurations All Lost Money on Japanese Horse Racing
HKUST researchers tested LLM agents on 324 Japanese horse races. Seven configurations improved their forecasts, but all 28 lost money on average.
KANLab: A TensorFlow Playground Fork That Shows KAN Training in the Browser
Lukas Niehaus and colleagues released KANLab: a small Kolmogorov–Arnold network trains in the browser, and every B-spline curve can be viewed and edited by hand. Open source, Apache-2.0.
CyberArena Pits 12 AI Agents Against Each Other in 466 Real-Time Attack-and-Defense Matches
The new CyberArena benchmark makes two LLM agents hack and patch services at the same time. It covers 12 models, 40 tasks and 466 matches, and it breaks the ceiling of single-agent tests.
BeliefWeaver Adds 5.1 pp to Agents on Long Tasks; 9B Student Adds 9.8
BeliefWeaver helps agents fix errors whose cause lies several steps earlier. A distilled 9B model gained +9.8 pp and cut teacher costs by 54.8%.
DisaggKD moves the teacher onto separate nodes and speeds up distillation of Qwen3.5-35B-A3B into 9B
The DisaggKD framework splits teacher and student across nodes, streams hidden states, and computes the KD loss without full logits. On Qwen3.5-35B-A3B → 9B it runs faster than a colocated setup.
Copy-Paste Language Models Add a Copy Head to LLMs With 135M–1.7B Parameters
Researchers added a lightweight Copy-Paste head to the LM head and tested it on models from 135M to 1.7B parameters: quality improved at every scale at near-zero cost.
LeanPRM: Lean Labels 152,000 Proof Steps Without Humans and Lifts Accuracy to 74.7%
LeanPRM's authors labeled 10,634 proofs through Lean with no annotators. A Qwen3-4B verifier rose from 48.2% to 74.7% step-level accuracy, beating PRM800K.
Researchers audited 6,309 agent skills: half have shell access
An analysis of 6,309 listings from two marketplaces: 50% have shell access, 13% can reach credentials, and 59% of declared permissions understate actual capabilities.
Dream2Discern Predicts Multi-Agent System Failures With Over 90% Accuracy
The Dream2Discern auditor trains on "hard" latent variants of already-recorded trajectories and reaches over 90% accuracy on AFTraj-2K and StepShield without new runs.
TabWM learns to predict database edits: F1 rises from 0.016 to 0.685
From 64 execution examples, TabWM predicts how an action will change the relational database of an unfamiliar system: full edit F1 of 0.685 versus 0.016 with no context.
EvoWeb Tests Coding Agents on 320 Sequential Requests Across 32 Web Apps
The new EvoWeb benchmark makes AI agents build out a single app step by step. The models' main weakness is integrating new features with code they have already written.
DuoDiscover, built on coding agents, speeds up vLLM 1.92x and beats ArchAgent v2
DuoDiscover teaches a fleet of coding agents to invent fast proxy evaluators, and it outperforms humans: vLLM runs 1.92x faster, and a CPU design takes 16.7 days instead of 40.
HACE Co-Trains a Domain Specialist and an LLM Agent Across Three Domains
The HACE framework co-evolves a domain model and an LLM agent as a pair. On the main metrics, it beat self-improvement methods in recommendations, biology and materials science.
Hacker-fixer loop closes holes in the verifiers of Terminal Bench, KernelBench and SETA
A new paper automates finding and fixing vulnerabilities in the verifier code of agentic environments. An early version is already accepted into the official Terminal Bench repository.
UCAS Researchers Teach Weak LLM Agents to Handle Skills: Up to +14.1 Points Without Fine-Tuning
Skills that boost GPT-5.6-luna by 31.1 points drag Gemma-4-E4B down by 8.8. Chinese researchers propose a meta-skill that rewrites instructions to match a model's actual capabilities.
OpenMobile-2 Open-Sources a Mobile Agent: 79.9 on AndroidWorld and 110 Apps
OpenMobile-2 ships with its environment, data and training recipes: 79.9 on AndroidWorld, 50.4 on MobileWorld, and SPA-Bench on real devices up from 31.9 to 59.6.
RealityBench: 26 VLMs Tested on Documenting Physical Work in Blender
A Blender benchmark of 11 environments shows top VLMs read text in frame but struggle with measurements and fine details. At frame acceptance, they match a single rater.
CompactSplat allocates 3D Gaussians by image content, not by pixel
A new feedforward method reconstructs 3D scenes from a few images using fewer Gaussians. Primitive density can be adjusted at inference, with no network retraining.
Git audit of two repositories: AI trailer found in only 1 of 39 commits that touch agent notes
Researchers audited two AI-agent repositories: in Project B, only 2.6% of commits that edited the agent's notes carry a Co-Authored-By trailer. Git-based disclosure is unreliable.
Wasserstein BaryBooster Turns Diffusion Realism into Event Forecasting Accuracy
Researchers built a benchmark for event sequence forecasting and found that no method wins on both accuracy and realism. Wasserstein BaryBooster offers a way out.
LiveMACEBench Tests Five LLM Agents on a Live Market for 30 Days
The new LiveMACEBench benchmark ran five frontier models on live financial markets for 30 days. Its finding: returns are a poor indicator of an agent's real skills.
DScout Speeds Up Qwen3-8B by 4.14× With a Diffusion Drafter
DScout trains a diffusion drafter for prefix acceptance and tunes its configuration per workload: 3.84× on Qwen3-4B and 4.14× on Qwen3-8B versus autoregressive decoding.
SOUL: Researchers Teach VLA Robots to “Unlearn” State Hallucinations
SOUL finds sparse features of a false “already grasped” state in VLA models and erases them without breaking other manipulation skills. Tested in simulation and on real robots.
miniTE speeds up RL in NVFP4 on B300: 1.39× over BF16 on Qwen3-30B-A3B
InfiX.ai has released miniTE, an open framework for 4-bit RL: the learner hits 33,540 tokens/s on a single B300 node, while Qwen3-30B-A3B quality stays close to BF16.