Hugging Face Daily Papers
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
OfficeTown Simulates 23 Organizations and Trains Agents for Office Work on 406 Tasks
Researchers built OfficeTown, a generative simulation of offices that produced a benchmark of 406 tasks and 4,836 criteria. Training on its data noticeably improved Qwen3.5-9B on GDPval.
MARS Transfers Text-Based Refusal to Video and Images Without Data: A Test on 5 MLLMs
MARS extracts a refusal direction from an LLM's text backbone and adds safety to multimodal models with no training and no unsafe images. Tested on five MLLMs.
Conditional Program shrinks DeepSeek's n-gram memory from 203 to 51 GB at 2 bits
Researchers propose Conditional Program: 2-bit compression of n-gram memory tables cuts KL error by 41–63% versus round-to-nearest and shrinks DeepSeek's table fourfold.
ERB: AI Agents Build Their Own Model Tests and Match the Reference Benchmarks in 75% of Pairs
The EvalResearchBench benchmark tested 9 research agents. Their evaluators rank models the same way as the target benchmarks in 75% of pairs, against a ceiling of 91%.
MambaAct speeds up π0.5 robots: action expert 10x smaller, latency down 64.8%
MambaAct replaces the transformer action expert in π0.5 with a compact Mamba block. It is 10x smaller, cuts denoising latency by 64.8% and raises success rates in LIBERO.
LM2LM: Qwen3.8-27B Beats GPT-6-sol at Ideas for Improving LLMs and Is 33% Cheaper
Researchers used RL to train an LLM to propose improvements for language models: Qwen3.8-27B beat GPT-6-sol on held-out tasks and cut data-generation costs by 33%.
RoboHarness: A Coding Agent on Qwen3.8-Flash-Next Beats VLA Models in BEHAVIOR
Researchers at Zhejiang University have unveiled RoboHarness, a visual-geometric control panel that lets an LLM agent outperform VLA models on BEHAVIOR Challenge 2025 tasks.
New Survey Maps AI in Humanitarian Work: Six Tasks, From Weather to Refugees
A survey on Hugging Face Daily Papers describes AI as a tool for six humanitarian tasks: weather forecasting, energy, forests, poverty, post-disaster damage and migration.
ARC-SORT hits 81.82 HOTA on SportsMOT and runs at 50 fps
The training-free ARC-SORT tracker scores 81.82 HOTA on SportsMOT with the fewest ID switches among 69 entries, and runs at about 50 fps on a laptop RTX 3060.
SciΠvot benchmark: the best LLM picks the right plan 63.5% of the time, versus 73.4% for humans
A benchmark of 248 senior-researcher corrections finds the best model picks the right call 63.5% of the time; fresh researchers hit 73.4%. Project context tends to hurt models more than help.
GlobeBench: The Best LLM Correctly Predicts Environment Responses in Only 49.57% of Tasks
The 1632-task GlobeBench found that the best of 20 models, GPT-6 Astra, simulates the environment's response correctly in only 49.57% of cases. No model solved 631 tasks.
A Frozen 16 MB Vocabulary Matches a Trained One in Small LLMs
Researchers replaced 168M embedding parameters in a Gemma-like model with a frozen 16 MB table and did not lose to the trainable version on three of four benchmarks.
LIBERO-Industry tests 5 robot policies on 216 factory tasks
The new LIBERO-Industry benchmark adds a conveyor, narrow action windows, and tight tolerances. Five generalist policies lose performance as conditions get harder.
LLM-Attacker Benchmarks Split-LLM Attacks: Gradients Raise Leakage to 37.77%
The LLM-Attacker framework standardizes privacy testing for split LLMs. On GPT-2, server-side gradients raised exact document recovery from 13.71% to 37.77%.
Data Rewriting Beats DP-SGD on the Privacy-Quality Trade-Off
A controlled comparison on synthetic medical records with "canaries": fine-tuning leaks the most, DP-SGD hurts quality, and private rewriting offers the best trade-off.
Kolab: 115 Hours of Recordings Show That AI Agents Don't Understand Their Human Colleagues
Minnesota NLP researchers recorded 50 professionals from 40 occupations working with AI agents. Frontier models overestimate how well that collaboration goes.
DialFork measured procedural hallucinations in agents across 36,278 dialogues
The DialFork benchmark catches agents that sound smooth and factually correct but drift off script. It includes 36,278 dialogues, 24 tasks, 12 domains, and two new metrics: FlowKL and FlowScore.
AlphaBet trains an LLM agent to forecast events: Brier score improves 14.4%
Researchers trained GLM-4.5-Air to search for data and give probabilities across several dates. Brier on Forecast-Dojo fell from 0.731 to 0.626, on FutureX from 0.612 to 0.528.
RealCLI: The Best Terminal AI Agent Scores 49% on a Test of Working With a Human
The new RealCLI benchmark checks not only whether an agent solved a terminal task, but also how it works with the user. The best of 33 configurations scored 49.2%, and only 37.6% on the joint metric.
SAE-IDF Matches Symbols Across Different Systems and Beats a Baseline by Up to 45 Points
SAE-IDF uses sparse autoencoders to match elements across heterogeneous symbolic systems. It beats a dense baseline on 12 tasks with Gemma-2-9B, winning by up to 45 points.
ReaDy speeds up hybrid LLM decoding 3.16x on a single B200
ReaDy feeds the recurrent states of hybrid linear models into a speculative drafter, averaging a 3.16x speedup on Qwen3.8-27B and a single B200.
DAGFormer Plans Connections Between Layers and Cuts Perplexity by 17–25%
DAGFormer adds direct cross-layer connections to the transformer and plans them with a topology predictor. On models up to 1B, WikiText-2 perplexity drops by 17–25%.
Researchers Find a 0.4–1.1% Window in Pythia Pretraining Where the Readout Reorganizes
A model's hidden states encode a task before its output shows it. The gap traces back to the unembedding matrix, which reorganizes during 0.4–1.1% of Pythia's pretraining.
semCER Weights OCR Errors by Severity: 61M LLM-Judge Ratings
The new semCER metric keeps CER's alignment but has an LLM judge score every recognition error. The authors released 61M rated edits across 16 languages.
SAE on Gemma-2-2B: Share of Matched Features Rises from 2% to 29% With 50M Tokens
Experiments on four SAE architectures and three models: reconstruction plateaus at small budgets, while feature reproducibility keeps growing with data, from 2% at 5M tokens to 29% at 50M.
REASON teaches a model to think in two loops and saves about half the compute
REASON splits inference into a fast loop and a slow loop: the system gathers extra data only where uncertainty remains, and spends about half the compute.
MULTIFINGROUND: LLMs Recognize Accounting Concepts With 99% Accuracy but Miss Facts
A new benchmark on filings from 90 issuers: models identify the concept with up to 99.4% accuracy, yet in 26.6–41.9% of cases they miss at least one related number.
Interlock teaches agents to finish the job: up to +6.64 points on JobBench
The Interlock framework synthesizes long-horizon tasks for LLM agents. Fine-tuning two Qwen3.6 models on their own trajectories gave +4.00–4.50 pp on AutomationBench.
BlenderWorld: GPT-6-Astra Beats Opus-5 in 3D Benchmark with a Score of 79.9
The new BlenderWorld benchmark ran 13 frontier models through 223 3D creation tasks in Blender. GPT-6-Astra leads with 79.9, but spatial coherence still lags.
WinARDet fixes looping in MLLM detectors: F1 rises from 32.16 to 39.36
The WinAR window-compression method stops multimodal models from repeating one box endlessly: on the dense RSDense benchmark, F1 rose to 39.36 and duplicates fell from 35.48% to 13.00%.