Research Papers
MiniMax-H3 Scores 41.97% on New Physical-World Reasoning Benchmark
A new evaluation framework tests whether the omni-modal model MiniMax-H3 can reason about physical events when each input modality...
VA-Bench Tests Whether Multimodal LLMs Can Actually Act on What They See
A new benchmark from Hugging Face evaluates the full observe-reason-act-revise loop for embodied spatial intelligence, revealing t...
Qwen3.8-LiveTranslate Cuts Interpretation Lag to 2.3 Seconds, Adds Speaker Separation
Alibaba's Qwen team unveils Qwen3.8-LiveTranslate, a simultaneous interpretation system built on an Interleave audio-text architec...
Hugging Face Releases VākQA, First Telugu Spoken Question Answering Benchmark
VākQA introduces 2,001 Telugu factoid question-answer pairs with 2.53 hours of speech audio, bilingual transcriptions, and human-v...
Srijika Restyles Indic Fonts by Reusing OpenType Layout, Not Rebuilding It
A new Hugging Face paper introduces Srijika, a pipeline that restyles shaping-complete template fonts to produce installable OpenT...
PACT Benchmark Tests Whether Enterprise AI Agents Bend the Rules Under Pressure
A new benchmark called PACT measures how well enterprise LLM assistants follow standing compliance rules when users, managers, or...
Hugging Face Paper: Vision-RL² Cuts Visual Tokens 4x While Boosting Fine-Grained MLLM Accuracy
A new Hugging Face paper shows that localizing a region of interest tolerates 3–4x stronger token compression than recognizing its...
Hugging Face Paper: Supervising Observations, Not Just Actions, Boosts RL Agent Exploration
A new paper introduces ActObs, a fine-tuning method that also supervises environment observation tokens in agent trajectories, yie...
Hugging Face Paper: Candidate Count Alone Misstates LLM Test-Time Scaling Costs
A new paper accepted to EESP@SC26 argues that the inference budget N only says how many candidates are generated, not how they are...
Hugging Face Paper Proposes SELF-INDEX, a Self-Evolving Search Index for LLM Agents
A new Hugging Face paper introduces SELF-INDEX, a framework that lets a retrieval index diagnose its own failures and revise its k...
Video DeltaNet Cuts Livestream Video Generation to 6.7 Seconds on Eight B200 GPUs
Hugging Face paper introduces Video DeltaNet (VDN), a hybrid attention design pairing local Softmax attention with bidirectional l...
WeVisDoc Introduces Two-Stage Data-Centric Framework for Robust Document Parsing
Hugging Face paper WeVisDoc proposes a two-stage, data-centric approach that first broadens document coverage and then uses diagno...