Research Papers
Video LLMs Are Still Costly: Survey Maps 125 Papers on Inference Efficiency
A new survey reviews inference-efficiency mechanisms for video and audiovisual large language models, organizing 125 papers by pip...
HybridAL Cuts Active Learning Training Time by Up to 49% by Switching Retraining to Fine-Tuning
A new EMNLP 2026 paper introduces HybridAL, an adaptive schedule that monitors online stabilization signals to decide when active...
OracleZoom Brings Reference-Constrained Recursive Super-Resolution to Extreme Zoom Levels
A new Hugging Face paper introduces OracleZoom, an on-policy self-distillation-inspired framework that keeps recursive super-resol...
Φ-Bench Tests Whether LLMs Can Engineer the Infrastructure That Runs Them
A new benchmark from Hugging Face–hosted research evaluates large language models on open-ended, long-horizon engineering of the L...
Hugging Face Paper Challenges Full Reasoning Traces in LLM Post-Training
A new paper argues that large language models benefit more from truncated reasoning trajectory endpoints than from complete traces...
PARSER Decouples Reading From Reasoning to Fix Long-Context LLM Agents
A new Hugging Face paper introduces PARSER, a scatter-gather agent architecture that reads document chunks in parallel while a lea...
SyncWorld Uses Visual Calibration to Turn World Models Into Zero-Shot Robot Simulators
A new Hugging Face paper introduces SyncWorld, an action-conditioned world model that uses visual calibration episodes to map robo...
Puppeteer: Hugging Face Paper Grounds Co-Speech Gestures in Objects and Posture
A new diffusion-based model called Puppeteer generates co-speech gestures in a causal latent space, conditioning on speech, motion...
SWE-Bench Pro Verified Targets Reward Hacking and Task Flaws in Coding Agent Evaluation
A new verified version of the SWE-Bench Pro benchmark aims to remove reward hacking and task-quality problems that may have inflat...
Hugging Face Paper Unveils T1, a 122B RL-Trained Terminal Agent for Long-Horizon Tasks
T1 is a 122B Mixture-of-Experts model trained with reinforcement learning to operate a real shell in a cloud sandbox for 300+ tool...
AgentGrad Uses Sequential Intervention to Sharpen Multi-Agent Prompt Optimization
A new Hugging Face paper proposes AgentGrad, a framework that identifies the responsible agent through sequential intervention and...
Hugging Face Paper Proposes World Model RL to Speed Up Autonomous Research Agents
A new paper introduces World Model RL (WMRL), which swaps costly sandbox execution for a learned world model and adds debiasing an...