Research Papers
S3Gym: New Benchmark Tests Whether LLMs Can Truly Self-Improve
Hugging Face researchers introduce S3Gym, an interactive benchmark that evaluates whether LLM agents can self-test, self-judge, an...
WHALE: A Simple Alternating Recipe for Joint Model-Harness Optimization in AI Agents
Researchers propose WHALE, a method that alternates between updating model weights and searching for better harness code, improvin...
New Benchmark Tests Whether LLMs Can Self-Evolve from Vague Goals
Researchers introduce ASPIRE, a benchmark for self-evolving LLM agents that start with only vague natural-language goals, revealin...
HarnessDev: Benchmarking LLMs on Building and Evolving Their Own Agent Harnesses
A new benchmark shifts evaluation from task outputs to the execution infrastructure agents build themselves, revealing wide variab...
Pixel Linguist II: New Design Principles for Robust Pixel Text Representation Learning
Hugging Face researchers introduce Pixel Linguist II, a vision encoder that learns to read and compress text directly from pixels....
Hugging Face Researchers Unveil Declarative Attention: Letting LLMs Skip Irrelevant Context
A new paper from Hugging Face introduces Declarative Attention, a protocol that lets language models declare which parts of the co...
Hugging Face Researchers Unveil CoGR: Co-Evolving Generative Retrieval with Reinforcement Learning
A new paper from Hugging Face introduces CoGR, a framework that trains LLMs to generate keywords for both queries and items, enabl...
EarlyEval: Cutting Agent Evaluation Costs with Early Outcome Prediction
Hugging Face researchers introduce EarlyEval, a framework that predicts agent success or failure early in a run, reducing evaluati...
Hugging Face Unveils SolarWM: Open Framework for Long-Horizon Video World Models
SolarWM provides a fully open framework and unified training recipe for building interactive video world models, enabling real-tim...
Hugging Face Researchers Unveil DisCo: Distilling GitHub Repos into Reusable AI Skills
A new research agent called DisCo distills operational knowledge from GitHub repositories into reusable skills, boosting autonomou...
Hugging Face Paper Proposes Control-Data Flow Separation for Stable Prompt Optimization in Multi-Agent LLMs
A new framework separates structured execution protocols from optimizable language content, preventing prompt optimization from co...
Hugging Face Paper Proposes Perception-Centered Architecture for Persistent AI Agents
A new research paper introduces Pera, a perception-centered architecture for persistent language agents that continuously adapt se...