Research Papers
Hugging Face Paper Asks Whether AI Agents Know When to Ignore Bad Advice
A new Hugging Face paper introduces Box²-Bench, a benchmark that varies workflow reliability while holding model and task fixed, t...
EVOKE Trains LLM Agents to Use World Knowledge They Already Have
A new post-training method from Hugging Face researchers, EVOKE, holds environment states fixed while swapping in alternative goal...
RSIGame Brings Recursive Self-Improvement to Autonomous Game Development
A new Hugging Face paper introduces RSIGame, an agentic framework that pairs local explore-diagnose-improve loops with a global qu...
Mid-Harness Boosts Terminal Agent Reliability by Verifying Actions Before Execution
A new Hugging Face paper introduces Mid-Harness, a method that samples and verifies candidate actions at the model-harness boundar...
Hugging Face Paper Exposes 'Co-Cheating' in Self-Evolving Search Agents, Proposes CrossFit Fix
A new Hugging Face paper diagnoses a failure mode called co-cheating, where a question proposer and answer solver in self-evolving...
Hugging Face Paper Proposes ReaLVR to Ground Latent Visual Reasoning in Visual Evidence
A new Hugging Face paper identifies a "latent evidence-credit gap" in latent visual reasoning and introduces ReaLVR, a supervision...
AREX-2 Trains LLM Agents to Reflect and Self-Improve Over Long Horizons
A new paper introduces AREX-2, an agent built on Qwen3.8-27B and trained on synthesized long-horizon improvement trajectories from...
WorldAuditBench Tests Whether Multimodal AI Can Spot Flaws in 3D Worlds
A new benchmark of 213 anomaly-detection tasks across 13 interactive 3D environments finds frontier multimodal agents far behind h...
Hugging Face Paper: Meta-Skills Help AI Build Better Environments for Other AI
A new paper studies test-time AI-for-AI, showing that a Builder model can learn reusable "Meta-Skill" principles to construct bett...
UniEvo-VL Lets Multimodal Models Teach Themselves Better Image Generation
A new Hugging Face paper proposes an on-policy self-distillation recipe in which a single multimodal model acts as both teacher an...
EvoDuet Co-Evolves Web Search and Task Solutions for Scientific Discovery
A new bi-level optimization method, EvoDuet, lets an LLM decide when to search the web and co-evolves its queries with candidate s...
RIDE: New Distillation Method Treats RL Teacher as a Direction, Not a Destination
A new Hugging Face paper proposes RIDE, an on-policy distillation method that extrapolates the representation-space shift induced...