News Around the World — 2026-10-15

AI

• Hugging Face Papers — Player-1: Learning to Play from Synthetic Experience — PPO training in 16 synthetic game environments with programmatic rewards beats supervised fine-tuning on 6,044 gameplay trajectories, and the gains transfer to unseen games: on MineDojo-300, Player-1-9B’s success rate rises from 10.54% (after SFT) to 38.33%, above the strongest baseline’s 12.63%.
• Hugging Face Papers — Recovering Reliable Supervision via Critique-Guided On-Policy Distillation — CriticOPD has the teacher critique the student’s wrong answers, then scores the student’s rewrite with that critique in context, lifting average accuracy across six math benchmarks by 4.57 points for Qwen3-1.7B (50.08 to 54.65) and 4.06 points for Qwen3-0.6B (34.80 to 38.86) over standard on-policy distillation.
• Hugging Face Papers — Doc2Harness: Benchmarking LLMs on Long-Horizon Harness Engineering of Coding Agents — A deterministic benchmark for building complete coding-agent harnesses from a dev document and an interface-only skeleton: across 19 models and 328 test cases, the top model (GPT-6 Astra) passes only 53.05%, the weakest passes 27.44% — even the strongest model fails nearly half the tests.

</content>