RUC's Agent-Reflex trains reflective agents in three stages, scoring 58.6 on long-horizon tool-use benchmarks
RUC-NLPIR · hf · 2026-10-14
RUC NLPIR proposes Agent-Reflex, a general training framework for proactive agent reflection under noisy MCP tool feedback and user interventions:
- RSD: injects tool-feedback perturbations, mines success-failure pairs from a diverse LLM pool, and distills reusable reflection meta-skills.
- ReSFT: converts skills into supervision for self-reflection, skill recovery, and user-centric correction.
- ORPO: combines agentic RL under noisy interactions with reflection-guided on-policy distillation.
Built on Qwen3-8B, it reaches 58.6 overall on five long-horizon tool-use benchmarks and 29.1 on AgentNoiseBench, beating environment-scaling and reflection baselines. Code is open-sourced.
More from coding & agent
- Airbnb case study: portable agent skill hits 99.2% procedural accuracy in 1-2 engineer days — _reachsumit · 2026-10-14
- Ship faster sites: have Opus send 4 subagents to audit 20 performance bottlenecks before launch — MartinGTobias · 2026-10-14
- Claude Code makes a SpaceX documentary video entirely via chat, no editing software — xiaohu · 2026-10-14
- Copilot Cowork review: a year behind rivals, riddled with UX gaps — koltregaskes · 2026-10-14
- Agent told to change one class time rewrote the waiver and reverted prices — StormController · 2026-10-14
- Open-source d2r-mobile runs Switch games natively on iPhone at 60fps — dotey · 2026-10-14