NVIDIA's FC-SWE: Failure-Conditioned RL Boosts Multi-Attempt Resolution for SWE Agents
nvidia · hf · 2026-10-19
NVIDIA proposes FC-SWE, failure-conditioned RL for long-horizon software engineering agents. Where GRPO stops after a failed attempt, FC-SWE resets the environment and feeds the failed patch plus verifier output as failure-context evidence into the next attempt, with rewards composed across attempts (e.g. (0,1) for fail-then-pass). Conditioning recovery on failure feedback boosts resolution rates across multiple attempts on several benchmarks.
More from coding & agent
- DecisionAct: open-source Python agent framework using decision models instead of LLM replies — AgitatedReveal07 · 2026-10-19
- 2-GPU R9700 vLLM rig hard power-offs: tensor-parallel all-reduce transient, not wattage, kills 1200W PSUs — brainchillzZ · 2026-10-19
- Nadella: Windows being redesigned as a safer OS to run AI agents — wateriscoding · 2026-10-19
- Krea2 vs MOSS vs Qwen2.1 vs Flux3 hands-on: Flux3 flops hard — CaptainBackwoods447 · 2026-10-19
- AI performance engineering meetup: GPU rooflines, agent trace-to-training loop, Homa transport — cfregly · 2026-10-19
- Asimov open-sources Python SDK and Edge control layer for its humanoid robot — freelerobot · 2026-10-19