ROSS: Relearning from Self-Generated Rollouts through Selective Supervision

Open paper
Benchmark Model Rank Results
code-generation-on-swe-bench-verifiedROSS (Qwen3.6-35B-A3B) (Codex CLI)–Accuracy: 68.4
instruction-following-on-ifbenchROSS (Qwen3.6-35B-A3B)–Accuracy: 47.76

Rank counts only results with a code link.