AI
ROSS: Relearning from Self-Generated Rollouts through Selective Supervision
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 27/09/2026, 20:00:00 UTC
Briefing created: 30/09/2026, 04:03:51 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 30/09/2026, 03:18:25 UTC.
Briefings
ROSS (Relearning from Self-Generated Rollouts through Selective Supervision) improves reinforcement learning performance by preserving historical rollouts while applying loss only to selected model continuations. On Qwen3.6-35B-A3B, ROSS increased the six-benchmark MOPD average from 58.40% to 62.20% and SWE-bench Verified from 64.20% to 68.40%. This method outperforms baselines across mathematics, code generation, instruction following, and software engineering, demonstrating that self-rollout training yields further gains through offline supervised fine-tuning.
Key points
- ROSS improves reinforcement learning performance by preserving historical rollouts while applying loss only to selected model continuations.
- On Qwen3.6-35B-A3B, ROSS increased the six-benchmark MOPD average from 58.40% to 62.20% and SWE-bench Verified from 64.20% to 68.40%.
Context and limitations
Results are limited to specific domain-specific reinforcement learning tasks and Qwen3.6-35B-A3B model; findings may not generalize to other architectures or tasks.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 30/09/2026, 04:03:51 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeify5zadn6qxu5ota5dh6upkqe65tts2622invdbwswiyazqmff6di