SFT-then-RL Outperforms Mixed-Policy Methods for LLM Reasoning
Mixed-policy LLM reasoning gains stem from buggy baselines; fixing optimizer and loss bugs makes standard SFT-then-RL outperform them by up to 22 points.
Published 2026Sydney Poster Session 1 · Tue, Dec 8, 10:00 AM–1:00 PM local time · Hall 1-4arXiv ↗OpenReview ↗
Only vote on papers you've read. Sign in with GitHub to vote.
Abstract
Recent mixed-policy optimization methods for LLM reasoning that interleave or blend supervised and reinforcement learning signals report improvements over the standard SFT-then-RL pipeline. We show that numerous recently published research papers rely on a faulty baseline caused by two distinct bugs: a CPU-offloaded optimizer bug in DeepSpeed that silently drops intermediate micro-batches during gradient accumulation (affecting multiple downstream frameworks including TRL, OpenRLHF and Llama-Factory), and a loss aggregation bug in OpenRLHF that incorrectly weights per-mini-batch losses. Together they suppress SFT performance, with the optimizer bug accounting for most of the gap and the loss aggregation bug contributing a smaller additional effect. Once corrected, the standard SFT-then-RL pipeline surpasses every published mixed-policy method we evaluate by +3.8 points on math benchmarks with Qwen2.5-Math-7B and by +22.2 points with Llama-3.1-8B. Even a truncated variant with just 50 RL steps outperforms mixed-policy methods on math benchmarks while using fewer FLOPs.