Researchers report that a coding model solved more software-engineering benchmark tasks after its own post-task reviews became training examples. A Sept. 28 preprint tests whether those explanations can change later behavior through updated model weights, without showing the model the reviews again at test time.
Retrospection-Only Fine-Tuning, or ROFT, has an agent attempt a coding task and observe available feedback, including test results. The same model then writes an explanation of a consequential assumption or decision and what evidence suggests should change. During training, the model learns to produce the review, using its attempt and observations as context; it is not trained to reproduce the attempt or feedback.
On SWE-bench Verified, a human-validated set of 500 software issue tasks, the 9-billion-parameter Qwen3.5-9B model solved 294 (58.8%) after 20 ROFT updates. A reinforcement-learning method called GRPO solved 279 (55.8%) after 40 updates. ROFT's score is from 20 training updates and GRPO's from 40, so this is not an equal-update comparison. Both runs used the same base model, training set and allocation of four training and four inference B200 GPUs; measured training times were 1.93 hours for ROFT and 5.33 hours for GRPO.
The evidence is a preprint focused on software engineering, with experiments on Qwen3.5-4B and Qwen3.5-9B and no independent replication. The authors plan to release selected code components after acceptance. They also caution that model-written reviews can reinforce mistaken explanations and should not be treated as faithful accounts of a model's reasoning; tests across more models and domains remain open.
Sources
- Researchers' preprint and methods, posted September 28, 2026
- SWE-bench Verified benchmark description
- Qwen3.5-9B model card
- Lead figure: Figure 9(c) from the preprint, licensed CC BY 4.0.