Self-distillation makes coding agents fake hindsight

Training a coding agent on privileged hints, like the correct fix or advice from a stronger model, can leave it worse than doing no training at all, per a paper posted to arXiv on September 26.

The stakes: anyone fine-tuning multi-turn agents on tool use is likely running a version of this setup right now, and the failure mode doesn't announce itself. Untrained baselines beat the trained model on some runs.

The cause is specific. Feed a privileged hint into the loss function and the model learns to act as if it observed something it never saw during the actual task. It's rehearsing an answer key, not learning to search for one.

The paper's fix, Privileged Self-Practice, keeps the hint but relocates it. When most rollouts on a task fail, an analyzer writes a short per-task instruction. The task gets resampled with that instruction in context. Training then runs on unmodified GRPO, the reinforcement learning method underneath, with no privileged information touching the gradient.

Across three student models, this was the only method that consistently beat plain GRPO on both AppWorld and SWE-bench Verified, with resolved-rate gains as high as 61%. The paper doesn't say what the analyzer step costs in extra compute per training round, or whether the gains hold at model sizes it didn't test.

The distinction driving all of this is where the hint lives. In the loss, it teaches imitation. In the sampler, it teaches the model to get itself unstuck.