Perplexity Trains Computer Agent on Real Mistakes With Hint-Guided Self-Distillation

The post-trained weights and training code were not released. The model runs only as a model option inside Perplexity Computer. The base model, GLM 5.2, is openly available on Hugging Face.

Standard rejection sampling fine-tuning (RFT) judges each session and imitates only the successful ones. A successful outcome does not mean every step was correct. An agent can recover from a bad tool call and still deliver the right answer. Imitating that full trajectory can reinforce the error. Discarding failed sessions also throws away clear evidence of avoidable mistakes.

Perplexity's method separates two decisions: which sessions hold behavior worth imitating, and which turns hold mistakes worth correcting. Each assistant turn gets one of three treatments. Non-error turns in successful sessions receive cross-entropy (CE) loss and are imitated. Error turns with a validated hint receive Kullback-Leibler (KL) divergence loss and are corrected, in any session. Remaining turns stay in the input but receive no loss and are kept as context. Successful sessions can supply both imitation and correction targets. Unsuccessful sessions supply only correction targets.

A hint is a short corrective instruction grounded in information the model already had. In one example, a search call set recency_filter to 'year.' The schema allowed only 'day,' 'week,' or 'month.' The hint names the failed call, includes the validation error, and suggests an allowed value or omitting the optional field.

The corrective part uses On-Policy Self-Distillation (OPSD). The trainer runs the same GLM 5.2 checkpoint twice on the recorded turn. The teacher pass sees the hint; the student pass does not. Both use teacher forcing, so no replacement answer is generated. The teacher's next-token probabilities are detached and act as a soft target through forward KL. The combined loss is (CE + λ × KL), divided by the number of imitated tokens. Setting λ to 0 recovers standard SFT. The CE term matters because correction-only training can let teacher and student agree by ignoring context.

The pipeline draws from training-eligible Computer sessions served by GLM 5.2. Sessions with personally identifiable information and users who opted out are excluded. An LLM judge keeps tasks rated 4 or 5 on a 5-point difficulty scale. Two LLM judges must both approve the final delivery for a session to count as successful. For user feedback, three LLM judges locate the responsible turn, and at least two must agree. This is important because the last assistant turn before a complaint is the root cause only about half the time. Each hint is also checked against information available before the mistake, which reduces hindsight bias. In one example, a user asked for their 'w3' on Paychex. The model assumed a W-2 typo and searched for the wrong form. The hint targets that earlier interpretation, not just the final answer.

Hints work before training. On 985 held-out tool-error turns, the unchanged base model avoided the original failure in 93.7% of cases with hints, up from 75.1%. The share taking the corrected action rose from 60.6% to 82.3%. On user-feedback turns, fixed or on-track rates rose from 40.0% to 75.0% for explicit evidence. For inferred intent, they rose from 32.5% to 80.0%.

Offline tool errors fell. Recorded tool-error rates were 2.79% for stock GLM 5.2 and 1.35% for RFT only. The RFT plus OPSD checkpoint reached 0.87%. Perplexity notes these checkpoints used different training data, so this is not a matched ablation. Task-level benchmark results on suites like BrowseComp and SpreadsheetBench were mixed.

Live results are narrower. Each A/B test used about 100,000 users per condition. An early checkpoint versus stock GLM 5.2 showed 2.82% versus 2.94% failures, which was not significant. The later checkpoint comparison produced the significant 21.2% drop, without hints at inference. Strong dissatisfaction moved from 2.58% to 2.54%, which was also not significant. Perplexity did not compare the later checkpoint directly against stock GLM 5.2 online.

Editor's Summary Perplexity Research's post-training method uses failed user sessions, validated hints, and on-policy self-distillation to correct agent mistakes. In live A/B tests, tool-call failures fell from 2.24% to 1.77%, a 21.2% relative reduction, while strong user dissatisfaction showed no significant change. The weights and code were not released, and the model runs only inside Perplexity Computer, with GLM 5.2 as the openly available base model.