Perplexity Trains Its Computer Agent on Real Mistakes With Hint-Guided Self-Distillation
Perplexity Research has published a new post-training study that trains a model inside Perplexity Computer on real user sessions โ including failed ones. The method combines rejection sampling fine-tuning with hint-guided self-distillation. In a live A/B test, tool-call failures dropped from 2.24% to 1.77% across two trained checkpoints, which Perplexity reports as a statistically significant 21.2% relative reduction.
Is it deployable? Not directly. Perplexity has not released the post-trained weights or the training code. The model runs only as a model option inside Perplexity Computer. The base model, GLM 5.2, remains openly available on Hugging Face.
Why Outcome-Only Filtering Falls Short
Standard rejection sampling fine-tuning (RFT) evaluates each session and imitates only the successful ones. But a successful outcome does not guarantee every step was correct. An agent can recover from a bad tool call and still produce the right answer โ and imitating that full trajectory can reinforce the underlying error. Discarding failed sessions also throws away clear evidence of avoidable mistakes.
Imitate, Correct, or Keep as Context
Perplexity separates two decisions: which sessions contain behavior worth imitating, and which turns contain mistakes worth correcting.
Each assistant turn receives one of three treatments:
- Imitate: non-error turns in successful sessions receive cross-entropy (CE) loss.
- Correct: error turns with a validated hint receive Kullback-Leibler (KL) divergence loss, in any session.
- Keep as context: remaining turns stay in the input but receive no loss.
Successful sessions can supply both imitation and correction targets. Unsuccessful sessions supply only correction targets.
How a Hint Becomes a Training Signal
A hint is a short corrective instruction grounded in information the model already had. In one example, the agent is told what it should have checked rather than simply being given the right answer. This makes the hint a targeted nudge toward a better reasoning path, not a memorized fix.
Because the hint is derived from the model's own context, the corrected turn becomes a self-distillation target: the model learns to produce the improved behavior on its own, without the hint at inference time.
2026 Context: Why This Matters Now
As of 2026, agentic AI systems are moving from demos into production workflows, and reliability is the bottleneck. A 21.2% relative reduction in tool-call failures is meaningful for any team running computer-use agents at scale. The approach also reflects a broader industry trend: training on real-world experience, including failures, rather than curated success-only datasets.
For teams evaluating GLM 5.2 or building their own computer agents, Perplexity's method offers a useful template โ even without released weights. The key insight is simple: treat failures as labeled correction data, and let the model learn from its own mistakes through hint-guided self-distillation.
via MarkTechPost