Perplexity Details Hint-Guided Self-Distillation for Training Its Computer Agent

On September 23, Perplexity officially published post-training research for its Computer agent (model). The core method is "hint-guided self-distillation," and online A/B testing showed tool-call failure rate dropped 21.2%, down to 1.77%. The training data pipeline excludes sessions containing personally identifiable information (PII) and sessions from users who opted out of training; the data governance details were also made public and are worth attention.

Confirmed

• The officially disclosed multi-stage post-training pipeline consists of SFT, RL, RFT, and a final stage of on-policy self-distillation (OPSD). A repost from @inductionheads noted the motivation: although the RL environment is realistic, it still lacks the diversity and complexity of real production use cases, creating a sim2real gap. Hence reinforcement learning is done first in synthetic environments, then training uses real user sessions to expose failures outside designed scenarios.
• Method motivation: @perplexityai explained that rejection sampling fine-tuning (RFT) only imitates successful sessions, which reinforces errors recovered by luck and discards evidence from failed ones. The new method combines RFT with hint guidance, learning from real user sessions—imitating good action trajectories while explicitly correcting avoidable mistakes (such as wrong tool calls). Even if the overall trajectory ends up succeeding, errors within it are still corrected so the model learns from its own mistakes.
• Online A/B test results: a later-trained checkpoint cut tool-call failure rate by 21.2%, bringing it down to 1.77%.
• Data pipeline: explicitly excludes sessions containing PII and sessions from users who opted out of training.

Why it matters

• The method offers a reusable paradigm for continuously improving agents from real production sessions, bridging the gap between synthetic RL environments and real use cases.
• Explicitly correcting hidden errors in successful trajectories breaks through the common RFT limitation of "learning only from successes," offering direct reference value for other agent teams.
• The simultaneously disclosed data exclusion rules (PII, user opt-out) demonstrate one practical approach to data governance in agent training.

2026-09-23 ~ 2026-09-23 · 5 related posts

• [source] Perplexity's hint-guided self-distillation cuts its agent's tool-call failures by 21.2% — perplexity_ai · 2026-09-23
• [source] Perplexity details training data pipeline: PII and opt-out sessions excluded — perplexity_ai · 2026-09-23
• [source] Perplexity details hint-guided self-distillation: agent tool-call failures cut from 2.24% to 1.77% — perplexity_ai · 2026-09-23
• Perplexity's post-training method cuts agent tool call failures by 21% — beffjezos · 2026-09-23
• Perplexity cuts tool-call failures 21.2% with hint-guided on-policy self-distillation — inductionheads · 2026-09-23