Perplexity Reveals Hint-Guided Self-Distillation Post-Training, Cutting Tool-Call Failure Rate 21.2%

On September 23, Perplexity published its post-training research for the Computer agent. The core method is hint-guided self-distillation, letting the model learn from its own errors in real user sessions. Online A/B tests showed the later checkpoint reduced tool-call failure rate by 21.2%, down to 1.77%. The training data pipeline explicitly excludes sessions containing PII and sessions from users who opted out of training.

Confirmed

• The multi-stage post-training pipeline consists of SFT, RL, RFT, and a final stage of on-policy self-distillation (OPSD). Perplexity's Denis Yarats introduced this pipeline and noted it originates from ICML'26 research.
• Motivation: RL environments, while realistic, still lack the diversity and complexity of real production use cases (a sim2real gap), so RL is done in synthetic environments first, then training on real user sessions exposes failures beyond designed scenarios.
• Perplexity explained that RFT only imitates successful sessions, reinforcing luckily-recovered errors and discarding evidence from failed sessions. The new method combines RFT with hint guidance, imitating good trajectories while explicitly correcting avoidable errors such as wrong tool calls—even within otherwise successful trajectories.
• Data pipeline: sessions with PII and user opt-out sessions are excluded.

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 overcomes the common RFT limitation of learning only from successes.
• The disclosed data exclusion rules (PII, opt-out) demonstrate a data-governance practice for agent training.

2026-09-23 ~ 2026-09-24 · 6 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
• Perplexity bridges sim2real with RFT and on-policy self-distillation (OPSD) post-training — siyan_zhao · 2026-09-24