Perplexity cuts tool-call failures 21.2% with hint-guided on-policy self-distillation
inductionheads · x · 2026-09-23
Perplexity shared new research on its multi-stage post-training pipeline: SFT, RL, RFT, and a final stage of on-policy self-distillation (OPSD). The motivation: RL environments are realistic but still lack the diversity and complexity of real production use cases, leaving a sim2real gap.
Key ideas of OPSD:
- Learn from problematic trajectories by constructing hints from tool-call errors and negative user feedback in follow-up conversations;
- Use self-distillation to fix those errors, combining RFT (forward KL/CE) for successful trajectories with OPSD (reverse KL) for problematic ones.
Applied to post-training their Computer model (demonstrated on GLM-5.2), a later checkpoint reduced tool-call failures by 21.2% relative to an earlier one in a live A/B test.
More from coding & agent
- minpts-check: one-run predictor tells if your HDBSCAN clusters survive min_samples changes — leland_mcinnes · 2026-11-12
- Ten wild Gemini 4 Carbon demos in 24 hours: CUDA meadows, V8 engines, mech stomps — socialwithaayan · 2026-11-12
- How to cut Fable 5.2 token costs by 90%: Arc routing plus a lean CLAUDE.md — EXM7777 · 2026-11-12
- Hamel Husain's complete AI evals framework for PMs and engineers — aakashgupta · 2026-11-12
- AI coding in one year: from unusable to full automation with no code review — angelaqdai · 2026-11-12
- Interactions Tool Searches Web Animations by Motion for Claude Code to Rebuild — yuwen_lu_ · 2026-11-12