Agentic on-policy distillation implicitly reshapes ReAct patterns and causes negative transfer
Deyang Kong · hf · 2026-11-18
A study from PKU, USTC, Tsinghua and Zhejiang University examines how on-policy distillation (OPD) implicitly reshapes LLM agents' ReAct interaction patterns across math, ALFWorld, and τ²-bench:
- General-reasoning supervision cuts math tool attempts from 70.4% to 14.4% while raising accuracy from 16.0% to 38.0%; code-interpreter supervision improves household actions and retail queries
- ALFWorld supervision drops τ²-bench accuracy from 41.5% to 6.6%, with frequent tool calls violating user-information dependencies
- Selective OPD guided by teacher preference shifts mitigates negative transfer, at a cost to source-domain learning
More from coding & agent
- Dev's Grok bot will auto-run a GTA 6 fan wiki, aggregating 5 platforms every 15 minutes — Rasmic · 2026-11-18
- AI-Generated Code Is Forcing Software to Become a Real Engineering Discipline — PaulShellDev · 2026-11-18
- Screenshot shows Claude Code injecting 'NOT USER INPUT' notices for background tasks — philhchen · 2026-11-18
- Cloudflare open-sources security-audit-skill, a 6-phase agent security auditor — _jaydeepkarale · 2026-11-18
- AI-era code is drowning teams in vulnerabilities—why scanners and manual reviews fall short — _jaydeepkarale · 2026-11-18
- MIT: capable frozen LLM agents autonomously invent their own self-improvement algorithms — MIT · 2026-11-18