跳到正文
研踪PaperTrail
← 研踪 PaperTrail

arXiv:2611.19244 · v2

Agentic On-Policy Distillation Implicitly Reshapes LLMs' ReAct Patterns

首次提交 2026-11-14 16:18
版本更新 2026-11-18 13:32
雷达观测 2026-11-19 19:08
arXiv ↗PDF ↗

摘要

On-policy distillation (OPD) transfers specialized capabilities to language-model agents, but its cross-domain effects on interaction remain unclear. We investigate how OPD reshapes ReAct interaction patterns across mathematics, ALFWorld, and τ2τ^2-bench, focusing on reasoning, actions, and observation acquisition. General-reasoning supervision reduces mathematical tool attempts from 70.4% to 14.4% while increasing accuracy from 16.7% to 38.0%; code-interpreter supervision improves household actions and retail queries. In contrast, ALFWorld supervision reduces τ2τ^2-bench accuracy from 41.5% to 6.6%, accompanied by frequent tool calls that violate user-information dependencies. To address negative transfer using only source-domain training trajectories, we explore selective OPD guided by teacher preference changes before and after domain-specific RL. Retaining the top 95% of positions raises τ2τ^2 accuracy from 11.5% to 29.9% with a source-learning trade-off.

AI 论文解读

基于摘要 · 非全文审阅

此版本尚未生成 AI 解读。解读由后台任务生成并按论文版本保存,登录后共享已有结果。

版本差异

本站首次收录为 v2,暂无更早版本可对照。

观测到的版本

相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。

v2 · 2026-11-18 13:32 · 最新观测版本

Agentic On-Policy Distillation Implicitly Reshapes LLMs' ReAct Patterns

On-policy distillation (OPD) transfers specialized capabilities to language-model agents, but its cross-domain effects on interaction remain unclear. We investigate how OPD reshapes ReAct interaction patterns across mathematics, ALFWorld, and τ2τ^2-bench, focusing on reasoning, actions, and observation acquisition. General-reasoning supervision reduces mathematical tool attempts from 70.4% to 14.4% while increasing accuracy from 16.7% to 38.0%; code-interpreter supervision improves household actions and retail queries. In contrast, ALFWorld supervision reduces τ2τ^2-bench accuracy from 41.5% to 6.6%, accompanied by frequent tool calls that violate user-information dependencies. To address negative transfer using only source-domain training trajectories, we explore selective OPD guided by teacher preference changes before and after domain-specific RL. Retaining the top 95% of positions raises τ2τ^2 accuracy from 11.5% to 29.9% with a source-learning trade-off.

查看此版本解读 →查看此版本原文 ↗