Skip to content

RUCBM/ExpInternalization

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

5 Commits
 
 
 
 
 
 

Repository files navigation

Rethinking Continual Experience Internalization for Self-Evolving LLM Agents

Paper

Official code for the paper Rethinking Continual Experience Internalization for Self-Evolving LLM Agents.

The pipeline turns an agent's interaction trajectories into a compact, principle-level experience pool and internalizes it into the model parameters, so the agent and its experience co-evolve across iterations.

Components

  • experience_extraction/ — agent rollout + experience summarization: a local model solves tasks (web reasoning / math), and the scored trajectories are distilled into a reusable experience pool. See its README for setup and usage.
  • evaluation/ — web-agent evaluation on WebWalkerQA, GAIA-Text-103, and BrowseComp-ZH, with and without inference-time experience injection. Built on the Tongyi DeepResearch inference framework. See its README.

More components (experience injection and context-distillation training) will be released here.

About

Official code for "Rethinking Continual Experience Internalization for Self-Evolving LLM Agents"

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors