YuyaoGe/SkillForge_Alfworld_7B

Overview

YuyaoGe/SkillForge_Alfworld_7B is a specialized language model, fine-tuned from Qwen2.5-7B-Instruct, developed by Yuyao Ge and collaborators. It implements the SkillForge pipeline, a novel approach to agent training where skills and the agent's policy co-evolve through a dynamic lifecycle. This model is specifically optimized for the ALFWorld environment, addressing the common issue in memory-augmented agents where outdated skills remain in the library, hindering performance.

Key Capabilities

• Dynamic Skill Management: SkillForge introduces a lifecycle (trial, active, stable, retired) for skills, scoring each skill against the policy's own rollouts to ensure only relevant and effective skills are utilized.
• Enhanced Performance on ALFWorld: Achieves 92.4% overall success on ALFWorld, outperforming the strongest memory-augmented RL baseline, SkillRL (89.9%), and the untuned base model (14.8%).
• Adaptive Learning: The system prevents performance degradation from stale instructions by actively retiring or mutating skills that the policy has outgrown.
• Co-evolutionary Training: Skills and the policy are trained together, with GRPO (Generalized Policy Optimization) managing the policy updates and a forging cycle managing skill states.

Good for

• Research in Agent Learning: Ideal for researchers exploring dynamic skill acquisition, memory management in agents, and co-evolutionary training paradigms.
• ALFWorld Benchmarking: Provides a strong baseline and a unique approach for tasks within the ALFWorld environment.
• Understanding Skill Lifecycles: Demonstrates the benefits of a non-append-only skill library for improving agent robustness and efficiency.

This repository contains the policy weights and tokenizer; the evolved skill library is not included but its contents and fitness trajectories are documented in the associated paper and SkillFurnace (Appendix C).