Reproduce “Policy Improvement-Guided Replay for Continual Reinforcement Learning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Qin-Wen Luo, Haoxuan Gao, Sheng Ren, Xiang Chen, Sheng-Jun Huang
CoLiR selects a limited set of earlier prompts for replay during sequential vision-language model training. It scores prompts using changes in likelihood for correct responses from the final task policy and incorrect responses from the initial policy. In the five-task MRCL evaluation, Qwen3-VL-4B achieved 69.35% mean final accuracy, versus 61.59% for CPO; this result covers the tested sequence and methods.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.