Reproduce “Coding Agent Memory Post-Training II: Unlocking the Memory Potential of Pre-trained File Operations for Long-Horizon Tasks via On-Policy Distillation” 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
Lirui Luo, Kelong Mao, Xinwei Yang, Rongqing Li, Luyu Chen, Kieran Wong, Yudong Guo, Xinrui Wang and 4 more
CAMP-II trains an agent to preserve task state in ordinary files across context replacements, using Bash and a frozen teacher to score the student’s sampled tokens, including memory writes. On CAMG-II’s 640-task test, its 4B policy averages 64.6, versus 51.8 for MemOPD with SFT. The result is limited to five environments; it remains below the teacher’s 76.0.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.