Reproduce “HERD: Harness-Evolving Recursive Distillation for Continual 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
Xun Sun, Thanh Thi Nguyen, Thuy-Trang Vu
HERD combines an external harness with weight updates to internalize useful experience for inference without harness context. It selects skills and high-residual records, then distills verified answers into weights under checkpoint and base-model drift constraints. On one 25-task TRACE+17 stream, it matches Seq-LoRA acquisition (0.713 vs. 0.715) and has lower forgetting (0.010 vs. 0.025), but some forgetting and capability loss remain.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.