Reproduce “Information-Gain Calibrated On-Policy Distillation for Deep Search Agents” 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
Chenxi Li, Xueyang Feng, Songyang Liu, Chenxu Wang, Xiaoying Tang, Chaozhuo Li, Litian Zhang
IG-OPD uses post-retrieval information gain to calibrate on-policy distillation for multimodal search agents. It reallocates supervision across turns using gain magnitude and adjusts search-action feedback using gain direction. Across six out-of-distribution benchmarks, its 4B student averages 55.0 accuracy versus 53.3 for standard OPD, but remains below its 8B teacher’s 56.1 average.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.