Reproduce “From Observation to Intervention: Distilling Self-Generated In-Context Demonstrations via Executable Causal Intervention” 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
Minming Ye, Min Liu, Haoang Chi, Ciheng Wu, Naiyang Guan
ECI2D tests self-generated demonstrations by checking whether a rationale’s executable reasoning structure produces correct answers on logically similar questions. On GSM-Symbolic-Plus-2, Qwen2.5-7B-Instruct achieves 65.28% accuracy with eight demonstrations, versus 58.12% for standard reinforced ICL and 62.10% for a same-format correctness-only control. The results support a local test of rationale faithfulness, not proof of recovered internal reasoning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.