Reproduce “Causal Chain of Thought: Mitigating Attention Leakage via Context Isolation” 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
Gabriele Dominici, Teague McMillan, Fazl Barez, Marc Langheinrich, Martin Gjoreski
CausalCoT tests whether restricting each reasoning step to its direct parents’ answers makes intermediate reasoning more consequential. It plans a directed acyclic graph, then answers nodes in separate contexts, with original input supplied only to root nodes. Across three datasets, GPT-5.4 achieves 48.43% Answer Change and 70.52% Predictable Change; these indicate intervention sensitivity, not internal faithfulness or guaranteed correctness.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.