Reproduce “Counterfactual Likelihood Tests for Indirect Influence in Private Reasoning Channels” 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

Alexander Boesgaard Lorup

The authors measure whether a proposer’s private reasoning influences a verifier’s fixed continuation by replacing the private block with a length-matched donor while holding public tokens and target fixed. A per-token negative log-likelihood increase indicates influence. Above a threshold of 0.05, A-to-B influence was 23.6%, versus 2.3% B-to-A; graph-cut controls yielded exact-zero scores for tested edges under this visibility mask.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.