Human-LLM Deliberation as Interactive Proof: Conditions for Verifiability Without Transparency
The paper argues that LLM claims can be verified without seeing inside the model, but only under explicit bounds on false passes and human checking error.
It treats the LLM as a prover and the human as a limited verifier who asks for supporting details and checks them locally. The result is an anytime-valid soundness guarantee: the chance of ever accepting a false claim stays below a chosen error level if the bounds hold across the interaction. Completeness over a finite horizon needs more: honest answers must be adequate, and the checks must make enough diagnostic progress. The authors stress that more checking can raise confidence, but only within the human verifier’s limits of effort, expertise, cognitive load, and fatigue. ArXiv · AI/CL/LG's note
It treats the LLM as a prover and the human as a limited verifier who asks for supporting details and checks them locally. The result is an anytime-valid soundness guarantee: the chance of ever accepting a false claim stays below a chosen error level if the bounds hold across the interaction. Completeness over a finite horizon needs more: honest answers must be adequate, and the checks must make enough diagnostic progress. The authors stress that more checking can raise confidence, but only within the human verifier’s limits of effort, expertise, cognitive load, and fatigue. ArXiv · AI/CL/LG's note
score 5