Rewindable proofs proposed for scalable oversight without a truthful prover
Original title: Efficient Scalable Oversight via Rewinding
The honest prover is linear in the original computation, and the verifier is nearly linear in input length after preprocessing.
The paper introduces rewindable interactive proofs, letting a weak verifier snapshot and restore a stronger AI prover during scalable oversight. The construction targets computations robust to changing a small fraction of oracle answers, uses no cryptography, and avoids assuming one prover is truthful. The authors implement the protocol, report a preliminary MNIST experiment, and say an RIP can be faster than a deployed proof system.
Unconfirmed: The MNIST experiment is described as preliminary.
- Rewinding restores a saved model state and continues the interaction.
- The verifier's near-linear time follows a one-time preprocessing step.
- The protocol applies when the verifier controls the prover's machine.
- Authors say the construction is more efficient than prior proofs without a truthful prover.
From the original article
1 min read at arxiv.orgScalable oversight asks how a weak verifier can check the outputs of computations performed by more powerful AI systems. One proposed approach uses relativizing interactive proofs, in which a weak verifier checks the output of a complex oracle-aided computation by interacting with one or more powerful provers.
By Ziyi Guan · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.