Neural Interactive Proofs
Lewis Hammond, Sam Adam-Day
TL;DR
This paper introduces neural interactive proofs, a framework where neural network-based agents interact to verify solutions, with theoretical analysis and experiments demonstrating potential applications in AI safety and problem-solving.
Contribution
It presents a unifying framework for neural interactive proofs, new protocols for interaction, and a theoretical comparison of approaches, supported by experiments in graph isomorphism and code validation.
Findings
Neural interactive proofs can verify complex tasks using neural networks.
New protocols improve interaction efficiency and robustness.
Experimental results demonstrate feasibility in toy and real-world tasks.
Abstract
We consider the problem of how a trusted, but computationally bounded agent (a 'verifier') can learn to interact with one or more powerful but untrusted agents ('provers') in order to solve a given task. More specifically, we study the case in which agents are represented using neural networks and refer to solutions of this problem as neural interactive proofs. First we introduce a unifying framework based on prover-verifier games, which generalises previously proposed interaction protocols. We then describe several new protocols for generating neural interactive proofs, and provide a theoretical comparison of both new and existing approaches. Finally, we support this theory with experiments in two domains: a toy graph isomorphism problem that illustrates the key ideas, and a code validation task using large language models. In so doing, we aim to create a foundation for future work on…
Peer Reviews
Decision·ICLR 2025 Poster
- I think the problem discussed by this work is important - I'm not aware of the NIP protocol being investigated in the past (though it is an obvious generalization of other things that have been studied). I think that the NIP protocol is plausibly better than other protocols that have been investigated, so it's a contribution to suggest it and empirically investigate it.
This paper has two main contributions: theory and empirical results. I'm concerned that the theoretical results aren't very important, they aren't crucially related to the empirical results, and the empirical results (while promising) aren't very detailed/strong. ## The theoretical results aren't very important Firstly, I was not persuaded that a lot of the theory is very useful to understand. I assume that the results are correct, but I don't think they're very informative about how to constr
This paper stands out for bridging theoretical and practical contributions while maintaining high standards of rigor throughout, as well as for the generality of the neural interactive proofs framework. 1. Originality: - Novel unification of interactive proofs with neural networks, bridging complexity theory and machine learning - Creative adaptation of game-theoretic concepts to create learnable verification protocols, particularly use of Nash and Stackelberg equilibria and their relat
I believe there are two central weaknesses of this paper, beyond what is listed in section 7 (lack of use of advanced methods, evaluation only on two domains, lack of evaluation of all protocols): The biggest weakness is that the scope, motivation, and impact of NIP are not made clear. Perhaps this is just a lack of context on interactive proofs and game theory on my part, but I cannot tell whether the combination of game-theoretic equilibria with generalization-from-examples for specifying th
The formulation of the problem into a zero-knowledge based verifier-prover interaction.
While I like the formulation, I have some concerns about this paper: 1. It is unclear on the zero-knowledge aspect. Why do we need to have the zero-knowledge in their interactions? The argument on the potential model stealing needs to be more carefully consolidated. 2. The actual technical contribution is limited. Neural interactive proof is technically explained with worst-case loss and Stackelberg game. First, I wasn't able to understand the paragraph under Proposition 2, due to the confus
Code & Models
Videos
Taxonomy
TopicsNeural Networks and Applications