Online Learnability of Chain-of-Thought Verifiers: Soundness and Completeness Trade-offs

Large language models with chain-of-thought generation have demonstrated great potential for producing complex mathematical proofs. However, their reasoning can often go astray, leading to increasing interest in formal and learned verifiers. A major challenge in learning verifiers, especially when their output will be used by the prover, is that this feedback loop may produce substantial distribution shift. Motivated by this challenge, we propose an online learning framework for learning chain-of-thought verifiers that, given a problem and a sequence of reasoning steps, check the correctness of the solution. Highlighting the asymmetric role of soundness (failure in catching errors in a proof) and completeness (flagging correct proofs as wrong) mistakes of the verifier, we introduce novel extensions of the Littlestone dimension which tightly characterize the mistake bounds for learning a verifier in the realizable setting. We provide optimal algorithms for finding the Pareto-frontier (the smallest total number of mistakes given a budget of soundness mistakes) as well as minimizing a linear combination of asymmetric costs. We further show how our learned verifiers can be used to boost the accuracy of a collection of weak provers, and enable generation of proofs beyond what they were trained on. With the mild assumption that one of the provers can generate the next reasoning step correctly with some minimal probability, we show how to learn a strong prover with small error and abstention rates.

Learning in Structured Stackelberg Games

We study structured Stackelberg games, in which both players (the leader and the follower) observe information about the state of the world at time of play. Importantly, this information may contain information about the follower, which the leader may use when deciding her strategy. Under this setting, we show that no-regret learning is possible if and only if the set of mappings from contexts to follower types that the leader uses to learn is not ``too complex''. Specifically, we find that standard learning theoretic measures of complexity do not characterize learnability in our setting and we give a new dimension which does, which we term the Stackelberg-Littlestone dimension. In the distributional setting, we give analogous results by showing that standard complexity measures do not characterize the sample complexity of learning, but a new dimension called the Stackelberg-Natarajan dimension does. We then show that an appropriate empirical risk minimization procedure achieves the corresponding sample complexity.