Prover-Verifier Games: Theory & Applications

• Prover-verifier games are interactive protocols where a powerful prover and a resource-limited verifier exchange messages to establish proof validity.
• They underpin complexity theory by characterizing classes like IP, MIP, and QMIP while enabling breakthroughs in zero-knowledge and device-independent proofs.
• Recent extensions incorporate neural and rational variants, advancing AI alignment, blockchain verification, and proof-complexity in combinatorial logic.

A prover-verifier game is a formal interactive protocol between two (or more) agents—a "prover" with access to greater computational power or privileged information, and a "verifier" restricted in computational resources. These protocols are central in complexity theory, cryptography, proof theory, neural verification, and emerging AI alignment contexts. The aim is to design interactive mechanisms where the verifier, through communication and possibly randomness, extracts robust guarantees of correctness or knowledge from the prover, despite adversarial or strategic incentives.

1. Formal Models and Complexity-Theoretic Foundations

Classically, prover-verifier games instantiate interactive proof systems where a verifier seeks to decide language membership with the help of a prover. In the single-prover model (IPs), the class IP captures PSPACE, while multi-prover interactive proofs (MIP), where non-communicating provers are interrogated in parallel, correspond to NEXP. Restricting the answer size, as in one-bit $k$-prover games, yields a detailed hierarchy parameterized by soundness error thresholds, linking $\mathsf{BPP}$, $\mathsf{SZK}$, $\mathsf{AM}$, $\mathsf{EXP}$, and $\mathsf{NEXP}$ as the allowed acceptance probability for cheating provers increases (Austrin et al., 2013). Quantum generalizations (MIP$^*$, QMIP$^*$) further extend the landscape, admitting entangled provers and yielding separations from classical complexity classes.

Interactive verifier-prover protocols also underlie constraint satisfaction hardness results, succinctly capturing the checkability of solutions or proofs via minimal communication, and form the foundation for the PCP theorem and related inapproximability results.

2. Game-Theoretic and Strategic Variants

Beyond adversarial verification, prover-verifier games are general frameworks for modeling negotiation, conflict, or cooperation—ranging from classical zero-sum refereed games to rational proofs and mechanism design settings. In non-cooperative rational interactive proofs (ncRIP), provers seek to maximize their individual utility under strong sequential equilibrium refinements, giving fine-grained control over protocol robustness via a utility gap parameter (Chen et al., 2017).

Modern AI and machine learning extend the paradigm to scenarios where both agents are neural networks, possibly with different capacities. Here, the prover may be an untrusted, more powerful model and the verifier a "trusted" but weaker system whose output must remain robust to persuasion or deception by the prover (Hammond et al., 2024). Game-theoretic solution concepts—Nash equilibrium, Stackelberg equilibrium, and subgame dominance—determine verifiability and incentive alignment. Leader-follower (Stackelberg) variants are critical for ensuring robustness, as only verifier-leading Stackelberg equilibria generally guarantee both completeness and soundness in learning-based settings (Anil et al., 2021).

3. Instantiations: Neural, Quantum, and Proof-Complexity Applications

Neural Prover-Verifier Games

With large-scale neural models, prover-verifier games address the need for outputs to be efficiently checkable by smaller, trusted systems. In "Neural Concept Verifier" (NCV), the classifier is decomposed into agents: a concept extractor (e.g., CLIP- or NCB-based), provers selecting concept subsets, and a verifier operating only on masked concept encodings. This framework unifies interpretable, nonlinear classification with verifiability at scale, providing high completeness and near-perfect soundness, even on high-dimensional data and logically complex regimes (Turan et al., 10 Jul 2025). Adversarial/protective provers (Merlin/Morgana roles) ensure the verifier cannot exploit spurious or shortcut features.

In LLMs, checkable reasoning is enforced by adversarial training of helpful and sneaky provers and small verifiers, leading to improved legibility—outputs that are not only correct but checkable by humans or weaker models (Kirchner et al., 2024). Mitigating the so-called "legibility tax" (accuracy drop for requiring checkability) is addressed by decoupling correctness maximization (solved by the main model) from translation into checkable form (handled by a faithful translator in a decoupled PVG) (Kim et al., 26 Feb 2026).

Quantum Prover-Verifier Games

Quantum interactive proof models extend the classical setting to quantum provers and/or quantum verifiers. Modern protocols compress multi-prover interactive proofs into nonlocal games with entangled provers, revealing that certain verification tasks cannot achieve efficient error reduction through parallel repetition due to quantum hedging phenomena (Arunachalam et al., 2013, Molina, 2012). Complexity-theoretic separations show, for example, that quantum multi-prover games with small question and answer size are strictly harder than the classical ones and are NEXP-complete to approximate (Ji, 2016). Constructing sum-of-squares hierarchies with careful certificate design establishes convergent bounds on quantum advantage in compiled nonlocal games (Cui et al., 23 Jul 2025).

Importantly, in cryptographic contexts, prover-verifier games realize zero-knowledge, proofs of knowledge, and device-independent certification. Recent advances provide classical proofs of quantum knowledge for ground states of local Hamiltonians via multi-prover nonlocal games, with explicit extractors reconstructing quantum witnesses when the game is won with high probability (Broadbent et al., 17 Mar 2025).

Proof Complexity and Logic

Prover-verifier games directly correspond to proof systems, especially in the setting of sequent calculi and bounded arithmetic. Pudlák-Buss style games are polynomially equivalent to proof systems operating over branching programs and their non-deterministic or alternating extensions (Das et al., 22 Aug 2025). Winning strategies in such games are dual to proofs, and game lower bounds translate to proof complexity lower bounds.

Pebble games with backtracking options encode the difficulty of proving combinatorial principles like the pigeonhole principle in bounded arithmetic, providing switching-lemma-free impossibility results for certain logical theories (Ken et al., 2024).

4. Information-Theoretic and Verification Guarantees

Central to prover-verifier games is the formalization and analysis of completeness and soundness:

• Completeness: Honest provers communicating correct proofs or evidence should enable the verifier to accept with high probability.
• Soundness: Cheating provers—subject to classical, quantum, or no-signaling constraints—cannot induce verifier acceptance above a bounded threshold.

In learning-based PVGs, "completeness" and "soundness" become classification performance under (respectively) honest and adversarial prover masking, with formal lower bounds on the mutual information between selected features and labels (Turan et al., 10 Jul 2025). In quantum protocols, the existence of hedging phenomena and no-signaling strategies limits the ability to drive soundness arbitrarily close to zero through repetition. Explicit mapping between verification and simulation shows that, in many quantum settings, efficient verification is as hard as efficient approximate simulation (França et al., 2020).

5. Methodological Variants and Open Directions

The design of prover-verifier games encompasses various protocol templates—simultaneous vs. sequential move (Nash vs. Stackelberg), single vs. multi-prover, adversarial vs. cooperative, classical vs. quantum, and checkability via symbolic, natural language, or latent representations.

Contemporary open problems include:

• Extending theorem-proving-based PVGs to structured prediction, reinforcement learning, or more general sequential tasks (Turan et al., 10 Jul 2025);
• Developing automated or human-aligned vocabularies of checkable concepts;
• Strengthening theoretical guarantees into the large-scale, multi-class, or multimodal regime;
• Implementing practical, efficient protocols for economic or blockchain settings with rational, strategic agents (Jain et al., 2016, Chen et al., 2017);
• Closing quantitative complexity gaps in quantum and one-bit multi-prover settings (Austrin et al., 2013, Cui et al., 23 Jul 2025);
• Characterizing the epistemic and proof-complexity power of new game forms, such as those implementing device-independent, zero-knowledge, or quantum proof-of-knowledge (Broadbent et al., 17 Mar 2025).

The prover-verifier paradigm, through its abstraction of trust, computation, and interaction, remains central to a wide array of foundational and applied domains, from classical and quantum complexity to the safe deployment and oversight of advanced AI systems.