Can verification scale to large models?

Lens adapted to this topic: How verification scales, and under what conditions

The constructive view is that verification is a genuine scaling axis for large-model systems. Candidate generation combined with verifiers can improve reasoning and agent performance without necessarily requiring larger models, while finer-grained scoring, repeated evaluation, criteria decomposition, and integration with static or formal tools can make verification more useful. This view treats scalability as task-dependent: objective checks and well-designed verifier pipelines are more promising than unconstrained general judgment.

• Verification can guide search and reinforcement learning, and theory suggests verifier-based methods can outperform verifier-free approaches under certain distributions of solution traces.
• Surveyed approaches use verifiers for process paths, final outcomes, or both, with prompt-based, discriminative, and generative designs.
• A general-purpose framework reports gains from finer-grained scores, repeated evaluation, and decomposed criteria across coding, robotics, and medical tasks.
• Combining LLMs with static analyzers and formal verification tools is presented as a way to strengthen code verification and safety assessment.
• Verification granularity affects both reasoning performance and compute efficiency, so invoking verification selectively may improve scaling economics.

Lens adapted to this topic: Why verification may not scale reliably

The skeptical view argues that verification cannot be treated as a universal substitute for capability or human judgment. Sampling more candidates only helps when the verifier is sufficiently accurate; imperfect tests can repeatedly select plausible but wrong answers. LLM judges may be inconsistent, vulnerable to prompt choices, and correlated with generators’ errors. In open-ended domains, the harder problem is defining correctness, so scalable verification requires layered deterministic checks and human review for consequential cases.

• With an imperfect verifier, false positives remain possible; repeated resampling therefore has an accuracy ceiling even with unlimited compute.
• Empirical work reports that verifier effectiveness falls on harder problems and that strong generators can produce errors that are harder for verifiers to detect.
• Smaller models can approach frontier-model accuracy on mathematical-proof verification but show substantially greater inconsistency, and prompt choice affects accuracy.
• A model verifying another model can reproduce shared assumptions or correlated errors, so agreement between models is not necessarily independent verification.
• In open-ended software work, computational checks depend on human-defined specifications; passing tests does not by itself establish the intended notion of correctness.