When to Trust the Cheap Check: Weak and Strong Verification for Reasoning

2026-02-19

AI Summary

This work addresses the lack of an effective decision mechanism for balancing weak verification (fast but unreliable) and strong verification (reliable but costly) in large language model inference. It formalizes, for the first time, a weak–strong verification framework and proposes a dynamic two-threshold policy that decides, based on the confidence of weak verification, whether to accept, reject, or escalate a query to strong verification. The approach makes no distributional assumptions about the query stream, models, or verifiers, and integrates online learning to rigorously control population-level false acceptance and rejection rates while minimizing reliance on strong verification. Theoretical analysis reveals that the optimal policy inherently exhibits a two-threshold structure and elucidates how calibration and sharpness govern the utility of weak verification.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Natural Language Processing: Fact-Checking / Misinformation Detection (NLP Focus)
• Reasoning under Uncertainty: Other Foundations of Reasoning under Uncertainty

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Economics, Online Markets and Human Computation: LLM based quality controls for crowd work

Abstract

Reasoning with LLMs increasingly unfolds inside a broader verification loop. Internally, systems use cheap checks, such as self-consistency or proxy rewards, which we call weak verification. Externally, users inspect outputs and steer the model through feedback until results are trustworthy, which we call strong verification. These signals differ sharply in cost and reliability: strong verification can establish trust but is resource-intensive, while weak verification is fast and scalable but noisy and imperfect. We formalize this tension through weak--strong verification policies, which decide when to accept or reject based on weak verification and when to defer to strong verification. We introduce metrics capturing incorrect acceptance, incorrect rejection, and strong-verification frequency. Over population, we show that optimal policies admit a two-threshold structure and that calibration and sharpness govern the value of weak verifiers. Building on this, we develop an online algorithm that provably controls acceptance and rejection errors without assumptions on the query stream, the language model, or the weak verifier.

Problem

• weak verification
• strong verification
• trust calibration
• verification policy
• LLM reasoning

Innovation

• weak verification
• strong verification
• verification policy
• online algorithm
• threshold structure