[Paper Review] Self-Verification Dilemma: Experience-Driven Suppression of Overused Checking in LLM Reasoning

Quanyu Long, Kai Jie Jiang+4 more|arXiv (Cornell University)|Feb 3, 2026

Explainable Artificial Intelligence (XAI)0 citations

TL;DR

The paper shows that many self-verification (recheck) steps in LLM reasoning are mostly confirmatory and offers an experience-driven test-time framework to selectively suppress redundant rechecks, reducing tokens while maintaining or improving accuracy.

ABSTRACT

Large Reasoning Models (LRMs) achieve strong performance by generating long reasoning traces with reflection. Through a large-scale empirical analysis, we find that a substantial fraction of reflective steps consist of self-verification (recheck) that repeatedly confirm intermediate results. These rechecks occur frequently across models and benchmarks, yet the vast majority are confirmatory rather than corrective, rarely identifying errors and altering reasoning outcomes. This reveals a mismatch between how often self-verification is activated and how often it is actually useful. Motivated by this, we propose a novel, experience-driven test-time framework that reduces the overused verification. Our method detects the activation of recheck behavior, consults an offline experience pool of past verification outcomes, and estimates whether a recheck is likely unnecessary via efficient retrieval. When historical experience suggests unnecessary, a suppression signal redirects the model to proceed. Across multiple model and benchmarks, our approach reduces token usage up to 20.3% while maintaining the accuracy, and in some datasets even yields accuracy improvements.

Motivation & Objective

• Quantify how frequently LLMs perform reflective self-verification during reasoning.
• Differentiate between rethink and recheck to understand the functional roles of reflection.
• Assess how often rechecks are corrective versus confirmatory and their impact on accuracy.
• Propose an offline experience-driven test-time framework to suppress low-utility rechecks without retraining models.
• Demonstrate the efficiency gains and accuracy trade-offs of the proposed approach across multiple models and math benchmarks.

Proposed method

• Empirical analysis of reflective steps in reasoning traces to categorize into rethink vs. recheck.
• Annotation of recheck outcomes as corrective or confirmatory using GPT-5 and human checks.
• Construction of an offline experience pool recording context and necessity of past rechecks.
• Development of a lightweight recheck activation detector (binary classifier with >97% accuracy).
• Retrieval of top-k similar experience units via BM25 to estimate the usefulness of current rechecks.
• Injection of a suppression signal when past experience suggests rechecks are unlikely to be beneficial, without altering model parameters.

Figure 1 : Reflective behaviors commonly observed in step-by-step mathematical reasoning. We illustrate three categories: rethink, where the model revises its strategy and explores an alternative line of reasoning; and recheck, where the model verifies already-derived intermediate results through re

Experimental results

Research questions

• RQ1 How frequently do LLMs exhibit reflective self-verification during reasoning across benchmarks and models?
• RQ2 What proportions of rechecks are corrective versus confirmatory, and how does this affect usefulness?
• RQ3 Can past verification experience be leveraged to selectively suppress redundant rechecks at test time without retraining?
• RQ4 What are the accuracy and efficiency trade-offs of applying experience-driven suppression (EDS) across diverse math benchmarks?

Key findings

• Reflective steps constitute a substantial portion of reasoning, often approaching or exceeding one third of steps across models and benchmarks.
• Rechecks are a large portion of reflections (about 40–58%), and on easier datasets they are more prevalent as local verification rather than strategy revision.
• Approximately 85–95% of rechecks are confirmatory and do not alter intermediate results or final answers.
• An offline experience pool enables estimating whether a current recheck will be beneficial, enabling selective suppression.
• EDS reduces average reasoning length by about 9% on average and up to 20.3% on MATH500, while maintaining or slightly improving accuracy across models/datasets.
• Compared to full suppression and aggressive truncation methods, EDS preserves necessary rethink and beneficial rechecks, achieving a favorable accuracy–efficiency trade-off.

Figure 2 : Percentage of steps classified as reflections.