CoChat
Sign Up
Cited by 5Topic ModelingNatural Language Processing TechniquesAdversarial Robustness in Machine Learning

Confidence v.s. Critique: A Decomposition of Self-Correction Capability for LLMs

Open access
Yang, Zhe, Zhang, Yichang, Wang, Yudong, Xu, Ziyao, Lin, Junyang, Sui, Zhifang·2025·Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)
DOI 10.18653/v1/2025.acl-long.203
Open this paper in CoChat — free accountOpen PDF

Abstract

Large Language Models (LLMs) can correct their self-generated responses, but a decline in accuracy after self-correction is also witnessed.To have a deeper understanding of selfcorrection, we endeavor to decompose, evaluate, and analyze the self-correction behaviors of LLMs.By enumerating and analyzing answer correctness before and after self-correction, we decompose the self-correction capability into confidence (being confident to correct answers) and critique (turning wrong answers to correct) capabilities, and propose two metrics from a probabilistic perspective to measure these 2 capabilities, along with another metric for overall self-correction capability evaluation.Based on our decomposition and evaluation metrics, we conduct extensive experiments and draw some empirical conclusions.For example, we find different models can exhibit distinct behaviors: some models are confident while others are more critical.We also find the trade-off between the two capabilities (i.e.improving one can lead to a decline in the other) when manipulating model self-correction behavior by prompts or in-context learning.Further, we find a simple yet efficient strategy to improve self-correction capability by transforming Supervision Fine-Tuning (SFT) data format, and our strategy outperforms vanilla SFT in both capabilities and achieves much higher accuracy after self-correction.Our code is publicly available on GitHub.

AI snapshot

Written from the paper’s full text.

TL;DR

The paper decomposes LLM self-correction into confidence (maintaining correct answers) and critique (turning wrong answers correct), with metrics CL, CS, and RSS. It finds a trade-off between the two under prompting/ICL and proposes CCT, a transformed SFT data strategy that improves both and raises post-self-correction accuracy over vanilla SFT.

Key findings

  • Self-correction does not necessarily improve accuracy: on GSM8k, GPT-3.5 Turbo improves by 2.7% after self-correction, while GPT-4 Turbo decreases by 1.5%; Llama3-8B-Instruct has Acc1 71.0% and Acc2 78.1%, whereas GPT-4 Turbo has Acc1 93.6% and Acc2 92.1%.
  • Across models, CL is generally high and CS relatively low; models with higher CS (e.g., Llama3-8B-Instruct) tend to have lower CL, suggesting difficulty achieving both simultaneously.
  • Prompting and ICL reveal a trade-off: confidence prompts/ICL examples raise CL but lower CS, while critique prompts/ICL examples raise CS but lower CL; simultaneous enhancement is hard without fine-tuning.
  • CCT outperforms vanilla SFT in post-self-correction accuracy: on GSM8k, CCT Acc2 44.2 vs. SFT 40.3; on MMLU, CCT Acc2 55.5 vs. SFT 48.9; on BoolQ, CCT Acc2 74.0 vs. SFT 63.8. On BoolQ, highest self-correction performance occurs when CLT data is about 40%.

Methods

  • The paper enumerates four self-correction scenarios—confident, unconfident, critical, stubborn—and defines Confidence Level (CL), Critique Score (CS), and Relative Self-Correction Score (RSS) from a probabilistic perspective, showing Acc2 = Acc1 × CL + (1 − Acc1) × CS.
  • Experiments cover classification and generation tasks: GSM8k, Humaneval, IFEval, MMLU, BoolQ, and CommonsenseQA, using open-source models (Llama3-8B/70B, Qwen2.5-7B/72B, DeepSeek-7B/67B, Mistral-7B-v3, GLM4-9B) and closed-source models (Qwen-Max, GPT-3.5 Turbo, GPT-4 Turbo).
  • Probability estimation uses logits for classification tasks and repeated sampling for generation tasks; open models are sampled 10 times per question, closed models 3 times, and datasets over 500 items are subsampled to 500. Behavior manipulation uses prompts and ICL; improvement experiments fine-tune Llama2-7B-Base with LoRA on GSM8k, MMLU, and BoolQ, comparing SFT, CLT, CST, and CCT.

Limitations

  • Stated: metric calculation relies on probability estimation and repeated sampling, making it more computationally expensive than non-probability evaluation; the decomposition is simplified and cannot distinguish stubborn wrong answers from questions too hard for the model, nor identify which question types trigger confidence or critique.
  • Stated: reasons for model-series similarities remain unknown; only static CLT/CST data mixing is explored; prompt manipulation is limited to simple prompts; classification probability estimation is relatively simple and could be optimized.
  • Evident from the text: the CCT improvement experiments are limited to Llama2-7B-Base with LoRA on three tasks, and the RSS derivation assumes CS ≤ Acc1 ≤ CL, which the paper says is empirically demonstrated but remains an assumption.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.18653/v1/2025.acl-long.203

Cites

0 citation errors · 13 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

  • Findingown data

    Models with higher CS values tend to have lower CL values.

    “models with higher CS values (e.g., Llama3-8-Instruct) tend to process lower CL values”
    Quote in the paperSubstantial claim
  • Findingown data

    Self-correction does not necessarily lead to an increase in accuracy.

    “Self-correction does not necessarily lead to an increase in Accuracy.”
    • Also stated in: § Confident：✓⇒✓ · hedged
      “self-correction usually but not necessarily leads to higher performance”
    • Also stated in: § Abstract · plain
      “a decline in accuracy after self-correction is also witnessed”
    Quote in the paperSubstantial claim
  • Findingown data

    With the increase of confident example number, CS decreases.

    “With the increase of confident example number, CL increases, and CS decreases.”
    Quote in the paperSubstantial claim
  • Findingown data

    A critique prompt/ICL example can cause higher CS.

    “Confidence prompt/ICL example can lead higer CL and lower CS; critique prompt/ICL example can cause lower CL and higher CS.”
    Quote in the paperSubstantial claim
  • Findingown data

    With the increase of confident example number, CL increases.

    “With the increase of confident example number, CL increases, and CS decreases.”
    Quote in the paperSubstantial claim
  • May combine two claims. 12
  • The quote only partly states this claim. 1
  • Classified as probably not a substantial claim. 5

  • May combine two claims.

    Findingown data

    CCT outperforms SFT in Acc2, CL, and CS, showing better self-correction capability.

    “CCT outperforms SFT in Acc 2 , CL, CS, showing better self-correction capability”
    • Also stated in: § Confident：✓⇒✓ · boosted
      “CCT outperforms SFT by a large margin on accuracy after self-correction, breaking the trade-off and achieving higher both CL and CS.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Acc1 and Acc2 exhibit an inverted U-shaped curve as the proportion of CLT data increases, and the model achieves its highest self-correction performance when the proportion of CLT data is approximately 40%.

    “Acc 1 and Acc 2 exhibit an inverted U-shaped curve (initially increasing and then decreasing), and the model achieves its highest self-correction performance when the proportion of CLT data is approximately 40%.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    On GSM8k, accuracy of GPT-3.5 Turbo is improved by 2.7% after self-correction, whereas accuracy of GPT-4 Turbo is decreased by 1.5%.

    “on the GSM8k dataset, accuracy of GPT-3.5 Turbo is improved by 2.7% after self-correction, whereas accuracy of GPT-4 Turbo is decreased by 1.5%.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    SFT achieves high Acc1 but low Acc2; CCT achieves high Acc2 but relatively low Acc1; and SFT+CCT can achieve both high Acc1 and Acc2.

    “SFT achieves high Acc 1 but low Acc 2 ; CCT achieves high Acc 2 but Acc 1 is relatively low; and SFT+CCT can achieve both high Acc 1 and Acc 2 .”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    In general, CL values are relatively high, while CS values are relatively low.

    “In general, the CL values are relatively high, while the CS values are relatively low.”
    • Also stated in: § Confident：✓⇒✓ · plain
      “confidence capability is generally better than critique capability for most models”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    A higher number of confidence examples increases confidence but diminishes critique capability, whereas more critique examples enhance CS but reduce CL.

    “a higher number of confidence examples increases confidence but diminishes critique capability, whereas more critique examples enhance CS but reduce CL.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Confidence prompt enhances CL across all tasks but diminishes CS.

    “confidence prompt enhances CL across all tasks but diminishes CS.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    CCT achieves the highest Acc2 across all three tasks (GSM8k, MMLU, BoolQ), significantly outperforming other methods.

    “CCT can lead to considerable accuracy improvements after self-correction and achieve the highest Acc 2 across all three tasks, significantly outperforming other methods”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    SFT+CCT can achieve both high Acc1 and Acc2.

    “SFT+CCT can achieve both high Acc 1 and Acc 2 .”
    Quote in the paperSubstantial claim

  • The quote only partly states this claim.

    Findingown data

    Models from the same series tend to show similar self-correction behaviors.

    “Models from the same series tend to show similar behaviors.”
    • Also stated in: § Confident：✓⇒✓ · plain
      “models from the same series tend to behave similarly”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    As the proportion of CLT data increases, CL consistently rises, while CS monotonically decreases.

    “as the proportion of CLT data increases, CL consistently rises, while the CS value monotonically decreases.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Critique prompt improves CS but reduces CL.

    “critique prompt improves CS but the price is a reduction in CL.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    CS reflects the extent to which the model persists in providing wrong answers.

    “CS reflects the extent to which the model persists in providing wrong answers.”
    Quote in the paperNot a substantial claim
    • The authors state this as an assumption: Questions are randomly sampled from the question set A with equal probability.
    • The authors state this as an assumption: The self-correction behavior of LLMs can be modeled by the stochastic process defined in §3.1.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Interpretationreasoning

    High CL indicates that the model is confident and will not change its correct answer even when challenged.

    “High CL also indicates the model is confident and will not change its correct answer even when challenged.”
    Quote in the paperNot a substantial claim
    • The authors state this as an assumption: Questions are randomly sampled from the question set A with equal probability.
    • The authors state this as an assumption: The self-correction behavior of LLMs can be modeled by the stochastic process defined in §3.1.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationnone given

    The higher CL is, the more confident the model is about its correct answers.

    “The higher CL is, the more confident the model is about its correct answers.”
    Quote in the paperNot a substantial claim
    • The authors state this as an assumption: Questions are randomly sampled from the question set A with equal probability.
    • The authors state this as an assumption: The self-correction behavior of LLMs can be modeled by the stochastic process defined in §3.1.

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    A lower CS value indicates a greater tendency for the model to stubbornly maintain erroneous responses.

    “A lower CS value indicates a greater tendency for the model to stubbornly maintain erroneous responses”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: Questions are randomly sampled from the question set A with equal probability.
    • The authors state this as an assumption: The self-correction behavior of LLMs can be modeled by the stochastic process defined in §3.1.

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    A higher CS value suggests a greater willingness of the model to correct these errors.

    “whereas a higher CS value suggests a greater willingness of the model to correct these errors.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: Questions are randomly sampled from the question set A with equal probability.
    • The authors state this as an assumption: The self-correction behavior of LLMs can be modeled by the stochastic process defined in §3.1.
17 of 17

  • The authors state this as an assumption.

    The inequality CS ≤ Acc1 ≤ CL holds.

    “we assume the following inequality holds: CS ≤ Acc 1 ≤ CL”
    “Experimental results in §4 also empirically demonstrate that this inequality is valid.”

  • The paper relies on this without support.

    The true values of P(ai), P(bi|ai), and P(bi|¬ai) are unattainable.

    “However, the actual values of these probabilities are unattainable.”

  • The authors state this as an assumption.

    Without loss of generality, the correct label is l0.

    “Without any loss of generality, assume the correct label is l 0”

  • The authors state this as an assumption.

    Questions are randomly sampled from the question set A with equal probability.

    “Randomly sampling a question q from A with equal probability.”

  • The authors state this as an assumption.

    The self-correction behavior of LLMs can be modeled by the stochastic process defined in §3.1.

    “Let's think about the stochastic process defined in §3.1:”
5 of 5

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • CST improves self-correction on GSM8k despite the contrary claim

    Minor

    Table 3 shows GSM8k CST improving from 33.1% before correction to 42.2% after correction, compared with SFT's 39.3% to 40.3%. Thus, the categorical statement that CST alone cannot improve self-correction capability is not supported by these results.

  • SFT does not have the best initial accuracy on every task

    Minor

    The claim holds for GSM8k and BoolQ, but Table 3 reports MMLU Acc 1 of 51.2% for CCT versus 48.6% for SFT.

  • Reported results violate the claimed CL bound

    Minor

    The paper says its §4 results validate Acc 1 ≤ CL. In Table 6, GPT-4 Turbo on CommonsenseQA has Acc 1 of 85.0% but CL of 81.7%, violating that inequality.

3 of 3

Cites

The publisher hasn’t deposited this paper’s reference list.

Research works printed but not linked: 2

  1. Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2021. Measuring massive multitask language under- standing. In International Conference on Learning Representations.
  2. Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9.

Cited by

Sorted by citations

Related

Authors
Yang, Z., Zhang, Y., Wang, Y., Xu, Z., Lin, J., & Sui, Z.
Affiliations
  • Peking University (CN)
  • Alibaba Group (China) (CN)
ORCID
Venue
Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)
Publisher
Association for Computational Linguistics
Published
2025
Pages
3998-4014
DOI
10.18653/v1/2025.acl-long.203
OpenAlex
W4412886705
Semantic Scholar
b1082a134ba499843becf0f2ba243e5bd675bbb7
Cited by
5
References
0
Citations by year
2026: 5
FWCI
5.41
Type
conference paper
Language
en
Open access
Licence
cc-by
Locations
Annual Meeting of the Association for Computational Linguistics (ACL) (publishedVersion, cc-by)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
  • Physical Sciences › Computer Science › Artificial Intelligence › Natural Language Processing Techniques
  • Physical Sciences › Computer Science › Artificial Intelligence › Adversarial Robustness in Machine Learning
Keywords
large language models, self-correction, confidence, critique, evaluation metrics, probabilistic metrics, prompt manipulation, in-context learning

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-07.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-07.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review