mlsanityData to 14 Oct 2026

Self-Correction Bench: Uncovering and Addressing the Self-Correction Blind Spot in Large Language Models

arXiv 2507.02778 ↗First posted 3 Jul 2025cs.CL cs.AI cs.LGcited by 20

On the map: LLM reasoning → Test-time compute and efficient thinking

Abstract (arXiv)

Although large language models (LLMs) have transformed AI, they still make errors and follow unproductive reasoning paths. Self-correction is vital for safety-critical applications, but studying it requires disentangling activation failure from knowledge deficiency: when a model fails to correct an error, is it because it cannot, or because it does not? We introduce Self-Correction Bench, a controlled evaluation framework that isolates this distinction by injecting the same error as either an external (user-attributed) or internal (model-attributed) error, keeping all other context identical. Testing 14 open-source non-reasoning models reveals a 64.5% Self-Correction Blind Spot: models correct external errors but fail on identical internal ones, proving the capability exists but is not activated. On models' own naturally generated errors, a measurable share of what a model fails to catch in its own output is caught when the identical error is presented externally. We trace the cause to post-training data composition: supervised fine-tuning datasets lack error-correction sequences, and fine-tuning with as few as 5,306 such traces already reduces the blind spot by 76.0%. Mechanistically, we identify a transferable conversational-role direction in representation space that causally gates self-correction. Appending "Wait" requires no training yet reduces the blind spot by 89.3%, and operates through a nearly independent pathway, indicating that correction activation is not reducible to this single mechanism.

Key facts. Cited by 20 papers in the map (data through 14 Oct 2026). Placed in Test-time compute and efficient thinking, one of 17 directions in LLM reasoning. Builds on 36 of its 54 references that are in the map. First posted on arXiv 3 Jul 2025.

Builds on

36 of 54 references.

Closest ancestors · 5

Same direction, newest first

  1. 2025s1: Simple test-time scaling31 Jan
  2. 2025Critique Fine-Tuning: Learning to Critique is More Effective than Learning to Imitate29 Jan
  3. 2024Training Language Models to Self-Correct via Reinforcement Learning19 Sep
  4. 2023Large Language Models Cannot Self-Correct Reasoning Yet3 Oct
  5. 2023Self-Refine: Iterative Refinement with Self-Feedback30 Mar

Other references · 31

  1. 2022Training language models to follow instructions with human feedbackcited by 15,971
  2. 2021LoRA: Low-Rank Adaptation of Large Language Modelscited by 14,380
  3. 2021Training Verifiers to Solve Math Word Problemscited by 10,109
  4. 2024DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Modelscited by 9,379
  5. 2021Measuring Mathematical Problem Solving With the MATH Datasetcited by 5,371
  6. 2023Reflexion: Language Agents with Verbal Reinforcement Learningcited by 4,276
  7. 2023Let's Verify Step by Stepcited by 3,789
  8. 2023Jailbroken: How Does LLM Safety Training Fail?cited by 1,651
  9. 2023A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questionscited by 1,601
  10. 2023Steering Llama 2 via Contrastive Activation Additioncited by 1,085
  11. 2024Tulu 3: Pushing Frontiers in Open Language Model Post-Trainingcited by 1,053
  12. 2024Refusal in Language Models Is Mediated by a Single Directioncited by 1,032
19 more
  1. 2023Steering Language Models With Activation Engineeringcited by 922
  2. 2020On Faithfulness and Factuality in Abstractive Summarizationcited by 900
  3. 2023A Multitask, Multilingual, Multimodal Evaluation of ChatGPT on Reasoning, Hallucination, and Interactivitycited by 894
  4. 2023Large Language Models Can Be Easily Distracted by Irrelevant Contextcited by 841
  5. 2023OpenAssistant Conversations -- Democratizing Large Language Model Alignmentcited by 721
  6. 2023Language Models can Solve Computer Taskscited by 372
  7. 2023Formalizing and Benchmarking Prompt Injection Attacks and Defensescited by 364
  8. 2023How Language Model Hallucinations Can Snowballcited by 311
  9. 2024When Can LLMs Actually Correct Their Own Mistakes? A Critical Survey of Self-Correction of LLMscited by 240
  10. 2024The Effect of Sampling Temperature on Problem Solving in Large Language Modelscited by 170
  11. 2023Benchmarking Cognitive Biases in Large Language Models as Evaluatorscited by 150
  12. 2024Cognitive Bias in Decision-Making with LLMscited by 106
  13. 2025PRMBench: A Fine-grained and Challenging Benchmark for Process-Level Reward Modelscited by 99
  14. 2022Capturing Failures of Large Language Models via Human Cognitive Biasescited by 98
  15. 2024Alice in Wonderland: Simple Tasks Showing Complete Reasoning Breakdown in State-Of-the-Art Large Language Modelscited by 68
  16. 2024Evaluating LLMs at Detecting Errors in LLM Responsescited by 50
  17. 2025Infinity Instruct: Scaling Instruction Selection and Synthesis to Enhance Language Modelscited by 50
  18. 2025On Linear Representations and Pretraining Data Frequency in Language Modelscited by 23
  19. 2023N-Critics: Self-Refinement of Large Language Models with Ensemble of Criticscited by 8

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2025: 1; 2026: 19.

20 papers.

Sort:

By area:

  1. 2026Why Reasoning Fails to Plan: A Planning-Centric Analysis of Long-Horizon Decision Making in LLM Agentscited by 23
  2. 2026The Last AI Built by Humans: Toward Genuine Recursive Self-Improvementcited by 11
  3. 2026The Illusion of Insight in Reasoning Modelscited by 9
  4. 2026OptiRepair: Closed-Loop Diagnosis and Repair of Supply Chain Optimization Models with LLM Agentscited by 5
  5. 2026Self-Improvement of Large Language Models: A Technical Overview and Future Outlookcited by 5
  6. 2025The Narrative Continuity Test: A Conceptual Framework for Evaluating Identity Persistence in AI Systemscited by 3
  7. 2026Structure Enables Effective Self-Localization of Errors in LLMscited by 3
  8. 2026Self-Attribution Bias: When AI Monitors Go Easy on Themselvescited by 3
  9. 2026Reasoning on the Manifold: Bidirectional Consistency for Self-Verification in Diffusion Language Modelscited by 2
  10. 2026Partially Correlated Verifier Cascades in LLM Harnesses: Concave Log-Odds, Polynomial Reliability, and Blind-Spot Ceilingscited by 2
  11. 2026Governing Reflective Human-AI Collaboration: A Framework for Epistemic Scaffolding and Traceable Reasoningcited by 1
  12. 2026ReFlect: An Effective Harness System for Complex Long-Horizon LLM Reasoningcited by 1
8 more
  1. 2026ESC: Emotional Self-Correction for Reliable Vision-Language Modelscited by 1
  2. 2026CUSUM-Shaped Inference-Time Monitoring and Targeted Re-Decoding for Quantized Small Language Model Reasoningcited by 1
  3. 2026ProCrit: Self-Elicited Multi-Perspective Reasoning with Critic-Guided Revision for Multimodal Sarcasm Detectioncited by 0
  4. 2026Spectral Origins of the Self-Correction Blind Spot in Autoregressive Generationcited by 0
  5. 2026Calibrated e-CUSUM Decoding for Quantized Reasoning Models: Why Token Log-Probability Is the Wrong Observable for Decoding Monitorscited by 0
  6. 2026Think Through a Bottleneck: Hourglass Reasoning for Rigorous Inductioncited by 0
  7. 2026When Do Models Admit They Are Wrong? Failure Disclosure Is Unstable Under Reinforcement Learningcited by 0
  8. 2026Anatomy of LLM Sycophancy: What a Flip Rate Hidescited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2024Small Language Models Need Strong Verifiers to Self-Correct Reasoning
  2. 2024Learning to Check: Unleashing Potentials for Self-Correction in Large Language Models
  3. 2025Language Models can perform Single-Utterance Self-Correction of Perturbed Reasoning
  4. 2025Synthetic Error Injection Fails to Elicit Self-Correction In Language Models
  5. 2024Internal Consistency and Self-Feedback in Large Language Models: A Survey
  6. 2026Contextual Drag: How Errors in the Context Affect LLM Reasoning
  7. 2025Can LLMs Correct Themselves? A Benchmark of Self-Correction in LLMs
  8. 2024Large Language Models have Intrinsic Self-Correction Ability