mlsanityData to 7 Oct 2026

ReBeCA: Unveiling Interpretable Behavior Hierarchy behind the Iterative Self-Reflection of Language Models with Causal Analysis

arXiv 2602.06373 ↗First posted 6 Feb 2026cs.CLcited by 0

Not on the map. Why?

Abstract (arXiv)

While self-reflection can enhance language model reliability, its underlying mechanisms remain opaque, with existing analyses often yielding correlation-based insights that fail to generalize. To address this, we introduce **ReBeCA** (self-**Re**flection **Be**havior explained through **C**ausal **A**nalysis), a framework for analyzing the interpretable behavioral hierarchy governing the self-reflection outcome. By modeling self-reflection trajectories as causal graphs, ReBeCA selects observed parent candidates and evaluates their stability through a three-stage ICP-based pipeline. In a controlled Qwen3 case study, we establish three critical findings: (1) Behavioral hierarchy: Semantic behaviors of the model influence final self-reflection results hierarchically: directly or indirectly; (2) Causation matters: Generalizability in self-reflection effects is limited to just a few semantic behaviors; (3) More neq better: The confluence of seemingly positive semantic behaviors, even among direct causal factors, yield no additive gain. ICP-based verification identifies sparse causal parents achieving up to 49.6% structural likelihood gains relative to the dense full-set association baseline across the studied task subsets. A controlled prompt-based behavioral intervention on a novel dataset provides out-of-distribution validation (p = .013, η2p = .071). The present study focuses on the Qwen3 family under fixed-round Self-Refine.

Key facts. No citing papers in the map yet (data through 7 Oct 2026). Builds on 34 of its 71 references that are in the map. First posted on arXiv 6 Feb 2026.

Builds on

34 of 71 references.

References · 34

  1. 2021Measuring Mathematical Problem Solving With the MATH Datasetcited by 5,203
  2. 2023Self-Refine: Iterative Refinement with Self-Feedbackcited by 3,317
  3. 2023Qwen Technical Reportcited by 3,273
  4. 2019Invariant Risk Minimizationcited by 1,832
  5. 2020COMET: A Neural Framework for MT Evaluationcited by 1,024
  6. 2023Large Language Models Cannot Self-Correct Reasoning Yetcited by 965
  7. 2023Teaching Large Language Models to Self-Debugcited by 963
  8. 2023CRITIC: Large Language Models Can Self-Correct with Tool-Interactive Critiquingcited by 751
  9. 2015Causal inference using invariant prediction: identification and confidence intervalscited by 533
  10. 2021Causal Abstractions of Neural Networkscited by 401
  11. 2025LLMs Get Lost In Multi-Turn Conversationcited by 400
  12. 2023Language Models can Solve Computer Taskscited by 371
22 more
  1. 2020Deep Learning Enabled Semantic Communication Systemscited by 282
  2. 2024When Can LLMs Actually Correct Their Own Mistakes? A Critical Survey of Self-Correction of LLMscited by 230
  3. 2023REFINER: Reasoning Feedback on Intermediate Representationscited by 203
  4. 2021Causal Inference in Natural Language Processing: Estimation, Prediction, Interpretation and Beyondcited by 177
  5. 2024A Survey on Uncertainty Quantification of Large Language Models: Taxonomy, Open Research Challenges, and Future Directionscited by 115
  6. 2024Pride and Prejudice: LLM Amplifies Self-Bias in Self-Refinementcited by 111
  7. 2023Causal-learn: Causal Discovery in Pythoncited by 108
  8. 2024Chain of Thoughtlessness? An Analysis of CoT in Planningcited by 100
  9. 2023Shepherd: A Critic for Language Model Generationcited by 96
  10. 2021Invariant Information Bottleneck for Domain Generalizationcited by 86
  11. 2022Robust Semantic Communications with Masked VQ-VAE Enabled Codebookcited by 53
  12. 2020Active Invariant Causal Prediction: Experiment Selection through Stabilitycited by 37
  13. 2023Causal Interpretation of Self-Attention in Pre-Trained Transformerscited by 30
  14. 2023Assumption violations in causal discovery and the robustness of score matchingcited by 28
  15. 2023RELIC: Investigating Large Language Model Responses using Self-Consistencycited by 23
  16. 2025BOUQuET: dataset, Benchmark and Open initiative for Universal Quality Evaluation in Translationcited by 21
  17. 2025Auditing Meta-Cognitive Hallucinations in Reasoning Large Language Modelscited by 19
  18. 2023Causal conditional hidden Markov model for multimodal traffic predictioncited by 14
  19. 2023On Learning Necessary and Sufficient Causal Graphscited by 14
  20. 2016An Empirical-Bayes Score for Discrete Bayesian Networkscited by 9
  21. 2017Dirichlet Bayesian Network Scores and the Maximum Relative Entropy Principlecited by 4
  22. 2022Towards Dynamic Causal Discovery with Rare Events: A Nonparametric Conditional Independence Testcited by 1

Built on by

Not cited yet.