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

AI summary

Overview

Research area: Natural Language Processing — causal analysis of large language model self-reflection pipelines (cs.CL).

Technical level: Advanced. The paper assumes familiarity with causal discovery, Invariant Causal Prediction, score-based structure learning (Greedy Equivalence Search), and LLM prompting frameworks such as Self-Refine.

Scope: The paper introduces ReBeCA, a causal-analysis framework that models multi-round self-reflection trajectories as causal graphs and identifies the sparse, stable semantic behaviors ("causal parents") that genuinely determine the outcome of iterative self-reflection, validated in a controlled case study on the Qwen3 model family under fixed-round Self-Refine.

What This Paper Is About

Self-reflection prompting lets language models critique and revise their own outputs, but prior analyses of why it works or fails are largely correlation-based, focused on isolated design factors or single reflection steps, and do not generalize across tasks or models. ReBeCA treats self-reflection as a structured multi-round trajectory and asks which observable behaviors causally drive the final outcome, rather than merely co-occurring with it. The goal is a reusable protocol that reliably selects and self-verifies these causal drivers without relying on a ground-truth causal graph, which does not exist for this domain.

Key Contributions

1. A causal-driven analytical framework (ReBeCA) that reframes self-reflection analysis around causal parents rather than correlations, explaining the self-Reflection Behavior through Causal Analysis.
2. A reusable three-stage ICP-based causal-parent selection and verification protocol that works without ground-truth causal graphs, combining Ensemble Causal Discovery, approximate structural uniqueness validation, and linear stability verification.
3. Consistency-Enhanced Self-Refine (CESR), a trajectory-generation procedure that multi-samples every intermediate feedback and refine step, clusters outputs by semantic equivalence, and selects the largest cluster's centroid as the representative branch to improve statistical significance of each trajectory.
4. A controlled Qwen3 case study plus a prompt-based intervention on AIME that empirically tests the identified causal parents out of distribution, yielding three findings about behavioral hierarchy, spurious correlation, and non-additivity.

Main Findings

• Behavioral hierarchy: Semantic behaviors influence the final self-reflection outcome hierarchically, either directly or indirectly. The effect is time-dependent: "Clarity and Organization" is causally decisive only at round t=2 and "Specificity" only at round t=4 for MATH, while for BOUQuET "Clarity and Organization" at round t=3 is the optimal causal parent for both outcome indicators. This contradicts the notion that uniform criteria apply across all rounds.
• Causation matters (correlations can be spurious): Only a sparse set of semantic behaviors genuinely affects the outcome. Stage 1 identified 12 independent candidate causal parent sets for MATH, and six each for the BOUQuET COMET and MetricX outcomes.
• Sparse parents beat dense association baselines: The selected parent set achieved a 22.36% CVLL gain over the empty set and a 49.59% gain over the full set on MATH; 6.32% and 15.10% respectively on BOUQuET (COMET); and 3.93% and 12.09% respectively on BOUQuET (MetricX). The abstract reports sparse causal parents achieving up to 49.6% structural likelihood gains relative to the dense full-set association baseline.
• Optimal CVLL values: MATH L*_BDs = −0.434 ± 0.563; BOUQuET L*_BDs|COMET = −0.978 ± 0.495 and L*_BDs|MetricX = −1.003 ± 0.536. Corresponding full-set values were −0.861 ± 1.227 (MATH), −1.152 ± 0.784 (COMET), and −1.141 ± 0.763 (MetricX).
• Linear stability mostly holds: Except for a single moderate effect, Levene's Test and One-Way ANOVA confirmed that coefficients and residuals of the causal parents display no substantial variation across the 20 re-stratified folds organized into 4 balanced groups of 5.
• More ≠ better (non-additivity): In the intervention study on 50 AIME problems across four LMs, targeting Clarity and Organization at t=2 or Specificity at t=4 individually improved results, while activating both jointly yielded no additive gain. On average each single-pattern condition reached 16 correct responses per model; excluding Qwen3-0.6B, single-pattern conditions averaged 21.0/50 versus 17.33/50 without intervention and 17.0/50 under joint activation.
• Out-of-distribution intervention validation: Cochran's Q test on the intervention produced p = .013 and η²p = .071, supporting the hypothesis that intervention groups diverge significantly in self-reflection outcome.

Methodology in Plain English

The researchers first build a reliable record of what a model does across multiple rounds of self-critique and revision. Because language model outputs are stochastic, a single trajectory is unrepresentative, so each feedback and refine step is sampled 20 times, the samples are grouped by semantic similarity using OpenAI's Text-Embedding-3-Large with cosine similarity and HDBSCAN, and the centroid of the largest cluster is kept as the representative output. This is CESR, a linear-time approximation that avoids enumerating an exponential number of possible trajectories.

Next, unstructured outputs are converted into structured data using semantic patterns — fixed-length binary vectors where 1 means a pattern is present and 0 means it is absent. Candidate patterns are mined with GPT-5 from a random subset of trajectories and then manually refined by removing patterns occurring under 5% of the time and consolidating overlapping ones. This binary format is chosen over embeddings for interpretability and because causal discovery algorithms scale poorly with dimensionality. After filtering, four candidate patterns remain for MATH and five for BOUQuET.

The heart of the method is a three-stage pipeline built on Invariant Causal Prediction, which relies on the idea that valid causal structures stay stable across distributions:

1. Stage 1 — select the optimal parent set. The data are split into five cross-validation folds; Ensemble Causal Discovery using Greedy Equivalence Search with four different score functions produces up to K × Q candidate parent sets; candidates are ranked by Cross-Validation Log-Likelihood using the Bayesian Dirichlet sparse (BDs) score, with temporal constraints supplied as adjacency-matrix masks via Causal-Learn.
2. Stage 2 — approximate structural uniqueness. The selected set must beat both the empty set and the full set on CVLL, giving boundary evidence that the chosen configuration is not arbitrary.
3. Stage 3 — linear stability verification. Assuming y ~ k · Pa_y^T + b, Levene's Test and One-Way ANOVA check whether coefficients and residuals stay stable across folds (H0 supported when p > .05 and η²p < .06).

Finally, a prompt-based intervention on 50 AIME problems tests causality directly: if a semantic pattern truly causes the outcome, instructing the model to emphasize it during a specific round should change the result. A 2 × 2 factorial design over CO at t=2 and Sp at t=4, evaluated with Cochran's Q, produced the significance result above. The paper reports that the study covers the Qwen3 family under fixed-round Self-Refine.

Why This Matters

Impact on research: The paper shifts self-reflection analysis from empirical "prompt engineering" to "causal behavior verification," and argues that generalizability — not accuracy — should be the bar, because high-accuracy claims on heterogeneous, uncertain self-reflection trajectories are vulnerable to spurious correlations. It also supplies a protocol that does not depend on ground-truth causal graphs, which do not exist in this domain. The case study data are released at https://tinyurl.com/ReBeCAStudy.

Real-world applications (implications drawn from the framework):

• Prompt and pipeline design for self-refining agents: Rather than tuning round counts or model size by trial and error, developers could target the specific behavior at the specific round that causally matters.
• Auditing and evaluating language model reliability: Distinguishing stable causal drivers from transient correlations gives an evaluation signal that is claimed to hold across analogous tasks.
• Multi-agent and tool-augmented reflection systems: The intervention design offers a way to test whether an added tool or feedback channel actually changes outcomes before deploying it.
• Multilingual and reasoning deployments: The study spans logical reasoning (MATH) and multilingual capacity (BOUQuET), suggesting the analysis can be applied to more than one task type.

Industry relevance: Any product that relies on iterative self-critique — code generation, agentic workflows, high-stakes question answering — pays a cost per reflection round. If stacking several seemingly positive behaviors yields no additive gain, effort and inference budget can be reallocated toward the sparse behaviors that actually move outcomes.

Future Directions

• Cross-family and cross-protocol evaluation. The current study covers four Qwen3 scales on two task types under fixed-round Self-Refine; the paper states that exact parent patterns may vary with model family, task, and reflection protocol, and that broader cross-family and adaptive-trajectory evaluation remains future work.
• Principled resampling. CESR applies a fixed resampling rate of 20 iterations per intermediate output; the paper notes no existing research offers a robust method for determining the optimal resampling count under arbitrary conditions, and calls for one.
• Automating semantic-pattern discovery. Candidate patterns are proposed by GPT-5 and manually refined, which risks overlooking critical but underexplored variables; the paper discusses Bag-of-Words plus unsupervised word embedding clustering as an alternative with notable limitations and calls for better automated methods that preserve interpretability.
• Ground-truth validation and confounding. Robust validation ultimately requires reliable ground truths that are currently non-existent for self-reflection, and the interpretation is conditional on observed-variable, faithfulness, and invariance assumptions — with hidden confounding, a selected pattern may proxy an unobserved mechanism.

Target Audience

Researchers and advanced practitioners in NLP and causal inference who study language model reasoning, self-correction, and agent behavior; engineers building self-refining or agentic LLM pipelines who want principled guidance instead of trial-and-error prompt tuning; and methodologists interested in applying causal discovery and invariance-based validation to behavioral phenomena in large models. The paper is written at an advanced technical level and assumes comfort with causal graphs, score-based structure learning, and statistical validation tests.

Authors’ abstract

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 ($p = .013, η^2_\mathrm{p} = .071$). The present study focuses on the Qwen3 family under fixed-round Self-Refine.