Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

RFEval: Benchmarking Reasoning Faithfulness under Counterfactual Reasoning Intervention in Large Reasoning Models

Yunseok Han, Yejoon Lee, Jaeyoung Do

TL;DR

RFEval introduces a formal framework for reasoning faithfulness based on stance consistency and causal influence, and presents a benchmark of $7{,}186$ instances across seven tasks to evaluate this property via output-level counterfactual interventions. Across $12$ open-source LRMs, unfaithfulness is pervasive, driven mainly by stance inconsistency, with sharper weaknesses in brittle domains like mathematics and coding; post-training RL-style objectives often degrade faithfulness, while simply increasing model size does not guarantee improvement. The study demonstrates that accuracy is neither necessary nor sufficient as a proxy for faithfulness, underscoring the need to report faithfulness alongside accuracy. It also provides a scalable methodology for auditing LRM reliability and calls for training objectives that promote a coherent, causally influential reasoning process, guiding safer and more trustworthy AI deployment.

Abstract

Large Reasoning Models (LRMs) exhibit strong performance, yet often produce rationales that sound plausible but fail to reflect their true decision process, undermining reliability and trust. We introduce a formal framework for reasoning faithfulness, defined by two testable conditions: stance consistency (a coherent stance linking reasoning to answer) and causal influence (the stated reasoning causally drives the answer under output-level interventions), explicitly decoupled from accuracy. To operationalize this, we present RFEval, a benchmark of 7,186 instances across seven tasks that probes faithfulness via controlled, output-level counterfactual interventions. Evaluating twelve open-source LRMs, we find unfaithfulness in 49.7% of outputs, predominantly from stance inconsistency. Failures are concentrated in brittle, convergent domains such as math and code, and correlate more with post-training regimes than with scale: within-family ablations indicate that adding current RL-style objectives on top of supervised fine-tuning can reduce reasoning faithfulness, even when accuracy is maintained. Crucially, accuracy is neither a sufficient nor a reliable proxy for faithfulness: once controlling for model and task, the accuracy-faithfulness link is weak and statistically insignificant. Our work establishes a rigorous methodology for auditing LRM reliability and shows that trustworthy AI requires optimizing not only for correct outcomes but also for the structural integrity of the reasoning process. Our code and dataset can be found at project page: https://aidaslab.github.io/RFEval/}{https://aidaslab.github.io/RFEval/

RFEval: Benchmarking Reasoning Faithfulness under Counterfactual Reasoning Intervention in Large Reasoning Models

TL;DR

RFEval introduces a formal framework for reasoning faithfulness based on stance consistency and causal influence, and presents a benchmark of instances across seven tasks to evaluate this property via output-level counterfactual interventions. Across open-source LRMs, unfaithfulness is pervasive, driven mainly by stance inconsistency, with sharper weaknesses in brittle domains like mathematics and coding; post-training RL-style objectives often degrade faithfulness, while simply increasing model size does not guarantee improvement. The study demonstrates that accuracy is neither necessary nor sufficient as a proxy for faithfulness, underscoring the need to report faithfulness alongside accuracy. It also provides a scalable methodology for auditing LRM reliability and calls for training objectives that promote a coherent, causally influential reasoning process, guiding safer and more trustworthy AI deployment.

Abstract

Large Reasoning Models (LRMs) exhibit strong performance, yet often produce rationales that sound plausible but fail to reflect their true decision process, undermining reliability and trust. We introduce a formal framework for reasoning faithfulness, defined by two testable conditions: stance consistency (a coherent stance linking reasoning to answer) and causal influence (the stated reasoning causally drives the answer under output-level interventions), explicitly decoupled from accuracy. To operationalize this, we present RFEval, a benchmark of 7,186 instances across seven tasks that probes faithfulness via controlled, output-level counterfactual interventions. Evaluating twelve open-source LRMs, we find unfaithfulness in 49.7% of outputs, predominantly from stance inconsistency. Failures are concentrated in brittle, convergent domains such as math and code, and correlate more with post-training regimes than with scale: within-family ablations indicate that adding current RL-style objectives on top of supervised fine-tuning can reduce reasoning faithfulness, even when accuracy is maintained. Crucially, accuracy is neither a sufficient nor a reliable proxy for faithfulness: once controlling for model and task, the accuracy-faithfulness link is weak and statistically insignificant. Our work establishes a rigorous methodology for auditing LRM reliability and shows that trustworthy AI requires optimizing not only for correct outcomes but also for the structural integrity of the reasoning process. Our code and dataset can be found at project page: https://aidaslab.github.io/RFEval/}{https://aidaslab.github.io/RFEval/
Paper Structure (94 sections, 16 equations, 30 figures, 27 tables)

This paper contains 94 sections, 16 equations, 30 figures, 27 tables.

Table of Contents

  1. Introduction
  2. Reasoning Faithfulness
  3. RFEval: Reasoning Faithfulness Evaluation Benchmark
  4. Benchmark Design and Tasks
  5. Benchmark Construction Pipeline
  6. Counterfactual Reasoning Generation
  7. Automatic LLM Validation and Human Review
  8. Results
  9. Evaluation Settings
  10. Models
  11. Implementation details
  12. LLM-based evaluation
  13. Main Results
  14. Analysis
  15. A1. Unfaithfulness is primarily driven by stance consistency failures, not from causal breakdown.
  16. ...and 79 more sections

Figures (30)

  • Figure 1: Examples of risks arising from unfaithful reasoning in LRMs, where the stated rationale conflicts with the final output. Such discrepancies can mislead users, and jeopardize safe deployment, especially in high-stakes settings, and obscure biases.
  • Figure 2: Illustration of the RFEval evaluation workflow. In the baseline setting, the input is fed to the target LRM and the evaluator extracts stances for reasoning, explanation, and answer $(r,e,a)$ and checks flaw identification. Under intervention, counterfactual reasoning $r’$ is appended and the same procedure is applied. In this example, the baseline output is stance-consistent ($\chi(o)=1$), whereas the intervened output is stance-inconsistent ($\chi(o’)=0$) and shows no causal influence ($\kappa(o,o’)=0$) because the final stance does not change.
  • Figure 3: Overall RF scores for each model. Reasoning faithfulness varies substantially across models with very different parameter counts, indicating that scale alone is not a reliable predictor of RF and that other factors (e.g., training regime and data) play a larger role.
  • Figure 4: (Left) Composition of RF violation types ($\neg\chi(o)$, $\neg\chi(o')$, $\neg\kappa$). (Right) Row-normalized heatmaps of where stance discontinuities occur (x-axis) at baseline and under intervention.
  • Figure 5: Ratio of satisfied conditions for causal influence: "Reasoning-only" (only reasoning stance changed), "Answer-only" (only final answer stance changed), and "Both" (both changed).
  • ...and 25 more figures

Theorems & Definitions (5)

  • Definition 2.1: Canonical Stance
  • Definition 2.2: Stance-Continuous
  • Definition 2.3: Stance Consistency
  • Definition 2.4: Causal Influence
  • Definition 2.5: Reasoning Faithfulness