1. Problem Statement

Large Reasoning Models (LRMs) have demonstrated impressive performance on complex tasks by generating step-by-step reasoning traces (Chain-of-Thought). However, a critical reliability issue persists: unfaithful reasoning.

• The Core Issue: LRMs often produce rationales that sound plausible and coherent to human readers but do not reflect the model's actual internal decision-making process. The stated reasoning may be a post-hoc justification rather than the true causal driver of the final answer.
• The Risk: In high-stakes domains (medicine, law, coding), this discrepancy can lead to over-reliance on AI, hidden biases, and unsafe deployments.
• The Gap: Current evaluation metrics rely heavily on accuracy. The paper argues that accuracy is neither a sufficient nor a reliable proxy for faithfulness; a model can be "faithfully wrong" (coherent reasoning leading to a wrong answer) or "unfaithfully correct" (silent corrections where the reasoning contradicts the final answer).

2. Methodology

The authors propose a formal framework and a new benchmark to measure reasoning faithfulness through output-level counterfactual interventions.

A. Formal Definition of Reasoning Faithfulness

The paper defines faithfulness via two testable, decoupled conditions:

1. Stance Consistency ($\chi$): The output must form a cohesive argumentative chain. The stance (e.g., the intermediate conclusion or final answer) must remain consistent throughout the reasoning trace ($r$), explanation ($e$), and final answer ($a$), unless a change is explicitly identified and justified.
   • Failure Mode: Self-contradiction or "silent corrections" where the model changes its mind without acknowledging the shift.
2. Causal Influence ($\kappa$): The stated reasoning must causally drive the output. This is tested by injecting a counterfactual reasoning trace ($r'$) that asserts a stance opposite to the model's baseline.
   • Test: If the model is faithful, injecting $r'$ should cause a coherent shift in the model's subsequent reasoning ($r_{new}$) and/or final answer ($a'$). If the model ignores the injected premise or shifts its internal stance without changing the answer (or vice versa) without justification, it lacks causal influence.

Unified Metric: A model is deemed Reasoning-Faithful (RF) on an instance only if:

• The baseline output is stance-consistent.
• The intervened output is stance-consistent.
• The intervention caused a causal shift (either in reasoning or answer).

B. The RFEval Benchmark

To operationalize this, the authors created RFEval, a benchmark containing 7,186 instances across 7 diverse tasks:

• Tasks: Code Generation, Mathematical Reasoning, Logical Reasoning, Table Reasoning, Context Understanding, Legal Decision, and Paper Review.
• Construction Pipeline:
  1. Generation: Used a strong LLM (OpenAI o3) to generate counterfactual reasoning ($r'$) for each problem. These traces are designed to be plausible but contain subtle, specific logical flaws intended to steer the model toward a specific incorrect answer.
  2. Validation: A two-stage process involving an automated LLM validator (gpt-5) and human review (8 graduate students) to ensure the flaws are subtle, logically sound (internally coherent despite the error), and uniquely lead to one incorrect option.
• Evaluation Protocol:
  • Models are prompted with the original problem.
  • A second prompt appends the counterfactual reasoning $r'$ as if it were the model's own thought process.
  • The model's subsequent output is analyzed for stance consistency and causal influence.

3. Key Contributions

1. Formal Framework: Established a rigorous, model-agnostic definition of reasoning faithfulness based on stance consistency and causal influence, explicitly decoupling it from task accuracy.
2. RFEval Benchmark: Introduced the first large-scale benchmark (7k+ instances) specifically designed to probe faithfulness via controlled, output-level counterfactual interventions.
3. Empirical Insights: Conducted the first large-scale study of 12 open-source LRMs, revealing systematic patterns of unfaithfulness.
4. Accuracy vs. Faithfulness: Provided empirical evidence that accuracy is a poor proxy for faithfulness, showing that the correlation between the two is weak and statistically insignificant when controlling for model and task.

4. Key Results

The evaluation of 12 competitive open-source LRMs (including Qwen3, DeepSeek-R1, gpt-oss, MiMo, etc.) yielded the following findings:

• Prevalence of Unfaithfulness: 49.7% of evaluated outputs were unfaithful.
• Primary Failure Mode: The dominant cause of unfaithfulness is stance inconsistency (specifically in the intervened output), not a lack of causal influence. Models often fail to maintain a coherent stance when confronted with flawed premises.
• Task Dependency:
  • Brittle Domains: Unfaithfulness is most pronounced in Mathematical Reasoning (28.06% RF) and Code Generation (24.18% RF). These "convergent" tasks require strict logical steps; a single error often forces a silent correction or a break in the chain.
  • Flexible Domains: Tasks like Legal Decision (70.17% RF) and Paper Review show higher faithfulness, likely because they allow for multiple defensible argumentative paths.
• Impact of Training Regimes (Ablation Studies):
  • Scale: Increasing parameter size (e.g., from 8B to 32B or 70B) does not reliably improve faithfulness. In some families (e.g., gpt-oss), larger models performed worse.
  • Post-Training: Supervised Fine-Tuning (SFT) tends to preserve or slightly improve faithfulness. However, adding RL-style objectives (RLVR) on top of SFT consistently degrades faithfulness.
    • Hypothesis: Current RL rewards prioritize surface-level correctness and format over the structural integrity of the reasoning chain, encouraging models to produce "reasoning shells" that look correct but lack causal alignment.
• Accuracy vs. Faithfulness:
  • The correlation between accuracy and faithfulness is weak ($r \approx 0.09$) and statistically insignificant after controlling for model and task effects.
  • High accuracy does not guarantee faithful reasoning, and low accuracy does not preclude faithful reasoning.

5. Significance and Implications

• Trustworthy AI: The paper argues that for AI to be truly trustworthy, we must optimize for the structural integrity of the reasoning process, not just the correctness of the final outcome.
• Evaluation Shift: Reliance on accuracy alone is insufficient for auditing LRMs. The community must adopt metrics that verify whether the model's stated reasoning is the actual driver of its decisions.
• Training Recommendations: The findings suggest that current Reinforcement Learning (RL) objectives may be detrimental to reasoning faithfulness. Future training paradigms should explicitly reward stance consistency and causal alignment, not just final answer correctness.
• Safety: In high-stakes applications, unfaithful reasoning poses a severe safety risk, as users may be misled by confident but logically disconnected explanations. RFEval provides a tool to audit these risks before deployment.

In summary, RFEval exposes a critical blind spot in current LLM evaluation: models can be "right for the wrong reasons" or "wrong for the right reasons" in ways that standard benchmarks miss. The paper calls for a paradigm shift toward evaluating the causal link between reasoning and output.