Skip to content

RFEval: Benchmarking Reasoning Faithfulness under Counterfactual Perturbations

Conference: ICLR 2026 arXiv: 2602.17053 Code: AIDASLab/RFEval Area: Causal Reasoning Keywords: reasoning_faithfulness, LRM_evaluation, counterfactual_intervention, benchmark

TL;DR

This paper proposes a formal framework for reasoning faithfulness (stance consistency + causal influence) and the RFEval benchmark (7,186 instances × 7 tasks). By applying output-level counterfactual interventions to evaluate 12 open-source LRMs, it finds that 49.7% of outputs are unfaithful and that accuracy is not a reliable proxy for faithfulness.

Background & Motivation

Large reasoning models (LRMs) such as DeepSeek-R1 and Qwen3 achieve strong performance on complex tasks, yet frequently produce plausible-sounding but unfaithful explanations—i.e., the stated reasoning does not reflect the model's actual decision process.

Core Problem: - When a model claims "I chose A because of X," is X truly the cause of choosing A? - In high-stakes domains such as medicine, law, and HR, unfaithful explanations can mislead users and obscure biases. - Existing evaluations focus primarily on accuracy, but accuracy does not equate to faithfulness.

Limitations of Prior Work: - Internal activation analysis (e.g., probing methods) requires model access and does not scale. - No systematic behavioral-level framework for faithfulness evaluation exists. - No unified benchmark enables cross-LRM comparison of reasoning faithfulness.

Method

Overall Architecture

This paper defines reasoning faithfulness at the behavioral level, assessed purely from the model's textual outputs without access to internal weights. The core idea: if reasoning is faithful, then changing the reasoning should change the answer.

Key Designs: Two Testable Conditions

Condition 1: Stance Consistency

For an LRM output \(o = (r, e, a)\) (reasoning, explanation, answer), stance consistency requires the entire output sequence to form a coherent argumentative chain:

\[\chi(o) := \bigwedge_{i=1}^{m} \iota(\langle c_{1:i-1}\rangle, c_i) \in \{0, 1\}\]

where \(\iota(u, v)\) is a stance continuity indicator: \(\iota = 1\) if and only if the stance of \(v\) is consistent with \(u\), or \(v\) explicitly identifies and justifies a departure.

Condition 2: Causal Influence

Given the model's original output \(o\) and the post-intervention output \(o'\) after injecting counterfactual reasoning \(r'\), causal influence requires that the reasoning or answer changes:

\[\kappa(o, o') := \mathbb{1}[S(r_{\text{new}}) \neq S(r)] \lor \mathbb{1}[S(a') \neq S(a)]\]

Unified Definition: Reasoning Faithfulness

\[\text{RF}(o, o') := \mathbb{1}[\chi(o) = 1 \land \chi(o') = 1 \land \kappa(o, o') = 1]\]

That is, both the original and intervened outputs must be stance-consistent, and the intervention must produce a genuine causal effect.

Contrastive Pre-condition

To ensure causal identifiability, evaluation is restricted to contrastive pairs (\(\delta = 1\)), where the injected counterfactual reasoning \(r'\) takes a stance opposite to the model's original stance. This eliminates ambiguity from "no-change" outcomes.

Benchmark Construction Pipeline

Counterfactual Reasoning Generation: Generated using OpenAI o3, with each prompt including 3 manually crafted few-shot examples to elicit subtle yet plausible reasoning flaws (e.g., arithmetic errors, logical fallacies).

Two-Stage Validation: 1. GPT-5 automatic filtering: checks for misleading sufficiency, logical coherence, subtle plausibility, and uniqueness (MCQA). 2. Manual review by 8 NLP/ML graduate students: PABAK = 0.710, reducing 8,499 candidates to 7,186 instances.

Evaluation Implementation: o3 serves as the evaluator for stance extraction, with human-validated F1 = 0.952.

Loss & Training

This paper presents an evaluation framework and involves no training loss. The core metric is:

\[\text{RF}^{\text{contrast}}(\mathcal{M}, \mathcal{D}) = \mathbb{E}\left[\text{RF}(o, o') \mid \delta(x, r'; \mathcal{M}) = 1\right]\]

along with contrastive coverage \(c(\mathcal{M}) = \Pr(\delta = 1)\).

Key Experimental Results

Main Results: Reasoning Faithfulness of 12 LRMs

Model Overall RF (%) Coverage CG MR LR TR CU LD PR
Qwen3-32B 73.29 0.78 24.66 47.87 88.62 89.84 77.66 89.90 91.49
LN-Super_v1 68.52 0.58 26.48 44.90 77.13 69.38 81.70 80.38 98.47
R1-Qwen-32B 64.24 0.75 29.02 32.57 70.79 82.47 63.16 91.04 75.13
R1-Qwen-7B 61.37 0.70 38.25 29.54 82.13 44.46 76.31 70.63 81.49
MiMo-RL-Zero 58.74 0.54 20.83 33.50 70.59 61.32 69.58 77.87 66.83
R1-Llama-70B 56.47 0.78 27.89 31.28 74.03 73.78 51.40 80.53 51.84
gpt-oss-20b 32.11 0.82 26.44 24.90 13.55 22.62 33.93 59.14 47.41
gpt-oss-120b 27.50 0.82 22.01 16.07 8.62 34.21 13.67 39.58 70.71

Key finding: 49.7% of evaluated instances are unfaithful. Even the best-performing model, Qwen3-32B, achieves only 73.29%.

Ablation Study: Sources of Unfaithfulness

Violation Type Proportion Description
\(\neg\chi(o')\) (post-intervention stance inconsistency) Dominant Model fails to respond coherently to the counterfactual premise
\(\neg\kappa\) (no causal influence) Secondary Reasoning changes but the answer does not follow
\(\neg\chi(o)\) (baseline stance inconsistency) Minor Original output is internally contradictory

Causal Influence Types: - Most models exhibit "Both" (reasoning and answer both change). - The gpt-oss series and Magistral-Small show more "Reasoning-only" cases (reasoning changes but answer does not). - Some Qwen/R1 models exhibit "Answer-only" cases (silent corrections—answer changes without the reasoning reflecting it).

Key Findings

  1. Task structure determines faithfulness: Math and code tasks (convergent, unique answers) show the highest unfaithfulness rates; legal and paper review tasks (supportive of multi-perspective argumentation) show the highest faithfulness.
  2. Scale does not determine faithfulness: gpt-oss faithfulness decreases from 20B to 120B (32.11% → 27.50%), while Qwen3 improves substantially from 8B to 32B (41.95% → 73.29%).
  3. Post-training paradigm is the key factor: Within the same model family, RLVR-style post-training may reduce faithfulness even when accuracy remains unchanged.
  4. Accuracy ≠ Faithfulness: After controlling for model and task, the correlation between accuracy and faithfulness is weak and statistically insignificant. High accuracy does not guarantee faithful reasoning.
  5. Failure locations exhibit family-specific patterns: The gpt-oss series breaks early in the intervention chain (\(r' \to r_{\text{new}}\)); Qwen/R1 models fail more often at the later stage (\(r_{\text{new}} \to a'\)).

Highlights & Insights

  • Elegant formal framework: Reasoning faithfulness is decomposed into two independently testable conditions (consistency + causality), making the framework both rigorous and operationalizable.
  • Clever counterfactual intervention design: By injecting opposing reasoning at the output level, the method avoids the need for internal model access.
  • Most important finding: RL post-training can reduce faithfulness without degrading accuracy—a cautionary signal for the current RLVR trend.
  • Practical value: An open-source benchmark of 7,186 instances and an evaluation framework are provided, directly applicable to LRM auditing.

Limitations & Future Work

  1. Only open-source models are evaluated; closed-source API models (e.g., GPT-5.2, Claude) are difficult to subject to standard interventions due to response integrity mechanisms.
  2. The framework relies on an LLM evaluator (o3) for stance extraction; although F1 reaches 0.952, this is not perfect.
  3. Counterfactual reasoning is generated by o3, which may not cover all types of reasoning flaws.
  4. Evaluation operates at the coarse-grained \((r, e, a)\) level, without fine-grained step-by-step analysis of the reasoning chain.
  5. Contrastive coverage—particularly in the Paper Review task, averaging only 0.35–0.45—means that a substantial proportion of instances are excluded due to stance alignment.
  • Jacovi & Goldberg (2020): Early conceptual framework defining faithful explanations.
  • Chen et al. (2025b); Arcuschin et al. (2025): Empirical evidence of unfaithful reasoning in LRMs.
  • Lanham et al. (2023): CoT faithfulness study, but not based on counterfactual intervention.
  • This paper's contributions lie in: (1) a formal definition, (2) large-scale systematic evaluation, and (3) characterization of the relationship between training paradigms and faithfulness.
  • Implication for LRM deployment: reporting accuracy alone is insufficient; faithfulness should be reported concurrently.

Rating

  • Novelty: ⭐⭐⭐⭐ — The formal framework and counterfactual intervention methodology are novel.
  • Experimental Thoroughness: ⭐⭐⭐⭐⭐ — 12 models × 7 tasks × 7,186 instances; large-scale and systematic.
  • Value: ⭐⭐⭐⭐ — The open-source benchmark is directly applicable to LRM auditing.
  • Writing Quality: ⭐⭐⭐⭐ — Formal definitions are clear, though the density of notation requires careful reading.
  • Overall: ⭐⭐⭐⭐ (4/5)