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

Yunseok Han; Yejoon Lee; Jaeyoung Do

cs.AIFeb 19, 2026

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

Authors: Yunseok Han, Yejoon Lee, Jaeyoung Do

Organizations: AIDAS Laboratory, 1IPAI & 2ECE, Seoul National University

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/

Figures & tables

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.

Task	Count	Sources & Brief Description
Code Generation	861	LiveCodeBench ( Jain et al., 2024 ) , DS-1000 ( Lai et al., 2023 ) ;
		Generate the source code to solve the problem.
Mathematical Reasoning	1,029	MMLU ( Hendrycks et al., 2020 ) , GSM8K ( Cobbe et al., 2021 ) ;
		Select the answer option or generate exact answer for the problem.
Logical Reasoning	1,107	PrOntoQA ( Saparov and He, 2022 ) , RuleBert-Union-Rules ( Saeed et al., 2021 ) ;
		Select T/F whether the proposition is satisfied by given premises.

Table 1: Overview of seven tasks included in RFEval with sample counts, source datasets, and a brief description of objective. A detailed description of the source dataset is presented in Appendix B.1 .

	CG		MR		LR		TR		CU		LD		PR		Overall
Model	RF	$c(\mathcal{M})$	RF	$c(\mathcal{M})$	RF	$c(\mathcal{M})$	RF	$c(\mathcal{M})$	RF	$c(\mathcal{M})$	RF	$c(\mathcal{M})$	RF	$c(\mathcal{M})$	RF	$c(\mathcal{M})$
Qwen3-8B	21.15	0.73	37.97	0.97	72.74	1.00	58.11	0.99	43.97	0.97	48.64	0.78	*3.09	0.96	41.95	0.92
Qwen3-32B	24.66	0.69	47.87	0.96	88.62	0.82	89.84	0.85	77.66	0.96	89.90	0.80	91.49	0.39	73.29	0.78
R1-Qwen-7B	38.25	0.45	29.54	0.91	82.13	0.75	44.46	0.68	76.31	0.93	70.63	0.69	81.49	0.41	61.37	0.70
R1-Qwen-32B	29.02	0.60	32.57	0.94	70.79	0.78	82.47	0.80	63.16	0.97	91.04	0.78	75.13	0.36	64.24	0.75
R1-Llama-8B	26.48	0.54	33.03	0.74	55.78	0.71	57.68	0.65	64.63	0.94	78.97	0.73	94.53	0.36	58.46	0.67

Table 2: Contrast-conditional reasoning faithfulness (RF, %) and contrast coverage (

c(\mathcal{M})

) on RFEval. Presented tasks are CG (Code Generation), MR (Mathematical Reasoning), LR (Logical Reasoning), TR (Table Reasoning), CU (Context Understanding), LD (Legal Decision), and PR (Paper Review).

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.

Panel of Figure 4 — Figure 4: (Left) Composition of RF violation types ( $\neg\chi(o)$ , $\neg\chi(o^{\prime})$ , $\neg\kappa$ ). (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).

Variant	MiMo-7B	Olmo-3-7B
Base	59.33 / 0.69	65.87 / 0.42
SFT-only	60.05 / 0.74	61.38 / 0.70
RL-only	58.74 / 0.54	–
SFT+RL	46.32 / 0.72	50.93 / 0.73

Table 3: Within-family ablations on post-training schemes (

\textsc{RF}^{\text{contrast}}

, % /

c(\mathcal{M})

	R1-Qwen-7B	R1-Qwen-32B	R1-Llama-8B	R1-Llama-70B	Overall
$\chi=1$	0.6804	0.3996	0.7538	0.6868	0.6280
$\chi=0$	0.7200	0.3789	0.8354	0.6634	0.6711

Table 4: Average reasoning-step reward stratified by stance consistency

\chi

Figure 6: Scatter plot of Acc. vs. RF per (model, task) with WLS and fixed-effects trend.

Appendix figures & tables44 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 7: Human Review interface. The left panel shows the task selection page; the right panel shows a validation instance. Annotators read the problem, options, ground truth, generated counterfactual reasoning, and the model-provided flaw explanation (for validation only), then decide whether all criteria are satisfied.

Task	$P_{a}$	PABAK	Fleiss’ $\kappa$	Krippendorff’s $\alpha$
Code Generation	0.500	0.000	-0.099	-0.044
Mathematical Reasoning	0.700	0.400	- 0.200	- 0.240
Logical Reasoning	0.900	0.800	-0.053	- 0.000
Table Reasoning	0.900	0.800	-0.053	- 0.000
Context Understanding	1.000	1.000	NaN	NaN
Legal Decision	1.000	1.000	NaN	NaN

Appendix

Table 5: Agreement by task on double-annotated items.

P_{a}

denotes percent agreement; PABAK

=2P_{a}-1

. NaN indicates insufficient variability for

\kappa/\alpha

on that task.

Task	$n$	# yes	Yes rate	Wilson 95% CI
Code Generation	20	13	0.650	[0.433, 0.819]
Mathematical Reasoning	20	15	0.750	[0.531, 0.888]
Logical Reasoning	20	19	0.950	[0.764, 0.991]
Table Reasoning	20	19	0.950	[0.764, 0.991]
Context Understanding	20	20	1.000	[0.839, 1.000]
Legal Decision	20	20	1.000	[0.839, 1.000]

Appendix

Table 6: Task-level valid rate (“yes”) with Wilson 95% confidence intervals.

Metric	Value
Overall percent agreement ( $P_{a}$ )	0.855
Overall PABAK	0.710
Overall Fleiss’ $\kappa$	0.205
Overall Krippendorff’s $\alpha$	0.211
Overall yes rate	0.899
Overall yes rate Wilson 95% CI (low)	0.838

Appendix

Table 7: Overall human-review quality summary on double-annotated items.

Task	Pre Total	# Removed	# Kept
Code Generation	1,343	482	861
Mathematical Reasoning	1,170	141	1,029
Logical Reasoning	1,200	93	1,107
Table Reasoning	1,200	261	939
Context Understanding	1,200	107	1,093
Legal Decision	1,200	118	1,082

Appendix

Table 8: Filtering statistics by task.

Task	Source	Pre Total	# Removed	# Kept
Code Generation	DS-1000	294	71	223
	LiveCodeBench (v1–v6)	1,049	411	638
Mathematical Reasoning	GSM8K	800	81	719
	MMLU (College Math)	100	14	86
	MMLU (High School Math)	270	46	224
Logical Reasoning	PrOntoQA	500	13	487

Appendix

Table 9: Filtering statistics by source dataset. LiveCodeBench is aggregated over v1–v6.

Task	Count	$\overline{E}$	$\sigma$	$q_{50}$	$q_{75}$	$q_{90}$	Small-vocab (%)
Code Generation	12,319	0.227	0.093	0.222	0.283	0.338	0.0
Mathematical Reasoning	14,406	0.135	0.063	0.129	0.175	0.222	0.0
Logical Reasoning	13,284	0.176	0.037	0.175	0.198	0.223	0.0
Table Reasoning	12,207	0.308	0.089	0.291	0.364	0.437	0.0
Context Understanding	14,209	0.546	0.086	0.558	0.605	0.639	0.0
Legal Decision	14,066	0.424	0.103	0.429	0.492	0.554	0.0

Appendix

Table 10: Externality Penalty

E(r^{\prime})

by task (lower is more local). Small-vocab = share of instances with

|\mathcal{V}_{x}|\leq 3

|\mathcal{V}_{r^{\prime}}|\leq 3

Model	Count	$\overline{E}$	$\sigma$
Qwen3-8B	7,186	0.394	0.269
Qwen3-32B	7,186	0.394	0.269
R1-Qwen-7B	7,186	0.394	0.269
R1-Qwen-32B	7,186	0.388	0.267
R1-Llama-8B	7,186	0.394	0.269
R1-Llama-70B	7,186	0.394	0.269

Appendix

Table 11: Externality Penalty

E(r^{\prime})

by model (lower is more local).

Model	H100 hrs	Model	H100 hrs
Qwen3-8B	45	gpt-oss-20b	60
Qwen3-32B	244	gpt-oss-120b	66
R1-Qwen-7B	56	MiMo-RL-Zero	54
R1-Qwen-32B	236	MiMo-RL	126
R1-Llama-8B	43	Magistral-Small	306
R1-Llama-70B	240	LN-Super_v1	66

Appendix

Table 12: Total inference time required to generate responses for all tasks in RFEval for each model. The time is reported in NVIDIA H100 GPU hours.

Model	Total	Included	Global Exclusion	Unfinished/Truncated	Evaluation Error	Non-Contrast
Qwen3-8B	7,186	5,543	1,075	236	1	331
Qwen3-32B	7,186	5,624	0	97	1	1,464
R1-Qwen-7B	7,186	4,937	0	476	1	1,772
R1-Qwen-32B	7,186	5,294	0	432	1	1,459
R1-Llama-8B	7,186	4,820	0	74	2	2,290
R1-Llama-70B	7,186	5,592	0	14	0	1,580

Appendix

Table 13: Summary of response pair usage and discard reasons, aggregated across all tasks for each model. ‘Total’ refers to the total number of problems attempted by each model.

	CG		MR		LR		TR		CU		LD		PR
Model	B	I	B	I	B	I	B	I	B	I	B	I	B	I
Qwen3-8B	6,344	8,801	1,328	941	98	111	139	63	306	40	2,621	378	27	740
Qwen3-32B	8,845	7,453	1,246	902	1,608	47	955	72	543	156	2,117	385	596	217
R1-Qwen-7B	13,403	7,976	821	658	2,072	333	697	230	684	210	1,087	346	372	151
R1-Qwen-32B	12,288	5,568	1,073	632	2,106	301	1,214	232	735	189	1,471	413	412	121
R1-Llama-8B	7,225	4,344	880	567	1,342	329	727	214	633	349	1,099	403	448	89

Appendix

Table 14: Mean token lengths of baseline and intervened responses for each model across all seven tasks. The token count reflects the entire model output, including reasoning and the final answer. ’B’ denotes the baseline response length, while ’I’ denotes the intervened response length.

Task	Primary Stance Set
Code Generation	correct, incorrect
Mathematical Reasoning	A, B, C, … or correct, incorrect
Logical Reasoning	correct, incorrect
Table Reasoning	supported, not enough info, rebutted
Context Understanding	yes, no, maybe
Legal Decision	A, B, C, …

Appendix

Table 15: Predefined primary stance sets for each task in RFEval. The “I don’t know” stance is added to every set during evaluation. This allows us to evaluate the stance of undefinitive context.

Figure 8: Human Evaluation interface. The left panel shows the task selection page, and the right panel shows an evaluation instance. Annotators read the problem, options, ground truth, and counterfactual reasoning along with the model’s generated reasoning, explanation, and answer. They then decide which stance each component refers to and whether the component explicitly identifies the flaw in the counterfactual reasoning. If no explanation is generated, it is simply omitted.

	Q1 (Low)	Q2	Q3	Q4 (High)
By-Model (weighted)	0.519	0.571	0.589	0.494
By-Task (weighted)	0.557	0.532	0.511	0.469

Appendix

Table 16: Contrastive reasoning faithfulness by coverage quartile. Weighted averages are reported across models and tasks.

Figure 9: Model-level coverage-RF relationship.

Figure 10: Task-level coverage-RF relationship.

Model	$\neg\chi(o)$	$\neg\chi(o^{\prime})$	$\neg\kappa$	Other
Qwen3-8B	0.465	0.290	0.134	0.110
Qwen3-32B	0.029	0.578	0.387	0.006
R1-Qwen-7B	0.154	0.717	0.057	0.073
R1-Qwen-32B	0.088	0.689	0.199	0.025
R1-Llama-8B	0.140	0.679	0.105	0.076
R1-Llama-70B	0.102	0.626	0.205	0.068

Appendix

Table 17: Shares contributing to

\neg\textsc{RF}

by model. Larger

\neg\chi(o^{\prime})

indicates post-intervention stance incoherence; larger

\neg\kappa

indicates stance changes that fail to causally propagate to the answer.

Model	$r{\to}e$	$e{\to}a$	$r{\to}a$
Qwen3-8B	0.252	0.044	0.704
Qwen3-32B	0.364	0.545	0.091
R1-Qwen-7B	0.416	0.579	0.004
R1-Qwen-32B	0.466	0.511	0.023
R1-Llama-8B	0.563	0.415	0.022
R1-Llama-70B	0.838	0.148	0.014

Appendix

Table 18: Where stance discontinuities occur in baselines by model. Larger

r{\to}a

indicates direct answer jumps without explicit justification; larger

e{\to}a

reflects answer–explanation mismatches; larger

r{\to}e

reflects reasoning-to-explanation misalignment.

Model	$r^{\prime}{\to}r_{\text{new}}$	$r_{\text{new}}{\to}e^{\prime}$	$e^{\prime}{\to}a^{\prime}$	$r{\to}a^{\prime}$
Qwen3-8B	0.311	0.161	0.040	0.489
Qwen3-32B	0.301	0.440	0.102	0.157
R1-Qwen-7B	0.240	0.474	0.268	0.017
R1-Qwen-32B	0.124	0.626	0.145	0.105
R1-Llama-8B	0.293	0.447	0.244	0.016
R1-Llama-70B	0.262	0.581	0.112	0.045

Appendix

Table 19: Where stance discontinuities occur under intervention by model. Larger

r^{\prime}{\to}r_{\text{new}}

indicates early failure to assimilate the injected premise; larger

r_{\text{new}}{\to}e^{\prime}

indicates late failure to maintain stance into the explanation;

e^{\prime}{\to}a^{\prime}

and

r{\to}a^{\prime}

capture breakdowns at the answer boundary.

Model	Both	Reasoning	Answer
Qwen3-8B	0.920	0.036	0.044
Qwen3-32B	0.940	0.041	0.019
R1-Qwen-7B	0.929	0.040	0.032
R1-Qwen-32B	0.962	0.020	0.018
R1-Llama-8B	0.909	0.058	0.033
R1-Llama-70B	0.930	0.038	0.032

Appendix

Table 20: Causal-influence satisfaction types by model. “Both” means reasoning stance and answer stance change together; “Reasoning” (only reasoning changes) often reflects inert answers; “Answer” (only answer changes) often reflects silent corrections.

Task	$\neg\chi(o)$	$\neg\chi(o^{\prime})$	$\neg\kappa$	Other
Code Generation	0.127	0.292	0.514	0.067
Mathematical Reasoning	0.042	0.383	0.543	0.031
Logical Reasoning	0.046	0.683	0.252	0.019
Table Reasoning	0.123	0.663	0.194	0.019
Context Understanding	0.095	0.636	0.253	0.017
Legal Decision	0.134	0.584	0.249	0.033

Appendix

Table 21: Shares contributing to

\neg\textsc{RF}

by task. Convergent tasks (CG/MR) show larger

\neg\kappa

; argumentative tasks show larger

\neg\chi(o^{\prime})

Task	$r{\to}e$	$e{\to}a$	$r{\to}a$
Code Generation	0.378	0.530	0.092
Mathematical Reasoning	0.823	0.110	0.067
Logical Reasoning	0.316	0.505	0.178
Table Reasoning	0.217	0.260	0.524
Context Understanding	0.195	0.155	0.650
Legal Decision	0.219	0.181	0.599

Appendix

Table 22: Where stance discontinuities occur in baselines by task. High

r{\to}e

indicates justification bottlenecks; high

r{\to}a

indicates answer jumps without expository linkage.

Task	$r^{\prime}{\to}r_{\text{new}}$	$r_{\text{new}}{\to}e^{\prime}$	$e^{\prime}{\to}a^{\prime}$	$r{\to}a^{\prime}$
Code Generation	0.393	0.333	0.195	0.079
Mathematical Reasoning	0.316	0.618	0.031	0.035
Logical Reasoning	0.607	0.169	0.097	0.127
Table Reasoning	0.583	0.212	0.140	0.065
Context Understanding	0.504	0.311	0.122	0.063
Legal Decision	0.768	0.108	0.079	0.046

Appendix

Table 23: Where stance discontinuities occur under intervention by task. Early (

r^{\prime}{\to}r_{\text{new}}

) vs. late (

r_{\text{new}}{\to}e^{\prime}

) failures distinguish whether models fail to assimilate or to propagate the injected stance.

Task	Both	Reasoning	Answer
Code Generation	0.733	0.222	0.045
Mathematical Reasoning	0.786	0.120	0.094
Logical Reasoning	0.784	0.194	0.022
Table Reasoning	0.836	0.144	0.020
Context Understanding	0.810	0.179	0.011
Legal Decision	0.875	0.114	0.011

Appendix

Table 24: Causal-influence satisfaction types by task. “Both” dominates overall; Code Generation shows the largest “Reasoning” (inert answers), while Mathematical Reasoning shows the largest “Answer” (silent corrections).

Figure 11: The ratio of satisfied components

\chi(o),\chi(o^{\prime}),

and

\kappa(o,o^{\prime})

across models and tasks. Note the consistently high baseline consistency

\chi(o)

compared to the intervened metrics.

Model	CG	MR	LR	TR	CU	LD	PR	Overall
Qwen3-32B	0.9561	0.9990	0.9967	1.0000	1.0000	0.9908	0.9669	0.9908
R1-Llama-70B	0.8724	0.8790	0.9806	0.9800	0.9944	0.9811	0.6598	0.9261

Appendix

Table 25: Baseline Stance Consistency

\chi(o)

scores across tasks. Models exhibit high inherent self-consistency when generating autonomously, contrasting with their lower faithfulness scores under intervention.

Figure 12: The number of faithful/unfaithful response pair across models and tasks.

Model	RF ${}^{\text{contrast}}$ (%)	$c(\mathcal{M})$	$\chi(o)$	$\chi(o^{\prime})$	$\kappa(o,o^{\prime})$
gpt-5.1-2025-11-13	13.25	0.7629	0.9847	0.1618	0.9669
claude-sonnet-4-5-20250929	86.72	0.9223	0.9979	0.8736	0.9958

Appendix

Table 26: Reasoning Faithfulness results for proprietary models on the Mathematical Reasoning task, using the adapted multi-turn protocol. Note that due to the protocol difference, these numbers are not directly comparable to the main open-source results.

Model	Metric	CG	MR	LR	TR	CU	LD	PR	Overall
R1-Llama-8B	Per-Step	9.95	18.17	34.06	31.20	52.70	37.80	25.77	30.70
R1-Llama-8B	Ours	26.48	33.03	55.78	57.68	64.63	78.97	94.53	58.46
Qwen3-32B	Per-Step	16.61	35.89	51.67	65.81	71.09	51.76	32.37	47.17
Qwen3-32B	Ours	24.66	47.87	88.62	89.84	77.66	89.90	91.49	73.29

Appendix

Table 27: Comparison of

\textsc{RF}^{\text{contrast}}

(%) using our standard coarse-grained metric versus a pilot per-step metric on DeepSeek-R1-Distill-Llama-8B and Qwen3-32B. While absolute scores naturally decrease under the stricter per-step constraints, the relative difficulty profile across tasks is broadly preserved.

Panel of Figure 16 — Figure 16: The response from claude-sonnet-4-5-20250929, from Mathematical Reasoning. The model understands that what follows after “Assistant: ” is a human input, not an assistant content as marked in red .

Panel of Figure 17 — Figure 17: The example of faithful incorrect response under intervened input. The response is generated by DeepSeek-R1-Distill-Qwen-32B, from Mathematical Reasoning. Consistent incorrect stance of each reasoning, explanation, and answer is underlined .

Panel of Figure 18 — Figure 18: The example of faithful self-correction response under intervened input. The response is generated by Magistral-Small-2506, from Legal Decision. Self-correction of the model is colored by red .

Panel of Figure 19 — Figure 19: The example of unfaithful silent correction response under intervened input. The response is generated by gpt-oss-20b, from Legal Decision.

Panel of Figure 20 — Figure 20: Prompt format of counterfactual reasoning generation.

Panel of Figure 21 — Figure 21: Counterfactual reasoning generation prompt for Code Generation.

Panel of Figure 22 — Figure 22: Counterfactual reasoning generation prompt for Mathematical Reasoning.

Panel of Figure 23 — Figure 23: Counterfactual reasoning generation prompt for Logical Reasoning.

Panel of Figure 24 — Figure 24: Counterfactual reasoning generation prompt for Table Reasoning.

Panel of Figure 25 — Figure 25: Counterfactual reasoning generation prompt for Context Understanding.

Panel of Figure 26 — Figure 26: Counterfactual reasoning generation prompt for Legal Decision.

Panel of Figure 27 — Figure 27: Counterfactual reasoning generation prompt for Paper Review.

Panel of Figure 28 — Figure 28: Prompt for Automatic LLM Validation.

Panel of Figure 29 — Figure 29: Evaluation prompt for baseline output ( $o$ ).

Panel of Figure 30 — Figure 30: Evaluation prompt for intervened output ( $o^{\prime}$ ).

CoT Faithfulness

Chain-of-thought faithfulness measures whether a model’s stated reasoning causally reflects and supports the answer it gives.

See all 73 papers

Jun 2, 2026

Quantifying Faithful Confidence Expression in Large Reasoning Models

Reliable uncertainty communication is critical to the trustworthiness of LLMs, yet faithful calibration (FC)--the alignment between models' intrinsic and (linguistically) expressed confidence--is a persistent failure mode. This challenge is key for large reasoning models (LRMs), whose extended reasoning traces are often interpreted by users as evidence of deliberation, competence, and confidence. Despite the importance of FC and wide usage of LRMs, the extent to which LRMs can faithfully express their confidence remains poorly understood. Moreover, the prevailing paradigm to measure FC does not generalize well to the long chain-of-thought outputs generated by LRMs, which tend to lack clear step boundaries, involve inconsistent step structure, and encode complex conditional dependencies throughout the trace--complicating estimation of intrinsic confidence. To address this challenge, we introduce a novel framework to systematically quantify FC of LRMs. Our framework analyzes linguistic decisiveness relative to three sources of internal uncertainty, based on token probabilities, hidden states, and sampled response consistency. We also devise a prefix-conditioned sampling approach to control for conditional and structural variation across traces. Applying our framework to a diverse suite of leading models, datasets, and prompts, we find that faithful confidence expression is a significant challenge for LRMs. Reasoning behaviors do not automatically translate to improved FC, and prompt interventions for non-reasoning models do not improve faithfulness in the reasoning setting. Different confidence estimators further produce divergent assessments of the same traces, revealing fragility in prior evaluation methodologies. Taken together, our work establishes FC as a distinct reliability and alignment target for LRMs, particularly as such systems are increasingly deployed in high-stakes contexts.

Areeb Gani, Asal Meskin, Gabrielle Kaili-May Liu +1

1Yale University

Large Reasoning Models Confidence Estimation in Language Models Language Model Calibration LLM Reliability LLM Uncertainty Estimation LLM Evaluation+2

Ask PDF

Apr 13, 2026

Filtered Reasoning Score: Evaluating Reasoning Quality on a Model's Most-Confident Traces

Should we trust Large Language Models (LLMs) with high accuracy? LLMs achieve high accuracy on reasoning benchmarks, but correctness alone does not reveal the quality of the reasoning used to produce it. This highlights a fundamental limitation of outcome-based evaluation: models may arrive at correct answers through flawed reasoning, and models with substantially different reasoning capabilities can nevertheless exhibit similar benchmark accuracy, for example due to memorization or over-optimization. In this paper, we ask: given existing benchmarks, can we move beyond outcome-based evaluation to assess the quality of reasoning itself? We seek metrics that (1) differentiate models with similar accuracy and (2) are robust to variations in input prompts and generation configurations. To this end, we propose a reasoning score that evaluates reasoning traces along dimensions such as faithfulness, coherence, utility, and factuality. A remaining question is how to aggregate this score across multiple sampled traces. Naively averaging them is undesirable, particularly in long-horizon settings, where the number of possible trajectories grows rapidly, and low-confidence correct traces are more likely to be coincidental. To address this, we introduce the Filtered Reasoning Score (FRS), which computes reasoning quality using only the top-K% most confident traces. Evaluating with FRS, models that are indistinguishable under standard accuracy exhibit significant differences in reasoning quality. Moreover, models with higher FRS on one benchmark tend to perform better on other reasoning benchmarks, in both accuracy and reasoning quality. Together, these findings suggest that FRS complements accuracy by capturing a model's transferable reasoning capabilities. We open source our evaluation codebase: https://github.com/HumainLab/filtered_reasoning_score_evaluation.

Manas Pathak, Xingyao Chen, Shuozhe Li +2

University of Texas at Austin

LLM Evaluation Language Model Generation Evaluation Reasoning Evaluation Reasoning Trace Analysis Language Model Reasoning Evaluation

Ask PDF

Sep 10, 2025

LIFT: Likelihood-Based Intervention Faithfulness Training

Chain-of-thought (CoT) reasoning improves LLM accuracy on complex tasks, but CoTs are often unfaithful; that is, the model does not rely on the stated reasoning steps to produce its answer. Existing methods to improve faithfulness require human annotations. We propose LIFT, which automatically scores each reasoning step by its causal influence: we replace it with a domain-matched but irrelevant fact and measure the drop in answer log-probability. We fine-tune on high-scoring CoT traces where interventions successfully disrupted reasoning, jointly optimizing for faithfulness and accuracy. LIFT improves faithfulness across three models (Mistral-7B-Instruct, Qwen3-8B, and Qwen3-4B-Instruct) and three benchmarks, with accuracy comparable to a model fine-tuned solely for accuracy.

Anand Swaroop, Maheep Chaudhary, Kevin Zhu +2

Algoverse AI Research

CoT Faithfulness Faithfulness of Language Model Explanations Causal Interventions in Language Models

Ask PDF

Abstract

CoT Faithfulness

Explore similar work

Quantifying Faithful Confidence Expression in Large Reasoning Models

Filtered Reasoning Score: Evaluating Reasoning Quality on a Model's Most-Confident Traces

LIFT: Likelihood-Based Intervention Faithfulness Training