Pith. sign in

REVIEW 4 major objections 4 minor 150 references

Reinforcement learning can directly train LLMs to generate self-explanations that faithfully reflect the factors actually driving their decisions, with the faithfulness correlation Phi-CCT rising from near zero to 0.664 in-distribution and

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 08:29 UTC pith:PPJ5HAHN

load-bearing objection Novel RL-for-faithfulness result is worth referee time, but the stale-label confound and reporting inconsistencies keep it from being conclusive. the 4 major comments →

arxiv 2607.21090 v1 pith:PPJ5HAHN submitted 2026-07-23 cs.LG cs.AIcs.CL

Training Large Language Models for Self-Explanation Faithfulness

Yeoktatt Cheah , Mar\'ia P\'erez-Ortiz , Noah Y. Siegel , Oana-Maria Camburu This is my paper
classification cs.LG cs.AIcs.CL
keywords self-explanation faithfulnesscounterfactual interventionsreinforcement learningPhi-CCTinfluence detectionmention alignmentlarge language modelsexplainability
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper argues that self-explanation faithfulness—whether a model's stated reasoning actually reflects the factors driving its decision—can be directly optimized rather than only measured. To do this, the authors turn the Phi-CCT faithfulness metric into a per-sample reinforcement-learning reward: a counterfactual prompt is created by inserting a single phrase, the insertion is labelled influential if the model's answer flips, and the model is rewarded when its explanation mentions exactly those influential insertions and omits the non-influential ones. They report that this RL procedure lifts Phi-CCT from near zero to 0.664 in-distribution and up to 0.691 on held-out tasks, while ordinary supervised fine-tuning teaches detection without verbalization. If correct, this gives a scalable training-time route to reducing unfaithful reasoning, since the reward is computed from the model's own behavior rather than from gold rationales.

Core claim

The central claim is that a model can be taught to disclose the factors that change its own decisions, and that this disclosure can be optimized directly with reinforcement learning. The operational definition of faithfulness is simple: for an inserted intervention Δ, influence is I=1{Y'≠Y} (the intervention flipped the decision), mention is M=1{Δ appears in the explanation Z'}, and a faithful explanation has M=I. The per-sample reward r=1{M⇔I} is aggregated into Phi-CCT = Corr(I,M) over a dataset. RL fine-tuning with this reward improved Phi-CCT from near zero to 0.664 for in-distribution data and to 0.691 for out-of-distribution tasks such as StrategyQA, while supervised fine-tuning improv

What carries the argument

The key object is the counterfactual probe: each training point is a factual prompt X paired with X'=X⊕Δ, and the reward r=1{M⇔I} is built from whether the inserted Δ flips the decision (I=1{Y'≠Y}) and whether the explanation mentions Δ (M). Because r is computed from the current model's generated explanations, it can be used as a per-sample reward in group-relative policy optimization (GRPO), with advantages normalized within a batch. The dataset-level evaluation uses Phi-CCT = Corr(I,M), which punishes both indiscriminate mentioning and indiscriminate silence, and class-balanced sampling prevents fixed strategies from earning high reward. This machinery is what lets a faithfulness metric,

Load-bearing premise

The load-bearing premise is that a single controlled insertion is the only causal difference between factual and counterfactual prompts, and that exact-token mention is a complete measure of disclosure; if either proxy fails, the reward is not actually measuring faithfulness, and because influence labels come from the frozen initial model, the policy can drift toward stale labels during training.

What would settle it

Resample influence labels from the current policy at every training step and compare Phi-CCT to the frozen-label variant; if the on-policy score falls to near zero, the frozen-label gains are mostly label drift. A second test is to replace exact-token mention detection with a paraphrase-aware detector and see whether trained models' Phi-CCT changes; if it rises sharply, the original metric was undervaluing genuine disclosure.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Self-explanations can be improved at training time without gold rationales or external judges, because the reward uses only the model's own decisions and its generated text.
  • Models trained this way can generalize the disclosure behavior to unseen datasets such as StrategyQA and ComVE, so the effect is not confined to the training distribution.
  • Deployment-time audits and oversight can use the same counterfactual probe: if an explanation omits an intervention that flips the model's decision, that omission is a detectable unfaithfulness signal.
  • Since RL training also improves influence introspection modestly, the learned disclosure rule appears to interact with the model's own detection ability rather than being a pure surface behavior.
  • The absence of increased overlap ratio or completion-length gaming indicates that faithfulness gains can be obtained without degenerate copying or silence collapse.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Because the influence labels are frozen from the initial model and not resampled, the cleanest next test is a fully on-policy version that recomputes I=1{Y'≠Y} from the current policy; the paper itself flags this. If the on-policy gain disappears, part of the reported Phi-CCT improvement would be label drift rather than improved disclosure.
  • Exact-token and morphological mention detection may undercount paraphrased references to an intervention; a paraphrase-tolerant mention detector would likely change both the measured scores and the cross-intervention transfer results.
  • The same reward design could be extended to multiple simultaneous interventions or to naturalistic paragraphs, where the assumption of a single controlled difference no longer holds; such an extension would test whether the mechanism scales beyond the paper's single-word and template-bias probes.
  • A broader multi-objective reward that also preserves task accuracy and calibration may be needed before training-time faithfulness can be used as a safety component, since the paper notes shortened explanations and possible calibration effects.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes an RL fine-tuning method that directly optimizes the faithfulness of LLM self-explanations. Faithfulness is operationalized through counterfactual interventions: for each prompt X and inserted intervention Δ, influence is I = 1{Y′ ≠ Y} and mention is M = 1{Δ mentioned in Z′}. The per-sample reward is r = 1{M ⇔ I}, trained with GRPO. Experiments on Llama3.1-8B and Qwen3-8B with random-word and user-bias interventions report large in-distribution Phi-CCT gains (up to 0.664) and OOD transfer (up to 0.691), and behavior analyses are used to argue against trivial reward hacking. The paper also distinguishes SFT-based influence introspection (detection) from RL-based disclosure.

Significance. If the central claim is established, this is a useful step toward directly training self-explanations to reflect a model's actual decision drivers, rather than only evaluating faithfulness at inference time. The paper has notable strengths: bootstrap CIs on the main metric, balanced datasets to prevent simple shortcut strategies, explicit behavioral checks for shortening and parroting, and an unusually candid discussion of the off-policy nature of the training labels. However, the load-bearing evidence is currently compromised by an internal inconsistency about whether evaluation labels are stale, and the main results section misreports the numbers in the appendix tables. The central claim—that RL teaches faithful disclosure rather than merely reward alignment to frozen labels—is defensible but not yet established.

major comments (4)
  1. [§4.2, §7, and the off-policy paragraph] The manuscript is internally inconsistent about the evaluation protocol. The §4.2 off-policy paragraph states that 'evaluation targets alignment with the trained model's current decisions,' while §7 ('Limitations') states that 'We train and evaluate faithfulness with respect to the original model's decisions y, y′ collected prior to training.' If the evaluation uses the original frozen decisions, then the reported Phi-CCT is not a measure of the trained model's faithfulness at all. If the evaluation uses the current decisions, then the acknowledged stale-label drift can inflate the reported gains: the model is rewarded for matching I computed at T0, and reported Phi-CCT may improve partly because the policy's decisions drift toward those stale labels. Either way, the current reporting does not establish the central claim. The paper must state the evaluation protocol precisely and, as a c
  2. [§4.2, RL training setup] The RL reward is computed on explanations generated after appending the counterfactual decision y′ to the counterfactual prompt x′, while the dataset construction in §4.1 and the evaluation protocol generate (y′, z′) jointly from x′. This train/eval conditioning gap gives the training policy access to the decision it must explain, which can encourage rationalization of a supplied answer rather than detection of influence. The paper should evaluate under matched conditions: either generate z′ without appending y′ during training, or append y′ during evaluation. Reporting both would clarify whether the learned disclosure transfers to the evaluation setting.
  3. [§5, first paragraph vs Tables 3 and 4] The main results section misreports the introspection results. It claims that SFT improves Llama3.1-8B on random insertions from -0.013 to 0.537, but Table 3(a) reports SFT-on-RI test correlation 0.114; the 0.536/0.537 figure is the RL Phi-CCT from Table 4(a). Similarly, the claimed Llama user-bias SFT value 0.664 is the RL Phi-CCT from Table 4(b); Table 3(b) reports 0.271. The numbers for Qwen also mix the two tables. Since the SFT-versus-RL contrast is a central contribution, the text must be corrected to match the tables, and the distinction between predicted-influence correlation (Table 3) and Phi-CCT (Table 4) must be maintained.
  4. [§3.1 and §5] Because the RL objective is a per-sample binary version of the same I/M agreement that Phi-CCT evaluates, a high Phi-CCT after RL is partly an optimization-success result, not an independent external validation. The paper should add a matched control, e.g., RL with a reward that is correlated with the intervention but not the exact M⇔I agreement (such as always-mention or always-silence), to show that the specific reward shape is what drives the behavioral change. The OOD and cross-intervention results already provide some evidence against pure reward hacking, but a control would make the causal claim much stronger.
minor comments (4)
  1. [Table 3] Several rows report identical train and test values (e.g., Llama base random-insertion -0.013 on both; Llama SFT-on-UB random-insertion 0.667; Llama SFT-on-RI user-bias 0.623). If train and test splits are actually the same for these conditions, state this explicitly; otherwise the duplication suggests an error.
  2. [§3.1 and notation] The reward is written as r = 1{M ⇐ ⇒ I}, which is typographically confusing. Use the standard indicator notation 1[M ⇔ I] or define the logical equivalence explicitly.
  3. [§5.2 and Figure 4] The behavioral claims—shortening for predicted negatives and absence of parroting—are based on visual inspection of Figure 4. Add effect sizes and significance tests or at least numeric summaries for the completion-length and overlap-ratio comparisons.
  4. [§1 and §5] The abstract and introduction describe the OOD result as 'reaching up to 0.691 on held-out tasks such as StrategyQA.' This number appears only for Qwen3-8B RL-on-UB on the combined OOD split; clarify that it is not a per-dataset StrategyQA figure and that OOD performance is not uniform across conditions.

Circularity Check

1 steps flagged

Mild definitional alignment between the RL reward and the Phi-CCT evaluation metric; no hidden circular derivation. The main flagged confound is an acknowledged off-policy staleness limitation, not a circularity.

specific steps
  1. self definitional [Section 3.1 (faithfulness definition, reward r=1{M⇔I}) and Section 5 (Phi-CCT results, e.g. 0.664)]
    "This sets up a training objective: a faithful self-explanation should mention Δ if and only if it influences the decision Y′. This is a proxy for Phi-CCT faithfulness metric (Siegel et al., 2025) defined as the correlation between influence and mention over an entire dataset: Phi-CCT= Corr(I,M). For our reward in RL training, we use a per-sample binary indicator of whether influence and mention agree: r=1{M ⇐ ⇒ I}."

    The evaluation metric Phi-CCT is the dataset-level correlation of the same binary I and M whose per-sample agreement r is the RL reward. Thus the headline in-distribution gain (e.g., 0.664 Phi-CCT) is, by construction, the result of training on a per-sample surrogate of the reported statistic; the metric is not an independent external check that 'faithful disclosure' improved. The construct validity of Corr(I,M) as faithfulness is inherited from the authors' own Phi-CCT/CCT papers and from the single-insertion causal model. The paper is transparent about this—'we bridge this gap by modifying existing faithfulness metrics into an RL training objective'—so this is a mild definitional alignment rather than a hidden tautology.

full rationale

The derivation is mostly self-contained: Phi-CCT is defined in the paper as Corr(I,M), and the RL reward is explicitly a per-sample binary version of that same agreement. This makes reward and evaluation aligned by construction, which is appropriate for a training objective but means the headline in-distribution number is not an independent validation of 'faithful disclosure'; I record that as a mild self-definitional step. The self-citations to Siegel et al. (2024; 2025) are not load-bearing in the logical sense because the metric is redefined in-text, and the method also builds on non-self work (Atanasova et al. 2023; Turpin et al. 2023). Independent evidence comes from held-out and OOD splits (Phi-CCT up to 0.691 on StrategyQA) and from behavioral analyses (completion length, overlap ratio), which are not part of the reward. The paper itself flags the most important confound: 'Reported Phi-CCT gains may therefore partly reflect the trained model's decisions drifting toward the stale labels, rather than purely improved disclosure.' That is an experimental-validity limitation—off-policy influence labels—not a circularity, because the reward does not use the evaluation outcome. Overall score 2 reflects the reward/evaluation alignment and reliance on the authors' own metric construct, without treating the derivation as equivalent to its inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 6 axioms · 0 invented entities

No free parameters are fitted to the reported Phi-CCT values; the numbers come from empirical evaluation of trained checkpoints. The load-bearing premises are the intervention-based proxy for internal causation, exact-token mention as disclosure, and the off-policy staleness of influence labels. No new entities such as particles, forces, or hidden states are introduced.

axioms (6)
  • domain assumption I=1{Y'≠Y} measures the causal influence of intervention Δ on the model's decision.
    The paper treats the controlled insertion Δ as the only difference between X and X' and attributes any answer flip to Δ. Used to define the reward and as a proxy for Phi-CCT.
  • domain assumption M=1{Δ mentioned in Z'} via morphological matching measures faithful disclosure.
    The reward and evaluation count only exact or morphologically related mentions; paraphrases or synonyms count as non-mention. The paper acknowledges this can underestimate mention rates.
  • domain assumption Deterministic decoding (T=0) gives stable decisions suitable for counterfactual pairing.
    The method assumes that Y vs Y' flips are attributable to Δ and not to sampling noise, because generations are produced at temperature 0.
  • domain assumption r=1{M⇔I} is an adequate per-sample surrogate for dataset-level Phi-CCT correlation.
    The reward is the binary agreement indicator, while Phi-CCT is the correlation over the whole dataset. This is a design choice, not a derivation.
  • ad hoc to paper Stale influence labels from the frozen initial model remain informative during RL training.
    Explicitly flagged in §4.2 and Limitations: 'training reward and evaluation metric are measuring subtly different objects... Reported Phi-CCT gains may therefore partly reflect the trained model's decisions drifting toward the stale labels.' This is a load-bearing assumption the paper does not quantify.
  • domain assumption Balanced 50/50 influential/non-influential splits remove class-prior reward hacking.
    The paper deliberately constructs balanced splits so that static mention/silence strategies score zero correlation, meaning reported Phi-CCT values are conditional on a constructed distribution rather than natural intervention frequency.

pith-pipeline@v1.3.0-alltime-deepseek · 29444 in / 14118 out tokens · 135053 ms · 2026-08-01T08:29:21.896778+00:00 · methodology

0 comments
read the original abstract

We propose a Reinforcement Learning (RL) method to directly optimize the faithfulness of self-explanations - the extent to which a model's generated reasoning accurately reflects its internal decision-making process. While existing work focuses on evaluating faithfulness or using inference-time prompting frameworks to improve an LLM's self-explanation's tractability, these approaches do not provide a mechanism to directly optimize a model's parameters to generate faithful self-explanations. We bridge this gap by modifying existing faithfulness metrics into an RL training objective. We investigate (1) if models can be trained to accurately detect factors that affect their decisions, and (2) whether RL can directly optimize for the disclosure of these factors thereby improving LLM self-explanations' faithfulness. We experiment with two intervention types: random-word insertions and user-bias insertions, using a per-sample reward derived from the Phi-CCT correlation metric. RL fine-tuned Llama3.1-8B and Qwen3-8B show substantial improvements on the Phi-CCT faithfulness metric, with in-distribution scores rising from near-zero to as high as 0.664, and out-of-distribution scores reaching up to 0.691 on held-out tasks such as StrategyQA. Cross-intervention generalization is weaker but more interesting: a priori we would not expect a model trained only on random word insertions to generalize to user-bias phrases, yet Llama3.1-8B shows non-zero transfer in this direction. The reverse direction and Qwen3-8B do not replicate this, indicating model-dependent and setup-dependent effects we cannot yet explain. Lastly we analyze model behavior to rule out reward gaming behaviors that often plague RL training. Ultimately, we show that models can be trained to implicitly identify influential factors and disclose them, offering a scalable path toward reducing unfaithful reasoning in LLMs.

Figures

Figures reproduced from arXiv: 2607.21090 by Mar\'ia P\'erez-Ortiz, Noah Y. Siegel, Oana-Maria Camburu, Yeoktatt Cheah.

Figure 1
Figure 1. Figure 1: Comparison of causal paths. In (a), the path [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Workflow of our RL training framework for improving faithfulness of self-explanations. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Models trained on Influence Introspection via SFT and models trained with Mention [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Analysis of models trained on Influence Introspection (SFT) and Mention Alignment (RL). [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Confusion matrices for all Llama3.1-8B-Instruct model variants evaluated on influence [PITH_FULL_IMAGE:figures/full_fig_p020_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Confusion matrices for all Qwen3-8B model variants evaluated on influence introspec [PITH_FULL_IMAGE:figures/full_fig_p021_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Confusion matrices of Llama3.1-8B-Instruct RL models evaluated on the influence [PITH_FULL_IMAGE:figures/full_fig_p022_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Confusion matrices of Qwen3-8B RL models evaluated on the influence-disclosure align [PITH_FULL_IMAGE:figures/full_fig_p023_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

150 extracted references · 18 canonical work pages · 12 internal anchors

  1. [1]

    2018 , publisher =

    The Book of Why: The New Science of Cause and Effect , author =. 2018 , publisher =

    2018

  2. [2]

    2009 , publisher =

    Causality: Models, Reasoning and Inference , author =. 2009 , publisher =

    2009

  3. [3]

    e-SNLI: Natural Language Inference with Natural Language Explanations , booktitle =

    Oana. e-SNLI: Natural Language Inference with Natural Language Explanations , booktitle =. 2018 , url =

    2018

  4. [4]

    CoRR , volume =

    Cunxiang Wang and Shuailong Liang and Yili Jin and Yilong Wang and Xiaodan Zhu and Yue Zhang , title =. CoRR , volume =. 2020 , url =. 2007.00236 , timestamp =

    Pith/arXiv arXiv 2020

  5. [5]

    Challenging BIG-Bench Tasks and Whether Chain-of-Thought Can Solve Them , booktitle =

    Mirac Suzgun and Nathan Scales and Nathanael Sch. Challenging BIG-Bench Tasks and Whether Chain-of-Thought Can Solve Them , booktitle =. 2023 , url =. doi:10.18653/V1/2023.FINDINGS-ACL.824 , timestamp =

    doi:10.18653/v1/2023.findings-acl.824 2023

  6. [6]

    Faithfulness vs

    Agarwal, Chirag and Tanneru, Sree Harsha and Lakkaraju, Himabindu , year =. Faithfulness vs. doi:10.48550/arXiv.2402.04614 , urldate =. 2402.04614 , primaryclass =

    Pith/arXiv arXiv doi:10.48550/arxiv.2402.04614

  7. [7]

    Quantus:

    Anna, Hedstr. Quantus:. 2022 , doi =

    2022

  8. [8]

    Reconsidering Faithfulness in Regular, Self-Explainable and Domain Invariant

    Steve Azzolin and Antonio Longa and Stefano Teso and Andrea Passerini , booktitle=. Reconsidering Faithfulness in Regular, Self-Explainable and Domain Invariant. 2025 , url=

    2025

  9. [9]

    arXiv preprint arXiv:2204.05514 , year=

    A comparative study of faithfulness metrics for model interpretability methods , author=. arXiv preprint arXiv:2204.05514 , year=

    Pith/arXiv arXiv

  10. [10]

    Chiappa, Silvia , year =. Path-. Proceedings of the AAAI Conference on Artificial Intelligence , volume =. doi:10.1609/aaai.v33i01.33017801 , urldate =

    doi:10.1609/aaai.v33i01.33017801

  11. [11]

    Framework for

    Dasgupta, Sanjoy and Frost, Nave and Moshkovitz, Michal , year =. Framework for. Proceedings of the 39th

  12. [12]

    De Cao, Nicola and Schlichtkrull, Michael Sejr and Aziz, Wilker and Titov, Ivan , editor =. How Do. Proceedings of the 2020. 2020 , month = nov, pages =. doi:10.18653/v1/2020.emnlp-main.262 , urldate =

    work page doi:10.18653/v1/2020.emnlp-main.262 2020

  13. [13]

    and Manzoor, Emaad and Pryzant, Reid and Sridhar, Dhanya and

    Feder, Amir and Keith, Katherine A. and Manzoor, Emaad and Pryzant, Reid and Sridhar, Dhanya and. Causal. 2022 , month = oct, journal =. doi:10.1162/tacl_a_00511 , urldate =

    doi:10.1162/tacl_a_00511 2022

  14. [14]

    Faithful

    Gat, Yair and Calderon, Nitay and Feder, Amir and Chapanin, Alexander and Sharma, Amit and Reichart, Roi , year =. Faithful. doi:10.48550/arXiv.2310.00603 , urldate =. 2310.00603 , primaryclass =

    Pith/arXiv arXiv doi:10.48550/arxiv.2310.00603

  15. [15]

    Transactions on Machine Learning Research , issn=

    Evaluating Interpretable Methods via Geometric Alignment of Functional Distortions , author=. Transactions on Machine Learning Research , issn=. 2025 , url=

    2025

  16. [16]

    Jacovi, Alon and Goldberg, Yoav , year =. Towards. doi:10.48550/arXiv.2004.03685 , urldate =. 2004.03685 , primaryclass =

    Pith/arXiv arXiv doi:10.48550/arxiv.2004.03685 2004

  17. [17]

    Algorithmic

    Karimi, Amir-Hossein and Sch. Algorithmic. Proceedings of the 2021. 2021 , month = mar, pages =. doi:10.1145/3442188.3445899 , urldate =

    arXiv 2021

  18. [18]

    Karimi, Amir-Hossein and Muandet, Krikamol and Kornblith, Simon and Sch. On the. Proceedings of the 40th. 2023 , month = jul, pages =

    2023

  19. [19]

    Text and

    Keith, Katherine and Jensen, David and O'Connor, Brendan , editor =. Text and. Proceedings of the 58th. 2020 , month = jul, pages =. doi:10.18653/v1/2020.acl-main.474 , urldate =

    doi:10.18653/v1/2020.acl-main.474 2020

  20. [20]

    Counterfactual

    Kusner, Matt J and Loftus, Joshua and Russell, Chris and Silva, Ricardo , year =. Counterfactual. Advances in

  21. [21]

    Causality can systematically address the monsters under the bench(marks)

    Causality Can Systematically Address the Monsters under the Bench(Marks) , author =. 2025 , month = feb, number =. doi:10.48550/arXiv.2502.05085 , urldate =. 2502.05085 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2502.05085 2025

  22. [22]

    Position:

    Loftus, Joshua , year =. Position:. doi:10.48550/arXiv.2406.02275 , urldate =. 2406.02275 , primaryclass =

    Pith/arXiv arXiv doi:10.48550/arxiv.2406.02275

  23. [23]

    Lyu, Qing and Apidianaki, Marianna and. Towards. 2024 , month = jun, journal =. doi:10.1162/coli_a_00511 , urldate =

    doi:10.1162/coli_a_00511 2024

  24. [24]

    A Review of Faithfulness Metrics for Hallucination Assessment in

    Malin, Ben and Kalganova, Tatiana and Boulgouris, Nikoloas , year =. A Review of Faithfulness Metrics for Hallucination Assessment in. doi:10.48550/arXiv.2501.00269 , urldate =. 2501.00269 , primaryclass =

    work page doi:10.48550/arxiv.2501.00269

  25. [25]

    Walk the

    Matton, Katie and Ness, Robert and Guttag, John and Kiciman, Emre , year =. Walk the. The

  26. [26]

    Parcalabescu, Letitia and Frank, Anette , year =. On. doi:10.48550/arXiv.2311.07466 , urldate =. 2311.07466 , primaryclass =

    Pith/arXiv arXiv doi:10.48550/arxiv.2311.07466

  27. [27]

    Ribeiro, Marco and Singh, Sameer and Guestrin, Carlos , editor =. ``. Proceedings of the 2016. 2016 , month = jun, pages =. doi:10.18653/v1/N16-3020 , urldate =

    doi:10.18653/v1/n16-3020 2016

  28. [28]

    Xai\_evals :

    Seth, Pratinav and Rathore, Yashwardhan and Singh, Neeraj Kumar and Chitroda, Chintan and Sankarapu, Vinay Kumar , year =. Xai\_evals :. doi:10.48550/arXiv.2502.03014 , urldate =. 2502.03014 , primaryclass =

    Pith/arXiv arXiv doi:10.48550/arxiv.2502.03014

  29. [29]

    Sia, Suzanna and Belyy, Anton and Almahairi, Amjad and Khabsa, Madian and Zettlemoyer, Luke and Mathias, Lambert , year =. Logical. Proceedings of the AAAI Conference on Artificial Intelligence , volume =. doi:10.1609/aaai.v37i8.26174 , urldate =

    work page doi:10.1609/aaai.v37i8.26174

  30. [30]

    Evaluating Input Feature Explanations through a Unified Diagnostic Evaluation Framework

    Sun, Jingyi and Atanasova, Pepa and Augenstein, Isabelle , year =. Evaluating. doi:10.48550/arXiv.2406.15085 , urldate =. 2406.15085 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2406.15085

  31. [31]

    , year =

    Turpin, Miles and Michael, Julian and Perez, Ethan and Bowman, Samuel R. , year =. Language. Thirty-Seventh

  32. [32]

    Counterfactual

    Wachter, Sandra and Mittelstadt, Brent and Russell, Chris , year =. Counterfactual. SSRN Electronic Journal , issn =. doi:10.2139/ssrn.3063289 , urldate =

    doi:10.2139/ssrn.3063289

  33. [34]

    Aligning as

    Xia, Yu and Yu, Tong and He, Zhankui and Zhao, Handong and McAuley, Julian and Li, Shuai , editor =. Aligning as. Proceedings of the 2024. 2024 , month = jun, pages =. doi:10.18653/v1/2024.naacl-long.262 , urldate =

    doi:10.18653/v1/2024.naacl-long.262 2024

  34. [35]

    Zhang, Weijia and Aliannejadi, Mohammad and Yuan, Yifei and Pei, Jiahuan and Huang, Jia-Hong and Kanoulas, Evangelos , year =. Towards. doi:10.48550/arXiv.2406.15264 , urldate =. 2406.15264 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2406.15264

  35. [36]

    Zheng, Xu and Shirani, Farhad and Chen, Zhuomin and Lin, Chaohao and Cheng, Wei and Guo, Wenbo and Luo, Dongsheng , year =. F-. The

  36. [37]

    Advances in neural information processing systems , volume=

    Chain-of-thought prompting elicits reasoning in large language models , author=. Advances in neural information processing systems , volume=

  37. [38]

    Advances in neural information processing systems , volume=

    A unified approach to interpreting model predictions , author=. Advances in neural information processing systems , volume=

  38. [39]

    ERASER : A Benchmark to Evaluate Rationalized NLP Models

    DeYoung, Jay and Jain, Sarthak and Rajani, Nazneen Fatema and Lehman, Eric and Xiong, Caiming and Socher, Richard and Wallace, Byron C. ERASER : A Benchmark to Evaluate Rationalized NLP Models. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. 2020. doi:10.18653/v1/2020.acl-main.408

    doi:10.18653/v1/2020.acl-main.408 2020

  39. [40]

    International Journal of Sustainable Development Through AI, ML and IoT , volume=

    Empowering the Patient Journey: The Role of Generative AI in Healthcare , author=. International Journal of Sustainable Development Through AI, ML and IoT , volume=. 2023 , month=

    2023

  40. [41]

    2025 , eprint=

    Faithfulness of LLM Self-Explanations for Commonsense Tasks: Larger Is Better, and Instruction-Tuning Allows Trade-Offs but Not Pareto Dominance , author=. 2025 , eprint=

    2025

  41. [42]

    CoRR , volume =

    Yinheng Li and Shaofei Wang and Han Ding and Hang Chen , title =. CoRR , volume =. 2023 , url =. doi:10.48550/ARXIV.2311.10723 , eprinttype =. 2311.10723 , timestamp =

    Pith/arXiv arXiv doi:10.48550/arxiv.2311.10723 2023

  42. [43]

    Leveraging Large Language Models for Learning Complex Legal Concepts through Storytelling , booktitle =

    Hang Jiang and Xiajie Zhang and Robert Mahari and Daniel Kessler and Eric Ma and Tal August and Irene Li and Alex Pentland and Yoon Kim and Deb Roy and Jad Kabbara , editor =. Leveraging Large Language Models for Learning Complex Legal Concepts through Storytelling , booktitle =. 2024 , url =. doi:10.18653/V1/2024.ACL-LONG.388 , timestamp =

    work page doi:10.18653/v1/2024.acl-long.388 2024

  43. [44]

    Do Prompt-Based Models Really Understand the Meaning of Their Prompts?

    Webson, Albert and Pavlick, Ellie. Do Prompt-Based Models Really Understand the Meaning of Their Prompts?. Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2022. doi:10.18653/v1/2022.naacl-main.167

    doi:10.18653/v1/2022.naacl-main.167 2022

  44. [45]

    2023 , eprint=

    Measuring Faithfulness in Chain-of-Thought Reasoning , author=. 2023 , eprint=

    2023

  45. [46]

    2022 , eprint=

    Selection-Inference: Exploiting Large Language Models for Interpretable Logical Reasoning , author=. 2022 , eprint=

    2022

  46. [47]

    Shao, Zhihong and Wang, Peiyi and Zhu, Qihao and Xu, Runxin and Song, Junxiao and Bi, Xiao and Zhang, Haowei and Zhang, Mingchuan and Li, Y. K. and Wu, Y. and Guo, Daya , month = apr, year =. doi:10.48550/arXiv.2402.03300 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2402.03300

  47. [48]

    Ouyang, Long and Wu, Jeff and Jiang, Xu and Almeida, Diogo and Wainwright, Carroll L. and Mishkin, Pamela and Zhang, Chong and Agarwal, Sandhini and Slama, Katarina and Ray, Alex and Schulman, John and Hilton, Jacob and Kelton, Fraser and Miller, Luke and Simens, Maddie and Askell, Amanda and Welinder, Peter and Christiano, Paul and Leike, Jan and Lowe, R...

    Pith/arXiv arXiv doi:10.48550/arxiv.2203.02155

  48. [49]

    Dodge, Jesse and Ilharco, Gabriel and Schwartz, Roy and Farhadi, Ali and Hajishirzi, Hannaneh and Smith, Noah , month = feb, year =. Fine-. doi:10.48550/arXiv.2002.06305 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2002.06305 2002

  49. [50]

    and Zeng, W

    Liu, W. and Zeng, W. and He, K. and Jiang, Y. and He, J. , title =. 2023 , archivePrefix =. 2312.15685 , url =

    Pith/arXiv arXiv 2023

  50. [51]

    and Yang, Z

    Luo, Y. and Yang, Z. and Meng, F. and Li, Y. and Zhou, J. and Zhang, Y. , title =. 2023 , archivePrefix =. 2308.08747 , url =

    Pith/arXiv arXiv 2023

  51. [52]

    and Kadavath, S

    Bai, Y. and Kadavath, S. and Kundu, S. and Askell, A. and Kernion, J. , title =. 2022 , archivePrefix =. 2204.05862 , url =

    Pith/arXiv arXiv 2022

  52. [53]

    and Davies, X

    Casper, S. and Davies, X. and Shi, C. and Gilbert, T. and Scheurer, J. , title =. 2023 , archivePrefix =. 2307.15217 , url =

    Pith/arXiv arXiv 2023

  53. [54]

    , title =

    Mroueh, Y. , title =. 2025 , archivePrefix =. 2503.06639 , url =

    arXiv 2025

  54. [55]

    arXiv preprint arXiv:2402.03300 , year=

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models , author=. arXiv preprint arXiv:2402.03300 , year=

    Pith/arXiv arXiv

  55. [56]

    Towards Faithful Multi-step Reasoning through Fine-Grained Causal-aware Attribution Reasoning Distillation

    Chu, Zheng and Chen, Jingchang and Wang, Zhongjie and Tang, Guo and Chen, Qianglong and Liu, Ming and Qin, Bing. Towards Faithful Multi-step Reasoning through Fine-Grained Causal-aware Attribution Reasoning Distillation. Proceedings of the 31st International Conference on Computational Linguistics. 2025

    2025

  56. [57]

    2023 , eprint=

    BUMP: A Benchmark of Unfaithful Minimal Pairs for Meta-Evaluation of Faithfulness Metrics , author=. 2023 , eprint=

    2023

  57. [58]

    Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages =

    Wu, Chenwang and Wang, Xiting and Lian, Defu and Xie, Xing and Chen, Enhong , title =. Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages =. 2023 , isbn =. doi:10.1145/3580305.3599240 , abstract =

    arXiv 2023

  58. [59]

    Proceedings of the 40th International Conference on Machine Learning , pages =

    On the Relationship Between Explanation and Prediction: A Causal View , author =. Proceedings of the 40th International Conference on Machine Learning , pages =. 2023 , editor =

    2023

  59. [60]

    2023 , eprint=

    A Meta-Evaluation of Faithfulness Metrics for Long-Form Hospital-Course Summarization , author=. 2023 , eprint=

    2023

  60. [61]

    Diagnostic Pathology , author =

    Challenges and barriers of using large language models (. Diagnostic Pathology , author =. 2024 , keywords =. doi:10.1186/s13000-024-01464-7 , abstract =

    doi:10.1186/s13000-024-01464-7 2024

  61. [62]

    AI Open , author =

    Large language models in law:. AI Open , author =. 2024 , keywords =. doi:10.1016/j.aiopen.2024.09.002 , abstract =

    doi:10.1016/j.aiopen.2024.09.002 2024

  62. [63]

    Position:

    Chen, Zichen and Chen, Jiaao and Chen, Jianda and Sra, Misha , month = feb, year =. Position:. doi:10.48550/arXiv.2502.15865 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2502.15865

  63. [64]

    2024 , date=

    Making Reasoning Matter: Measuring and Improving Faithfulness of Chain-of-Thought Reasoning , author=. 2024 , date=. 2402.13950 , archivePrefix=

    Pith/arXiv arXiv 2024

  64. [65]

    2025 , eprint=

    SCOPE: A Self-supervised Framework for Improving Faithfulness in Conditional Text Generation , author=. 2025 , eprint=

    2025

  65. [66]

    Reasoning Models Don't Always Say What They Think , abstract =

    Chen, Yanda and Benton, Joe and Radhakrishnan, Ansh and Uesato, Jonathan and Denison, Carson and Schulman, John and Somani, Arushi and Hase, Peter and Wagner, Misha and Roger, Fabien and Mikulik, Vlad and Bowman, Sam and Leike, Jan and Kaplan, Jared and Perez, Ethan , langid =. Reasoning Models Don't Always Say What They Think , abstract =. 2025 , date =

    2025

  66. [67]

    2025 , institution =

    Reasoning Models Don't Always Say What They Think , author =. 2025 , institution =

    2025

  67. [68]

    Context-faithful Prompting for Large Language Models

    Zhou, Wenxuan and Zhang, Sheng and Poon, Hoifung and Chen, Muhao. Context-faithful Prompting for Large Language Models. Findings of the Association for Computational Linguistics: EMNLP 2023. 2023. doi:10.18653/v1/2023.findings-emnlp.968

    doi:10.18653/v1/2023.findings-emnlp.968 2023

  68. [69]

    2023 , eprint=

    Faithful Chain-of-Thought Reasoning , author=. 2023 , eprint=

    2023

  69. [70]

    2023 , eprint=

    Chain-of-Thought Prompting Elicits Reasoning in Large Language Models , author=. 2023 , eprint=

    2023

  70. [71]

    2024 , eprint=

    Making Reasoning Matter: Measuring and Improving Faithfulness of Chain-of-Thought Reasoning , author=. 2024 , eprint=

    2024

  71. [72]

    Distillation with Explanations from Large Language Models

    Zhang, Hanyu and Wang, Xiting and Ao, Xiang and He, Qing. Distillation with Explanations from Large Language Models. Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024). 2024

    2024

  72. [73]

    2024 , eprint=

    Inference-Time Intervention: Eliciting Truthful Answers from a Language Model , author=. 2024 , eprint=

    2024

  73. [74]

    2024 , eprint=

    Non-Linear Inference Time Intervention: Improving LLM Truthfulness , author=. 2024 , eprint=

    2024

  74. [75]

    arXiv preprint arXiv:2305.18029 , year=

    Faithfulness tests for natural language explanations , author=. arXiv preprint arXiv:2305.18029 , year=

    Pith/arXiv arXiv

  75. [76]

    Advances in Neural Information Processing Systems , volume=

    Language models don't always say what they think: Unfaithful explanations in chain-of-thought prompting , author=. Advances in Neural Information Processing Systems , volume=

  76. [77]

    arXiv preprint arXiv:2307.13702 , year=

    Measuring faithfulness in chain-of-thought reasoning , author=. arXiv preprint arXiv:2307.13702 , year=

    Pith/arXiv arXiv

  77. [78]

    arXiv preprint arXiv:2404.18655 , year=

    Revealing the parametric knowledge of language models: A unified framework for attribution methods , author=. arXiv preprint arXiv:2404.18655 , year=

    Pith/arXiv arXiv

  78. [79]

    arXiv preprint arXiv:2406.15085 , year=

    Evaluating input feature explanations through a unified diagnostic evaluation framework , author=. arXiv preprint arXiv:2406.15085 , year=

    Pith/arXiv arXiv

  79. [80]

    arXiv preprint arXiv:2505.22823 , year=

    Self-Critique and Refinement for Faithful Natural Language Explanations , author=. arXiv preprint arXiv:2505.22823 , year=

    Pith/arXiv arXiv

  80. [81]

    arXiv preprint arXiv:2404.03189 , year=

    The probabilities also matter: A more faithful metric for faithfulness of free-text explanations in large language models , author=. arXiv preprint arXiv:2404.03189 , year=

    Pith/arXiv arXiv

Showing first 80 references.