The paper investigates whether the tokens highlighted by large language models (LLMs) as key to their predictions are behaviorally consistent when those tokens are manipulated. Using sufficiency, comprehensiveness, and counterfactuality metrics, the authors evaluate closed‑source GPT‑4o‑mini and open‑source models Gemma3, Granite8B, and DeepSeek on IMDB and Steam review datasets. Results show that GPT‑4o‑mini consistently follows expected patterns, Gemma3 and Granite8B maintain sufficiency but falter under stricter interventions, and DeepSeek variants exhibit structural inconsistencies, indicating that explanation reliability varies across model families and sizes.
By Alisson Rosa Pereira, Bruno Iochins Grisci
arXiv:2602.02639v2 Announce Type: replace-cross
Abstract: LLM self-explanations are often presented as a promising tool for AI oversight, yet their faithfulness to the model's true reasoning process...
By Harry Mayne, Justin Singh Kang, Dewi Gould, Kannan Ramchandran, Adam Mahdi, Noah Y. Siegel
The paper investigates how quantization affects large language models’ self‑explanations, examining natural language explanations and counterfactual examples across three quantization techniques and bit widths. Results show moderate declines in explanation quality (up to 4.4%) and faithfulness (up to 3.9%), with user studies indicating up to an 8.5% drop in coherence and trustworthiness. Larger models are less resilient in quality but remain more faithful, and no single quantization method consistently outperforms others across accuracy, quality, and faithfulness.
By Qianli Wang, Nils Feldhus, Pepa Atanasova, Fedor Splitt, Simon Ostermann, Sebastian M\"oller, Vera Schmitt
arXiv:2503. 13445v3 Announce Type: replace-cross Abstract: When asked to explain their decisions, LLMs can often give explanations which sound plausible to humans.
By Noah Y. Siegel, Nicolas Heess, Maria Perez-Ortiz, Oana-Maria Camburu
arXiv:2605. 11632v2 Announce Type: replace-cross Abstract: Self-generated counterfactual explanations (SCEs) are minimally modified inputs (minimality) generated by large language models (LLMs) that flip their own predictions (validity), offering a causally grounded approach to unraveling black-box LLM behavior.
By Yilong Wang, Qianli Wang, Bohao Chu, Yihong Liu, Jing Yang, Simon Ostermann
iFlip is an iterative refinement method for generating counterfactual examples using large language models. It incorporates three feedback types—model confidence, feature attribution, and natural language—to guide successive edits. Experiments show iFlip outperforms five state‑of‑the‑art baselines, achieving a 57.8% higher validity rate and improving model performance through counterfactual data augmentation.
By Yilong Wang, Qianli Wang, Nils Feldhus
arXiv:2607. 21090v1 Announce Type: cross 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.
By Yeoktatt Cheah, Mar\'ia P\'erez-Ortiz, Noah Y. Siegel, Oana-Maria Camburu
arXiv:2606. 28615v1 Announce Type: cross Abstract: Large language models (LLMs) are increasingly deployed in high-stakes domains, where free-text explanations such as chain-of-thought and post-hoc rationales are used to justify model outputs.
By Nhi Nguyen, Shauli Ravfogel, Rajesh Ranganath
arXiv:2608. 16747v1 Announce Type: cross Abstract: Many areas of AI research, such as language model interpretability and chain of thought faithfulness, seek to explain model behaviors.
By Adam Karvonen, Euan Ong, Subhash Kantamneni, Samuel Marks
The paper investigates whether language models can self‑attribute the influence of external information on their answers and proposes a counterfactual training method called CREDIT to improve this ability. Experiments show that most state‑of‑the‑art models rarely attribute influential information and often attribute non‑influential information, and that CREDIT significantly enhances attribution in several open models, outperforming other fine‑tuning and prompting approaches. CREDIT also distinguishes between answers that are genuinely changed by user suggestions and those that would have been given regardless, with some transfer to unseen information types and conflicting prompts.
By Zikai Zhou, Jiacheng Sang, Yilin Xu, Yufei Jin, Monica S. Lam
arXiv:2608. 16627v1 Announce Type: cross Abstract: Natural language explanations (NLEs) are increasingly used as inputs, for example, as few-shot rationales that influence model behavior in in-context learning (ICL).
By Mahdi Dhaini, Adam Dejl, Juraj Vladika, Volkan \"Ozer, Barbara Plank, Gjergji Kasneci
The paper discusses how Large Language Models can produce natural language self‑explanations that appear plausible but may not accurately reflect the model’s reasoning. It critiques current evaluation methods for such explanations and offers practical guidelines to assess their plausibility and faithfulness. Additionally, it argues that evaluation should also consider the actionability of these explanations, showing how they can aid decision‑making for various stakeholders.
By Elize Herrewijnen, Benedetta Muscato, Gizem Gezici, Fosca Giannotti