Qianli Wang

Hengyuan Zhang, Zhihao Zhang, Mingyang Wang et al.

Mechanistic Interpretability (MI) has emerged as a vital approach to demystify the opaque decision-making of Large Language Models (LLMs). However, existing reviews primarily treat MI as an observational science, summarizing analytical insights while lacking a systematic framework for actionable intervention. To bridge this gap, we present a practical survey structured around the pipeline: "Locate, Steer, and Improve." We formally categorize Localizing (diagnosis) and Steering (intervention) methods based on specific Interpretable Objects to establish a rigorous intervention protocol. Furthermore, we demonstrate how this framework enables tangible improvements in Alignment, Capability, and Efficiency, effectively operationalizing MI as an actionable methodology for model optimization. The curated paper list of this work is available at https://github.com/rattlesnakey/Awesome-Actionable-MI-Survey.

Qianli Wang, Tatiana Anikina, Nils Feldhus et al.

Natural language explanations (NLEs) are vital for elucidating the reasoning behind large language model (LLM) decisions. Many techniques have been developed to generate NLEs using LLMs. However, like humans, LLMs might not always produce optimal NLEs on first attempt. Inspired by human learning processes, we introduce Cross-Refine, which employs role modeling by deploying two LLMs as generator and critic, respectively. The generator outputs a first NLE and then refines this initial explanation using feedback and suggestions provided by the critic. Cross-Refine does not require any supervised training data or additional training. We validate Cross-Refine across three NLP tasks using three state-of-the-art open-source LLMs through automatic and human evaluation. We select Self-Refine (Madaan et al., 2023) as the baseline, which only utilizes self-feedback to refine the explanations. Our findings from automatic evaluation and a user study indicate that Cross-Refine outperforms Self-Refine. Meanwhile, Cross-Refine can perform effectively with less powerful LLMs, whereas Self-Refine only yields strong results with ChatGPT. Additionally, we conduct an ablation study to assess the importance of feedback and suggestions. Both of them play an important role in refining explanations. We further evaluate Cross-Refine on a bilingual dataset in English and German.

Jingyi Sun, Qianli Wang, Pepa Atanasova et al.

Chain-of-Thought (CoT) faithfulness, i.e., whether CoTs genuinely reflect large language models' (LLM) underlying behavior, is typically evaluated under two disjoint paradigms: contextual faithfulness, measured by perturbing the input or CoT trace, and parametric faithfulness, assessed by intervening on a model's parametric knowledge. Yet prior work compares them only descriptively. We fill this gap by proposing FaithMate, a unified preference-alignment interface for optimizing models towards either faithfulness paradigm. It enables us to investigate the interplay between the two paradigms, examining whether and to what extent faithfulness gains generalize within and across paradigms. Across three models, two datasets, and six faithfulness metrics, we find that the two paradigms are positively coupled, yet asymmetric: optimizing towards parametric faithfulness yields consistent gains across both paradigms, whereas the contextual counterpart delivers more variable gains. Within the contextual paradigm, faithfulness gains on one metric do not consistently transfer to others, implying that existing contextual metrics capture disjoint facets of faithfulness and exposing inherent trade-offs. These findings imply that CoT faithfulness is not a monolithic objective and therefore requires multifaceted optimization and evaluation.

Nils Feldhus, Qianli Wang, Tatiana Anikina et al.

While recently developed NLP explainability methods let us open the black box in various ways (Madsen et al., 2022), a missing ingredient in this endeavor is an interactive tool offering a conversational interface. Such a dialogue system can help users explore datasets and models with explanations in a contextualized manner, e.g. via clarification or follow-up questions, and through a natural language interface. We adapt the conversational explanation framework TalkToModel (Slack et al., 2022) to the NLP domain, add new NLP-specific operations such as free-text rationalization, and illustrate its generalizability on three NLP tasks (dialogue act classification, question answering, hate speech detection). To recognize user queries for explanations, we evaluate fine-tuned and few-shot prompting models and implement a novel Adapter-based approach. We then conduct two user studies on (1) the perceived correctness and helpfulness of the dialogues, and (2) the simulatability, i.e. how objectively helpful dialogical explanations are for humans in figuring out the model's predicted label when it's not shown. We found rationalization and feature attribution were helpful in explaining the model behavior. Moreover, users could more reliably predict the model outcome based on an explanation dialogue rather than one-off explanations.

Bohao Chu, Qianli Wang, Hendrik Damm et al.

How can system-generated responses be efficiently verified, especially in the high-stakes biomedical domain? To address this challenge, we introduce eTracer, a plug-and-play framework that enables traceable text generation by grounding claims against contextual evidence. Through post-hoc grounding, each response claim is aligned with contextual evidence that either supports or contradicts it. Building on claim-level grounding results, eTracer not only enables users to precisely trace responses back to their contextual source but also quantifies response faithfulness, thereby enabling the verifiability and trustworthiness of generated responses. Experiments show that our claim-level grounding approach alleviates the limitations of conventional grounding methods in aligning generated statements with contextual sentence-level evidence, resulting in substantial improvements in overall grounding quality and user verification efficiency. The code and data are available at https://github.com/chubohao/eTracer.

Qianli Wang, Nils Feldhus, Pepa Atanasova et al.

Quantization is widely used to accelerate inference and streamline the deployment of large language models (LLMs), yet its effects on self-explanations (SEs) remain unexplored. SEs, generated by LLMs to justify their own outputs, require reasoning about the model's own decision-making process, a capability that may exhibit particular sensitivity to quantization. As SEs are increasingly relied upon for transparency in high-stakes applications, understanding whether and to what extent quantization degrades SE quality and faithfulness is critical. To address this gap, we examine two types of SEs: natural language explanations (NLEs) and counterfactual examples, generated by LLMs quantized using three common techniques at distinct bit widths. Our findings indicate that quantization typically leads to moderate declines in both SE quality (up to 4.4\%) and faithfulness (up to 2.38\%). The user study further demonstrates that quantization diminishes both the coherence and trustworthiness of SEs (up to 8.5\%). Compared to smaller models, larger models show limited resilience to quantization in terms of SE quality but better maintain faithfulness. Moreover, no quantization technique consistently excels across task accuracy, SE quality, and faithfulness. Given that quantization's impact varies by context, we recommend validating SE quality for specific use cases, especially for NLEs, which show greater sensitivity. Nonetheless, the relatively minor deterioration in SE quality and faithfulness does not undermine quantization's effectiveness as a model compression technique.

Qianli Wang, Van Bach Nguyen, Yihong Liu et al.

Counterfactuals refer to minimally edited inputs that cause a model's prediction to change, serving as a promising approach to explaining the model's behavior. Large language models (LLMs) excel at generating English counterfactuals and demonstrate multilingual proficiency. However, their effectiveness in generating multilingual counterfactuals remains unclear. To this end, we conduct a comprehensive study on multilingual counterfactuals. We first conduct automatic evaluations on both directly generated counterfactuals in the target languages and those derived via English translation across six languages. Although translation-based counterfactuals offer higher validity than their directly generated counterparts, they demand substantially more modifications and still fall short of matching the quality of the original English counterfactuals. Second, we find the patterns of edits applied to high-resource European-language counterfactuals to be remarkably similar, suggesting that cross-lingual perturbations follow common strategic principles. Third, we identify and categorize four main types of errors that consistently appear in the generated counterfactuals across languages. Finally, we reveal that multilingual counterfactual data augmentation (CDA) yields larger model performance improvements than cross-lingual CDA, especially for lower-resource languages. Yet, the imperfections of the generated counterfactuals limit gains in model performance and robustness.

Jing Yang, Nils Feldhus, Salar Mohtaj et al.

Despite advances in Natural Language Generation (NLG), evaluation remains challenging. Although various new metrics and LLM-as-a-judge (LaaJ) methods are proposed, human judgment persists as the gold standard. To systematically review how NLG evaluation has evolved, we employ an automatic information extraction scheme to gather key information from NLG papers, focusing on different evaluation methods (metrics, LaaJ and human evaluation). With extracted metadata from 14,171 papers across four major conferences (ACL, EMNLP, NAACL, and INLG) over the past six years, we reveal several critical findings: (1) Task Divergence: While Dialogue Generation demonstrates a rapid shift toward LaaJ (>40% in 2025), Machine Translation remains locked into n-gram metrics, and Question Answering exhibits a substantial decline in the proportion of studies conducting human evaluation. (2) Metric Inertia: Despite the development of semantic metrics, general-purpose metrics (e.g., BLEU, ROUGE) continue to be widely used across tasks without empirical justification, often lacking the discriminative power to distinguish between specific quality criteria. (3) Human-LaaJ Divergence: Our association analysis challenges the assumption that LLMs act as mere proxies for humans; LaaJ and human evaluations prioritize very different signals, and explicit validation is scarce (<8% of papers comparing the two), with only moderate to low correlation. Based on these observations, we derive practical recommendations to improve the rigor of future NLG evaluation.

Nils Feldhus, Tanja Baeumel, Elena Golimblevskaia et al.

LLM-as-a-judge has become the dominant paradigm for grading model outputs at scale, yet the same model assigns systematically different scores when its output format changes (e.g., a 1-5 rating vs. a True/False label). Existing diagnoses of these format-induced inconsistencies stop at the input-output level. Using Position-aware Edge Attribution Patching (PEAP), we causally investigate the internal mechanism in Gemma-3, Qwen2.5, and Llama-3. We find that judgments across structured understanding and open-ended preference tasks share a sparse, generalized Latent Evaluator sub-graph in the mid-to-late multi-layer perceptrons (MLPs); zero-ablating it collapses judgment while preserving world knowledge in architecturally modular models. By structurally decoupling abstract judging from output formatting, we provide a mechanistic account of format-induced inconsistency on the open-weight models we study: a continuous judgment signal computed in the shared trunk is mapped through fragile, format-specific terminal branches, enabling format-independent preference to be isolated downstream of the requested output format. Our findings imply that benchmark-level reliability comparisons across formats are partially measuring formatter geometry rather than evaluation quality.

Yilong Wang, Qianli Wang, Bohao Chu et al.

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. Yet extending them beyond English remains challenging: existing methods struggle to produce valid SCEs in non-dominant languages, and a persistent trade-off between validity and minimality undermines explanation quality. We introduce Macro, a preference alignment framework that applies Direct Preference Optimization (DPO) to multilingual SCE generation, using a composite scoring function to construct preference pairs that effectively translate the trade-off into measurable preference signals. Experiments across four LLMs and seven typologically diverse languages show that Macro improves validity by 12.55\% on average over the chain-of-thought baseline without degrading minimality, while avoiding the severe minimality violations of the translation-based baseline. Compared to supervised fine-tuning, Macro achieves superior performance on both metrics, confirming that explicit preference optimization is essential for balancing this trade-off. Further analyses reveal that Macro increases cross-lingual perturbation alignment and mitigates common generation errors. Our results highlight preference optimization as a promising direction for enhancing multilingual model explanations.

Qianli Wang, Tatiana Anikina, Nils Feldhus et al.

Interpretability tools that offer explanations in the form of a dialogue have demonstrated their efficacy in enhancing users' understanding (Slack et al., 2023; Shen et al., 2023), as one-off explanations may fall short in providing sufficient information to the user. Current solutions for dialogue-based explanations, however, often require external tools and modules and are not easily transferable to tasks they were not designed for. With LLMCheckup, we present an easily accessible tool that allows users to chat with any state-of-the-art large language model (LLM) about its behavior. We enable LLMs to generate explanations and perform user intent recognition without fine-tuning, by connecting them with a broad spectrum of Explainable AI (XAI) methods, including white-box explainability tools such as feature attributions, and self-explanations (e.g., for rationale generation). LLM-based (self-)explanations are presented as an interactive dialogue that supports follow-up questions and generates suggestions. LLMCheckupprovides tutorials for operations available in the system, catering to individuals with varying levels of expertise in XAI and supporting multiple input modalities. We introduce a new parsing strategy that substantially enhances the user intent recognition accuracy of the LLM. Finally, we showcase LLMCheckup for the tasks of fact checking and commonsense question answering.

Qianli Wang, Mingyang Wang, Nils Feldhus et al.

Quantization methods are widely used to accelerate inference and streamline the deployment of large language models (LLMs). While prior research has extensively investigated the degradation of various LLM capabilities due to quantization, its effects on model explainability and interpretability, which are crucial for understanding decision-making processes, remain unexplored. To address this gap, we conduct comprehensive experiments using three common quantization techniques at distinct bit widths, in conjunction with two explainability methods, counterfactual examples and natural language explanations, as well as two interpretability approaches, knowledge memorization analysis and latent multi-hop reasoning analysis. We complement our analysis with a thorough user study, evaluating selected explainability methods. Our findings reveal that, depending on the configuration, quantization can significantly impact model explainability and interpretability. Notably, the direction of this effect is not consistent, as it strongly depends on (1) the quantization method, (2) the explainability or interpretability approach, and (3) the evaluation protocol. In some settings, human evaluation shows that quantization degrades explainability, while in others, it even leads to improvements. Our work serves as a cautionary tale, demonstrating that quantization can unpredictably affect model transparency. This insight has important implications for deploying LLMs in applications where transparency is a critical requirement.

Satwinder Singh, Qianli Wang, Zihan Zhong et al.

In this paper, we present a speaker-independent dysarthric speech recognition system, with a focus on evaluating the recently released Speech Accessibility Project (SAP-1005) dataset, which includes speech data from individuals with Parkinson's disease (PD). Despite the growing body of research in dysarthric speech recognition, many existing systems are speaker-dependent and adaptive, limiting their generalizability across different speakers and etiologies. Our primary objective is to develop a robust speaker-independent model capable of accurately recognizing dysarthric speech, irrespective of the speaker. Additionally, as a secondary objective, we aim to test the cross-etiology performance of our model by evaluating it on the TORGO dataset, which contains speech samples from individuals with cerebral palsy (CP) and amyotrophic lateral sclerosis (ALS). By leveraging the Whisper model, our speaker-independent system achieved a CER of 6.99% and a WER of 10.71% on the SAP-1005 dataset. Further, in cross-etiology settings, we achieved a CER of 25.08% and a WER of 39.56% on the TORGO dataset. These results highlight the potential of our approach to generalize across unseen speakers and different etiologies of dysarthria.

Qianli Wang, Nils Feldhus, Simon Ostermann et al.

Counterfactual examples are widely used in natural language processing (NLP) as valuable data to improve models, and in explainable artificial intelligence (XAI) to understand model behavior. The automated generation of counterfactual examples remains a challenging task even for large language models (LLMs), despite their impressive performance on many tasks. In this paper, we first introduce ZeroCF, a faithful approach for leveraging important words derived from feature attribution methods to generate counterfactual examples in a zero-shot setting. Second, we present a new framework, FitCF, which further verifies aforementioned counterfactuals by label flip verification and then inserts them as demonstrations for few-shot prompting, outperforming two state-of-the-art baselines. Through ablation studies, we identify the importance of each of FitCF's core components in improving the quality of counterfactuals, as assessed through flip rate, perplexity, and similarity measures. Furthermore, we show the effectiveness of LIME and Integrated Gradients as backbone attribution methods for FitCF and find that the number of demonstrations has the largest effect on performance. Finally, we reveal a strong correlation between the faithfulness of feature attribution scores and the quality of generated counterfactuals, which we hope will serve as an important finding for future research in this direction.

Luis Felipe Villa-Arenas, Ata Nizamoglu, Qianli Wang et al.

In this work, we introduce a methodology for alignment designed to enhance the ability of large language models (LLMs) to articulate their reasoning (self-explanation) even in the absence of annotated rationale explanations. Our alignment methodology comprises three key components: explanation quality assessment, self-instruction dataset generation, and model alignment. Additionally, we present a novel technique called Alignment with Anchor Preference Pairs, which improves the selection of preference pairs by categorizing model outputs into three groups: consistently correct, consistently incorrect, and variable. By applying tailored strategies to each category, we enhance the effectiveness of Direct Preference Optimization (DPO). Our experimental results demonstrate that this approach significantly improves explanation quality while maintaining accuracy compared to other fine-tuning strategies.

Yilong Wang, Qianli Wang, Nils Feldhus

Counterfactual examples are minimal edits to an input that alter a model's prediction. They are widely employed in explainable AI to probe model behavior and in natural language processing (NLP) to augment training data. However, generating valid counterfactuals with large language models (LLMs) remains challenging, as existing single-pass methods often fail to induce reliable label changes, neglecting LLMs' self-correction capabilities. To explore this untapped potential, we propose iFlip, an iterative refinement approach that leverages three types of feedback, including model confidence, feature attribution, and natural language. Our results show that iFlip achieves an average 57.8% higher validity than the five state-of-the-art baselines, as measured by the label flipping rate. The user study further corroborates that iFlip outperforms baselines in completeness, overall satisfaction, and feasibility. In addition, ablation studies demonstrate that three components are paramount for iFlip to generate valid counterfactuals: leveraging an appropriate number of iterations, pointing to highly attributed words, and early stopping. Finally, counterfactuals generated by iFlip enable effective counterfactual data augmentation, substantially improving model performance and robustness.

Qianli Wang, Tatiana Anikina, Nils Feldhus et al.

Conversational explainable artificial intelligence (ConvXAI) systems based on large language models (LLMs) have garnered considerable attention for their ability to enhance user comprehension through dialogue-based explanations. Current ConvXAI systems often are based on intent recognition to accurately identify the user's desired intention and map it to an explainability method. While such methods offer great precision and reliability in discerning users' underlying intentions for English, a significant challenge in the scarcity of training data persists, which impedes multilingual generalization. Besides, the support for free-form custom inputs, which are user-defined data distinct from pre-configured dataset instances, remains largely limited. To bridge these gaps, we first introduce MultiCoXQL, a multilingual extension of the CoXQL dataset spanning five typologically diverse languages, including one low-resource language. Subsequently, we propose a new parsing approach aimed at enhancing multilingual parsing performance, and evaluate three LLMs on MultiCoXQL using various parsing strategies. Furthermore, we present Compass, a new multilingual dataset designed for custom input extraction in ConvXAI systems, encompassing 11 intents across the same five languages as MultiCoXQL. We conduct monolingual, cross-lingual, and multilingual evaluations on Compass, employing three LLMs of varying sizes alongside BERT-type models.

Qianli Wang, Van Bach Nguyen, Nils Feldhus et al.

Counterfactual examples are widely employed to enhance the performance and robustness of large language models (LLMs) through counterfactual data augmentation (CDA). However, the selection of the judge model used to evaluate label flipping, the primary metric for assessing the validity of generated counterfactuals for CDA, yields inconsistent results. To decipher this, we define four types of relationships between the counterfactual generator and judge models: being the same model, belonging to the same model family, being independent models, and having an distillation relationship. Through extensive experiments involving two state-of-the-art LLM-based methods, three datasets, four generator models, and 15 judge models, complemented by a user study (n = 90), we demonstrate that judge models with an independent, non-fine-tuned relationship to the generator model provide the most reliable label flipping evaluations. Relationships between the generator and judge models, which are closely aligned with the user study for CDA, result in better model performance and robustness. Nevertheless, we find that the gap between the most effective judge models and the results obtained from the user study remains considerably large. This suggests that a fully automated pipeline for CDA may be inadequate and requires human intervention.

Qianli Wang, Tatiana Anikina, Nils Feldhus et al.

Conversational explainable artificial intelligence (ConvXAI) systems based on large language models (LLMs) have garnered significant interest from the research community in natural language processing (NLP) and human-computer interaction (HCI). Such systems can provide answers to user questions about explanations in dialogues, have the potential to enhance users' comprehension and offer more information about the decision-making and generation processes of LLMs. Currently available ConvXAI systems are based on intent recognition rather than free chat, as this has been found to be more precise and reliable in identifying users' intentions. However, the recognition of intents still presents a challenge in the case of ConvXAI, since little training data exist and the domain is highly specific, as there is a broad range of XAI methods to map requests onto. In order to bridge this gap, we present CoXQL, the first dataset in the NLP domain for user intent recognition in ConvXAI, covering 31 intents, seven of which require filling multiple slots. Subsequently, we enhance an existing parsing approach by incorporating template validations, and conduct an evaluation of several LLMs on CoXQL using different parsing strategies. We conclude that the improved parsing approach (MP+) surpasses the performance of previous approaches. We also discover that intents with multiple slots remain highly challenging for LLMs.