论文检索

Tianhui Cai, Yifan Liu, Zewei Zhou, Haoxuan Ma, Seth Z. Zhao, Zhiwen Wu, Xu Han, Zhiyu Huang, Jiaqi Ma

Understanding and adhering to traffic regulations is essential for autonomous vehicles to ensure safety and trustworthiness. However, traffic regulations are complex, context-dependent, and differ between regions, posing a major challenge to conventional rule-based decision-making approaches. We present an interpretable, regulation-aware decision-making framework, DriveReg, which enables autonomous vehicles to understand and adhere to region-specific traffic laws and safety guidelines. The framework integrates a Retrieval Augmented Generation (RAG)-based Traffic Regulation Retrieval Agent, which retrieves relevant rules from regulatory documents based on the current situation, and a Large Language Model (LLM)-powered Reasoning Agent that evaluates actions for legal compliance and safety. Our design emphasizes interpretability to enhance transparency and trustworthiness. To support systematic evaluation, we introduce DriveReg Scenarios Dataset, a comprehensive dataset of driving scenarios across Boston, Singapore, and Los Angeles, with both hypothesized text-based cases and real-world driving data, specifically constructed and annotated to evaluate models’ capacity for regulation understanding and reasoning. We validate our framework on the DriveReg Scenarios Dataset and real-world deployment, demonstrating strong performance and robustness across diverse environments.

Andrew Blair-Stanek, Nils Holzenberger, Benjamin Van Durme

We investigate whether large language models can discover and analyze U.S. tax-minimization strategies. This real-world domain challenges even seasoned human experts, and progress can reduce tax revenue lost from well-advised, wealthy taxpayers. We evaluate the most advanced LLMs on their ability to (1) interpret and verify tax strategies, (2) fill in gaps in partially specified strategies, and (3) generate complete, end-to-end strategies from scratch. This domain should be of particular interest to the LLM reasoning community: unlike synthetic challenge problems or scientific reasoning tasks, U.S. tax law involves navigating hundreds of thousands of pages of statutes, case law, and administrative guidance, all updated regularly. Notably, an LLM identified an apparently novel tax strategy, highlighting these models' potential to revolutionize tax agencies' fight against tax abuse.

Junwen Zheng, Xinran Xu, Li Rong Wang, Chang Cai, Lucinda Siyun Tan, Dingyuan Wang, Hong Liang Tey, Xiuyi Fan

Deep learning has demonstrated expert-level performance in melanoma classification, positioning it as a powerful tool in clinical dermatology. However, model opacity and the lack of interpretability remain critical barriers to clinical adoption, as clinicians often struggle to trust the decision-making processes of black-box models. To address this gap, we present a Cross-modal Explainable Framework for Melanoma (CEFM) that leverages contrastive learning as the core mechanism for achieving interpretability. Specifically, CEFM maps clinical criteria for melanoma diagnosis—namely Asymmetry, Border, and Color (ABC)—into the Vision Transformer embedding space using dual projection heads, thereby aligning clinical semantics with visual features. The aligned representations are subsequently translated into structured textual explanations via natural language generation, creating a transparent link between raw image data and clinical interpretation. Experiments on public datasets demonstrate 92.79% accuracy and an AUC of 0.961, along with significant improvements across multiple interpretability metrics. Qualitative analyses further show that the spatial arrangement of the learned embeddings aligns with clinicians’ application of the ABC rule, effectively bridging the gap between high-performance classification and clinical trust.

Hanwei Zhang, Luo Cheng, Rui Wen, Yang Zhang, Lijun Zhang, Holger Hermanns

Explainable AI (XAI) is crucial for building transparent and trustworthy machine learning systems, especially in high-stakes domains. Concept Bottleneck Models (CBMs) have emerged as a promising ante-hoc approach that provides interpretable, concept-level explanations by explicitly modeling human-understandable concepts. However, existing CBMs often suffer from poor locality faithfulness, failing to spatially align concepts with meaningful image regions, which limits their interpretability and reliability. In this work, we propose SL-CBM (CBM with Semantic Locality), a novel extension that enforces locality faithfulness by generating spatially coherent saliency maps at both concept and class levels. SL-CBM integrates a 1 × 1 convolutional layer with a cross-attention mechanism to enhance alignment between concepts, image regions, and final predictions. Unlike prior methods, SL-CBM produces faithful saliency maps inherently tied to the model’s internal reasoning, facilitating more effective debugging and intervention. Extensive experiments on image datasets demonstrate that SL-CBM substantially improves locality faithfulness, explanation quality, and intervention efficacy while maintaining competitive classification accuracy. Our ablation studies highlight the importance of contrastive and entropy-based regularization for balancing accuracy, sparsity, and faithfulness. Overall, SL-CBM bridges the gap between concept-based reasoning and spatial explainability, setting a new standard for interpretable and trustworthy concept-based models.

Nirmalendu Prakash, Yeo Wei Jie, Amir Abdullah, Ranjan Satapathy, Erik Cambria, Roy Ka-Wei Lee

Refusal on harmful prompts is a key safety behaviour in instruction‑tuned large language models (LLMs), yet the internal causes of this behaviour remain poorly understood. We study two public instruction tuned models—Gemma‑2-2B‑IT and LLaMA‑3.1-8B‑IT using sparse autoencoders (SAEs) trained on residual‑stream activations. Given a harmful prompt, we search the SAE latent space for feature sets whose ablation flips the model from refusal to compliance, demonstrating causal influence and creating a jailbreak. Our search proceeds in three stages: 1. Refusal Direction - Finding a refusal mediating direction and collecting SAE features close to that direction, followed by 2. Greedy Filtering - to prune this set to obtain a minimal set and finally 3. Interaction Discovery - a factorization‑machine (FM) model that captures non‑linear interactions among the remaining active features and the minimal set. This pipeline yields a broad set of jailbreak-critical features, offering insight into the mechanistic basis of refusal. Moreover, we also find evidence of redundant features which remain dormant unless earlier features are suppressed. Our findings highlight the potential for fine-grained auditing and targeted intervention in safety behaviours by manipulating the interpretable latent space.

Ninell Oldenburg, Ruchira Dhar, Anders Søgaard

In this paper, we argue that current AI research operates on a spectrum between two different underlying conceptions of intelligence: Intelligence Realism, which holds that intelligence represents a single, universal capacity measurable across all systems, and Intelligence Pluralism, which views intelligence as diverse, context-dependent capacities that cannot be reduced to a single universal measure. Through an analysis of current debates in AI research, we demonstrate how the conceptions remain largely implicit yet fundamentally shape how empirical evidence gets interpreted across a wide range of areas. These underlying views generate fundamentally different research approaches across three areas. Methodologically, they produce different approaches to model selection, benchmark design, and experimental validation. Interpretively, they lead to contradictory readings of the same empirical phenomena, from capability emergence to system limitations. Regarding AI risk, they generate categorically different assessments: realists view superintelligence as the primary risk and search for unified alignment solutions, while pluralists see diverse threats across different domains requiring context-specific solutions. We argue that making explicit these underlying assumptions can contribute to a clearer understanding of disagreements in AI research.

Zixuan Liu, Siavash H. Khajavi, Guangkai Jiang, Xinru Liu

Reward-model-based fine-tuning is a central paradigm in aligning Large Language Models with human preferences. However, such approaches critically rely on the assumption that proxy reward models accurately reflect intended supervision, a condition often violated due to annotation noise, bias, or limited coverage. This misalignment can lead to undesirable behaviors, where models optimize for flawed signals rather than true human values. In this paper, we investigate a novel framework to identify and mitigate such misalignment by treating the fine-tuning process as a form of knowledge integration. We focus on detecting instances of proxy-policy conflicts, cases where the base model strongly disagrees with the proxy. We argue that such conflicts often signify areas of shared ignorance, where neither the policy nor the reward model possesses sufficient knowledge, making them especially susceptible to misalignment. To this end, we propose two complementary metrics for identifying these conflicts: a localized Proxy-Policy Alignment Conflict Score (PACS) and a global Kendall-Tau Distance measure. Building on this insight, we design an algorithm named Selective Human-in-the-loop Feedback via Conflict-Aware Sampling (SHF-CAS) that targets high-conflict QA pairs for additional feedback, refining both the reward model and policy efficiently. Experiments on two alignment tasks demonstrate that our approach enhances general alignment performance, even when trained with a biased proxy reward. Our work provides a new lens for interpreting alignment failures and offers a principled pathway for targeted refinement in LLM training.

Minkyu Kim, Suan Lee, Jinho Kim

Transparent models, which provide inherently interpretable predictions, are receiving significant attention in high-stakes domains. However, despite much real-world data being collected as time series, there is a lack of studies on transparent time series models. To address this gap, we propose a novel transparent neural network model for time series called Generalized Additive Time Series Model (GATSM). GATSM consists of two parts: 1) independent feature networks to learn feature representations, and 2) a transparent temporal module to learn temporal patterns across different time steps using the feature representations. This structure allows GATSM to effectively capture temporal patterns and handle varying-length time series while preserving transparency. Empirical experiments show that GATSM significantly outperforms existing generalized additive models and achieves comparable performance to black-box time series models, such as recurrent neural networks and Transformer. In addition, we demonstrate that GATSM finds interesting patterns in time series.

Rémi Kazmierczak, Steve Azzolin, Eloïse Berthier, Anna Hedström, Patricia Delhomme, David Filliat, Nicolas Bousquet, Goran Frehse, Massimiliano Mancini, Baptiste Caramiaux 等

We introduce PASTA (Perceptual Assessment System for explanaTion of Artificial Intelligence), a novel human-centric framework for evaluating eXplainable AI (XAI) techniques in computer vision. Our first contribution is the creation of the PASTA-dataset, the first large-scale benchmark that spans a diverse set of models and both saliency-based and concept-based explanation methods. This dataset enables robust, comparative analysis of XAI techniques based on human judgment. Our second contribution is an automated, data-driven benchmark that predicts human preferences using the PASTA-dataset. This scoring called PASTA-score method offers scalable, reliable, and consistent evaluation aligned with human perception. Additionally, our benchmark allows for comparisons between explanations across different modalities, an aspect previously unaddressed. We then propose to apply our scoring method to probe the interpretability of existing models and to build more human interpretable XAI methods.

Rom Himelstein, Amit LeVi, Brit Youngmann, Yaniv Nemcovsky, Avi Mendelson

Safety-aligned large language models (LLMs) are becoming increasingly widespread, especially in sensitive applications where fairness is essential and biased outputs can cause significant harm. However, evaluating the fairness of models is a complex challenge, and approaches that do so typically utilize standard question-answer (QA) styled schemes. Such methods often overlook deeper issues by interpreting the model's refusal responses as positive fairness measurements, which creates a false sense of fairness. In this work, we introduce the concept of silenced biases, which are unfair preferences encoded within models' latent space and are effectively concealed by safety-alignment. Previous approaches that considered similar indirect biases often relied on prompt manipulation or handcrafted implicit queries, which present limited scalability and risk contaminating the evaluation process with additional biases. We propose the Silenced Bias Benchmark (SBB), which aims to uncover these biases by employing activation steering to reduce model refusals during QA. SBB supports easy expansion to new demographic groups and subjects, presenting a fairness evaluation framework that encourages the future development of fair models and tools beyond the masking effects of alignment training. We demonstrate our approach over multiple LLMs, where our findings expose an alarming distinction between models' direct responses and their underlying fairness issues.

Parian Haghighat, Hadis Anahideh, Cynthia Rudin

The existence of multiple, equally accurate models for a given predictive task leads to predictive multiplicity, where a ``Rashomon set'' of models achieve similar accuracy but diverge in their individual predictions. This inconsistency undermines trust in high-stakes applications where we want consistent predictions. We propose three approaches to reduce inconsistency among predictions for the members of the Rashomon set. The first approach is outlier correction. An outlier has a label that none of the good models are capable of predicting correctly. Outliers can cause the Rashomon set to have high variance predictions in a local area, so fixing them can lower variance. Our second approach is local patching. In a local region around a test point, models may disagree with each other because some of them are biased. We can detect and fix such biases using a validation set, which also reduces multiplicity. Our third approach is pairwise reconciliation, where we find pairs of models that disagree on a region around the test point. We modify predictions that disagree, making them less biased. These three approaches can be used together or separately, and they each have distinct advantages. The reconciled predictions can then be distilled into a single interpretable model for real-world deployment. In experiments across multiple datasets, our methods reduce disagreement metrics while maintaining competitive accuracy.

Neta Glazer, Yael Segal-Feldman, Hilit Segev, Aviv Shamsian, Asaf Buchnick, Gill Hetz, Ethan Fetaya, Joseph Keshet, Aviv Navon

Interpretability methods have recently gained significant attention, particularly in the context of large language models, enabling insights into linguistic representations, error detection, and model behaviors such as hallucinations and repetitions. However, these techniques remain underexplored in automatic speech recognition (ASR), despite their potential to advance both the performance and interpretability of ASR systems. In this work, we adapt and systematically apply established interpretability methods such as logit lens, linear probing, and activation patching, to examine how acoustic and semantic information evolves across layers in ASR systems. Our experiments reveal previously unknown internal dynamics, including specific encoder-decoder interactions responsible for repetition hallucinations and semantic biases encoded deep within acoustic representations. These insights demonstrate the benefits of extending and applying interpretability techniques to speech recognition, opening promising directions for future research on improving model transparency and robustness.

Brage Eilertsen, Røskva Bjørgfinsdóttir, Francielle Vargas, Ali Ramezani-Kebrya

The opaque nature of deep learning models presents significant challenges for the ethical deployment of hate speech detection systems. To address this limitation, we introduce Supervised Rational Attention (SRA), a framework that explicitly aligns model attention with human rationales, improving both interpretability and fairness in hate speech classification. SRA integrates a supervised attention mechanism into transformer-based classifiers, optimizing a joint objective that combines standard classification loss with an alignment loss term that minimizes the discrepancy between attention weights and human-annotated rationales. We evaluated SRA on hate speech benchmarks in English (HateXplain) and Portuguese (HateBRXplain) with rationale annotations. Empirically, SRA achieves 2.4× better explainability compared to current baselines, and produces token-level explanations that are more faithful and human-aligned. In terms of fairness, SRA achieves competitive fairness across all measures, with second-best performance in detecting toxic posts targeting identity groups, while maintaining comparable results on other metrics. These findings demonstrate that incorporating human rationales into attention mechanisms can enhance interpretability and faithfulness without compromising fairness.

Bikash Dutta, Adit Jain, Rishabh Ranjan, Mayank Vatsa, Richa Singh

Large Audio Language Models (LALMs) are transforming AI by processing and generating human language directly from audio. As these models proliferate in real-world applications, it becomes critical to evaluate their performance to ensure equitable and safe use across diverse linguistic and cultural contexts. We present the first comprehensive study of cultural bias in LALMs, extending text-based harm frameworks to the audio modality to analyze how linguistic diversity influences model behavior and uncover challenges in interpreting audio nuances. To address this, we introduce the Audio Cultural Intelligence Dataset (ACID), a multilingual audio–text benchmark spanning 1,315 hours across diverse languages and cultural contexts, and we conduct a systematic evaluation of 10 open-source and two closed-source models. Our results reveal substantial performance disparities across languages and cultural settings and show that biases manifest distinctly when models process audio inputs. These findings highlight the need to evaluate LALMs not only for technical accuracy but also for fair and culturally sensitive behavior, motivating the development of inclusive datasets and culturally aware training practices for safer and more equitable audio language models.

Ming-Chang Chiu, Shicheng Wen, Pin-Yu Chen, Xuezhe Ma

In vision-language models (VLMs), the ability to perceive and interpret color and physical environment is crucial for achieving contextually accurate understanding and interaction. However, despite advances in multimodal modeling, there remains a significant lack of specialized datasets that rigorously evaluate a model's capacity to discern subtle color variations and spatial context---critical elements for situational comprehension and reliable deployment across real-world applications. Toward that goal, we curate MegaCoin, a high-quality, human-labeled dataset based on \emph{real} images with various contextual attributes. MegaCoin consists of two parts: MegaCoin-Instruct, which serves as a supervised fine-tuning (SFT) dataset for VLMs; and MegaCoin-Bench, an annotated test set that can be used as a stand-alone QA dataset. MegaCoin provides three annotated features for 220,000 real images: foreground color, background color, and description of an object's physical environment, constituting 660k human annotations. In addition, MegaCoin can be applied to benchmark domain generalization (DG) algorithms. We explore benchmarking DG methods in the linear probing setup for VLM and show some new insights. Last but not least, we show that VLMs, including GPT-4o, have subpar color recognition capabilities, and fine-tuning with MegaCoin can result in improved performance on visual evaluation tasks. In certain cases, MegaCoin fine-tuned small-scale open-source models such as LLaVA and Bunny can outperform closed-source GPT-4o. We hope the utilities of MegaCoin can shed light on the directions VLMs can improve and provide a more complex platform for domain generalization algorithms.

Mahdi Alehdaghi, Rajarshi Bhattacharya, Pourya Shamsolmoali, Rafael M. O. Cruz, Eric Granger

As AI systems become more capable, it is important that their decisions are understandable and aligned with human expectations. A key challenge is the lack of interpretability in deep models. Existing methods such as GradCAM generate heatmaps but provide limited conceptual insight, while prototype-based approaches offer example-based explanations but often rely on rigid region selection and lack semantic consistency. To address these limitations, we propose PCMNet, a Part-Prototypical Concept Mining Network that learns human-comprehensible prototypes from meaningful regions without extra supervision. By clustering these into concept groups and extracting concept activation vectors, PCMNet provides structured, concept-level explanations and enhances robustness under occlusion and adversarial conditions, which are both critical for building reliable and aligned AI systems. Experiments across multiple benchmarks show that PCMNet outperforms state-of-the-art methods in interpretability, stability, and robustness. This work contributes to AI alignment by enhancing transparency, controllability, and trustworthiness in modern AI systems.

Xueqi Ma, Xingjun Ma, Sarah Monazam Erfani, Danilo Mandic, James Bailey

Developing open-set classification methods capable of classifying in-distribution (ID) data while detecting out-of-distribution (OOD) samples is essential for deploying graph neural networks (GNNs) in open-world scenarios. Existing methods typically treat all OOD samples as a single class, despite real-world applications—especially high-stake settings like fraud detection and medical diagnosis—demanding deeper insights into OOD samples, including their probable labels. This raises a critical question: Can OOD detection be extended to OOD classification without true label information? To answer this question, we introduce a Coarse-to-Fine open-set Classification (CFC) method that leverages large language models (LLMs) for text-attributed graphs. CFC consists of three key components: (1) A coarse classifier that utilizes LLM prompts for OOD detection and outlier label generation; (2) A GNN-based fine classifier trained with OOD samples from (1) for enhanced OOD detection and ID classification; and (3) Refined OOD classification achieved through LLM prompts and post-processed OOD labels. Unlike methods relying on synthetic or auxiliary OOD samples, CFC employs semantic OOD data-instances that are genuinely out-of-distribution based on their inherent meaning, thus improving interpretability and practical utility. CFC enhances OOD detection by 10% compared to state-of-the-art approaches on text-attributed graphs and in the text domain, while achieving up to 70% accuracy in OOD classification on graph datasets.

Xiao Zhao, Chang Liu, Ruiteng Ji, Zheyuan Zhang, Mingxu Zhu, Linna Song, Zhe Ren, Luo Qingliang, YuHang Gao, Zhaolong Du 等

Recent advances in vision language models (VLMs) have demonstrated remarkable potential in embodied navigation tasks. However, existing robot-centric datasets primarily focus on traditional 3D tasks such as perception and prediction, lacking adequate support for vision-language tasks. Vision-language-navigation (VLN) is a key capability for achieving human-like and interpretable navigation in complex environments. In this study, we present CoT-VLNBench, the first large-scale benchmark and dataset designed for chain-of-thought (CoT) reasoning in quadruped robot navigation. Our dataset encompasses a diverse range of indoor and outdoor scenes, multi-step navigation trajectories, and rich natural language instructions, all annotated with fine-grained CoT reasoning traces. Specifically, it contains 175K frames, 5.25M 3D bounding boxes, and 875K vision–question–answer (VQA) pairs. This comprehensive resource enables thorough evaluation of embodied agents’ perceptual and step-by-step reasoning abilities. Furthermore, we propose a novel CoT-VLN model, a state-of-the-art 7B VLN model that integrates visual, linguistic, and reasoning modules, to facilitate interpretable and effective navigation. Extensive experiments demonstrate that our approach significantly outperforms existing non-VLMs baselines on the new benchmark, underscoring the importance of CoT-VLN in embodied navigation. We hope that CoT-VLNBench will serve as a valuable resource to advance research at the intersection of robotics, vision, language, and reasoning.

Ziwei Wang, Leyang Yang, Xiaoxuan Tang, Sheng Zhou, Dajun Chen, Wei Jiang, Yong Li

Advances in Multimodal Large Language Models have significantly enhanced Graphical User Interface (GUI) automation. Equipping GUI agents with reliable episodic reasoning capabilities is essential for bridging the gap between users’ concise task descriptions and the complexities of real-world execution. Current methods integrate Reinforcement Learning (RL) with System-2 Chain-of-Thought, yielding notable gains in reasoning enhancement. For long-horizon GUI tasks, historical interactions connect each screen to the goal-oriented episode chain, and effectively leveraging these clues is crucial for the current decision. However, existing native GUI agents exhibit weak short-term memory in their explicit reasoning, interpreting the chained interactions as discrete screen understanding, i.e., unawareness of the historical interactions within the episode. This history-agnostic reasoning challenges their performance in GUI automation. To alleviate this weakness, we propose a History-Aware Reasoning (HAR) framework, which encourages an agent to reflect on its own errors and acquire episodic reasoning knowledge from them via tailored strategies that enhance short-term memory in long-horizon interaction. The framework mainly comprises constructing a reflective learning scenario, synthesizing tailored correction guidelines, and designing a hybrid RL reward function. Using the HAR framework, we develop a native end-to-end model, HAR-GUI-3B, which alters the inherent reasoning mode from history-agnostic to history-aware, equipping the GUI agent with stable short-term memory and reliable perception of screen details. Comprehensive evaluations across a range of GUI-related benchmarks demonstrate the effectiveness and generalization of our method.

Bernd Finkbeiner, Niklas Metzger, Satya Prakash Nayak, Anne-Kathrin Schmuck

Universal Safety Controllers (USCs) are a promising logical control framework that guarantees the satisfaction of a given temporal safety specification when applied to any realizable plant model. Unlike traditional methods, which synthesize one logical controller over a given detailed plant model, USC synthesis constructs a generic controller whose outputs are conditioned by plant behavior, called prophecies. Thereby, USCs offer strong generalization and scalability benefits over classical logical controllers. However, the exact computation and verification of prophecies remain computationally challenging. In this paper, we introduce an approximation algorithm for USC synthesis that addresses these limitations via learning. Instead of computing exact prophecies, which reason about sets of trees via automata, we only compute under- and over-approximations from (small) example plants and infer computation tree logic (CTL) formulas as representations of prophecies. The resulting USC generalizes to unseen plants via a verification step and offers improved efficiency and explainability through small and concise CTL prophecies, which remain human-readable and interpretable. Experimental results demonstrate that our learned prophecies remain generalizable, yet are significantly more compact and interpretable than their exact tree automata representations.