论文检索

Bertille Tierny, Arthur Charpentier, Francois Hu

Linear models are widely used in high-stakes decision-making due to their interpretability, but fairness constraints like Demographic Parity (DP) create opaque effects on model coefficients and predictive bias distribution. We propose a post-processing framework that can be applied on top of any linear model to decompose bias into direct (sensitive-attribute) and indirect (correlated-features) components. Our method analytically characterizes how DP reshapes each coefficient, enabling transparent feature-level interpretation.

Gokaramaiah Thota, Nagaraju K, Sathya Babu Korra

Deep learning has advanced medical imaging, but limited interpretability hinders clinical adoption. Class activation maps (CAM) provide visual explanations, yet methods such as Score-CAM are computationally expensive, requiring a forward pass for each activation map and limiting real-time applicability despite their high fidelity. To overcome this limitation, LowRank-CAM is proposed, which aggregates activation maps into a global matrix and applies singular value decomposition (SVD) to extract dominant spatial modes. The resulting top-r low-rank attention masks, with r

Benjamin Richards, Pushpa Kumar Balan

Deep learning models are emerging as strong alternatives to numerical weather prediction, yet their internal representations remain poorly understood. We analyze the latent space of Microsoft’s Aurora model to test whether its embed- dings align with known physical processes. First, we show that land–sea distinctions are strongly captured, with errors mainly at coastlines. Second, we examine extreme surface temperatures using percentile-based thresholds, finding that embeddings reveal a gradient from moderate to severe events, though recall degrades at the rarest percentiles. These results suggest that Aurora’s encoder encodes physically consistent features but underestimates rare extremes. Our study combines deep learning forecasting, interpretable representation learning, and classical ML probing, illustrating how cross-disciplinary AI methods can yield insight into foundation models

Ashikur Nobel, Jacob Matos, Honggang Wang, Hua Fang

We present iDT-diet, an intelligent digital twin prototype designed to model the long-term influence of diet quality on health biomarkers and chronic conditions. The system integrates three novel components: (i) a random forest learning model enhanced with Choquet LASSO feature selection for capturing complex, nonlinear interactions in temporal health data; (ii) a translation module that converts predictive outputs into natural language narratives of physical and biomarker states; and (iii) a generative 3D visualization engine that produces dynamic, personalized digital twins reflecting evolving health trajectories. This integration uniquely links advanced machine learning, interpretable communication, and immersive visualization within a single framework. While the current implementation focuses on retrospective digital twin generation, the system architecture supports real-time data integration, enabling continuous monitoring, predictive simulation, and personalized recommendation delivery for diet and lifestyle management.

Mahi Mann, Avinash Anand, Rajiv Ratn Shah

Mental health monitoring faces challenges from fragmented data and opaque risk scores. We present BRI-MH, an in- terpretable multimodal framework combining behavioral sig- nals with cognitive features from large language models to produce a weekly Behavioral Risk Index. Unlike prior work with isolated or black-box scores, BRI-MH offers transpar- ent, actionable insights and links continuous monitoring to adaptive feedback and therapeutic support, bridging digital phenotyping and clinical care.

Pengfei Liu, Rui Kong, Zongzhang Zhang

Hand-craft reward engineering requires domain knowledge with numerous trials and errors, while Preference-based Reinforcement Learning (PbRL) avoids manual reward design but often suffers from limited interpretability and unstable training. To address these issues, we propose a novel preference alignment framework. Our approach leverages large language models to generate sub-reward functions informed by prior knowledge and further align human preferences by optimizing the weights combining these sub-rewards. For policy learning, we introduce Policy Optimization via Pareto Regularization (POPR) which regularizes updates along Pareto-optimal directions. Experiments show that our framework improves reward quality and policy stability, achieving superior performance to expert-designed rewards across most tasks.

Yow-Fu Liou, Yu-Chien Tang, An-Zi Yen

When evaluating large language models (LLMs) for question answering tasks, a common protocol is multiple-choice question-answering (MCQA), where the model selects from a fixed set of choices. In contemporary robustness testing, researchers typically perturb instructions or introduce confusion into factual statements; however, model behavior also hinges on choice compliance: whether models remain within the canonical set A-D. We formalize this setting by asking whether the model continues to respect the interface's rules when the problem presents a tempting alternative. Our approach is interface-preserving: we append a single selectable option E while keeping the question and A-D unchanged. Then, we introduce three types of malicious option injection to assess LLMs' robustness. Experimental results highlight the vulnerability of LLMs on contradict type content of the additional option E. Our evaluation framework can effectively serve as a low-cost audit of rule adherence on existing datasets and black-box models, surfaces off-policy items, and supports interpretable model comparison for deployment.

Yousung Lee, Dongsoo Har

Dynamic head pruning in Vision Transformers (ViTs) improves efficiency by removing redundant attention heads, but existing pruning policies are often difficult to interpret and control. In this work, we propose a novel framework by integrating Sparse Autoencoders (SAEs) with dynamic pruning, leveraging their ability to disentangle dense embeddings into interpretable and controllable sparse latents. Specifically, we train an SAE on the final-layer residual embedding of the ViT and amplify the sparse latents with different strategies to alter pruning decisions. Among them, per-class steering reveals compact, class-specific head subsets that preserve accuracy. For example, bowl improves accuracy (76%→82%) while reducing head usage (0.72→0.33) via heads h2 and h5. These results show that sparse latent features enable class-specific control of dynamic pruning, effectively bridging pruning efficiency and mechanistic interpretability in ViTs.

Jun Jang, Thai Le, Bo Wang

With the rise of generative artificial intelligence (GenAI), academic dishonesty in classrooms has skyrocketed, yet the existing solutions for detecting such dishonesty often fall short. Standard "AI detectors" merely analyze one text at a time, failing to account for students' previous writings, which risks erroneous predictions. Meanwhile, existing token-based authorship verification (AV) models fail to analyze the nuances in writing styles that truly distinguish authorship. To fill this existing gap, we propose a novel AV framework that combines token-level stylometric features (e.g., POS tag patterns) with handcrafted stylistic features (e.g., sentence structure variation) to construct a comprehensive feature set. Using both benchmark corpora and real-world high school student essays, we trained multiple machine learning classifiers using the proposed feature set. Our initial experiments show that our approach outperforms the standard token-only baselines by over 25%, while offering interpretable, style-based insights. These preliminary results highlight the importance of nuanced stylistic features and suggest that a holistic AV system can provide educators with more reliable and transparent detection tools. Looking ahead, we plan to extend this work with large language models and multi-agent approaches to further enhance robustness and adaptability.

Yao Gao, Feng Li, Jeroen Van Dessel, Yi Sun, Robin Willaert

What if the next generation of human-computer interaction is not a screen... but a conversation? Large Language Models (LLMs) offer a new paradigm for interacting with computers through text, but they lack shape reasoning capabilities. We introduce Textual Anatomy Encoding (TAE), a workflow that connects LLMs with 3D anatomies. TAE employs clinician-validated semantic annotations and rule-based prompts to achieve deterministic and interpretable landmark localization. The results indicate that TAE enables LLMs to move beyond textual knowledge, achieving an accurate understanding of anatomical localization. This framework opens opportunities for diagnosis, surgical planning, and scalable medical annotation, positioning LLMs as a foundation for next-generation human–computer interaction in healthcare.

Taqiya Ehsan, Shuren Xia, Jorge Ortiz

Embodied agents must reason causally, as correlation-based models fail under intervention and distribution shift. This challenge arises in domains like robotics and cyber-physical systems, where agents balance efficiency and comfort under uncertainty. We introduce POLICYGRID, unifying causal discovery and control by treating each action as both decision and experiment. Leveraging constraint-based search, neural causal models, and language model priors with interventional validation, POLICYGRID yields adaptive, interpretable policies. Across synthetic, real-world, and live deployments, it achieves superior causal recovery (F1 = 0.89) and 2.8× better multi-objective performance than correlation-based baselines, demonstrating safe, generalizable decision-making.

Ashim Dhor, Smily Bharadwaj

Automated cancer segmentation in Whole Slide Images (WSIs) has been dominated by a paradigm of static pattern recognition, where even advanced methods leveraging Transformers, Multiple Instance Learning, or topology-aware losses remain fundamentally descriptive and correlational. To address this limitation, we reframe WSI segmentation from a descriptive task to one of causal process modeling. We introduce Topo-GraT, a novel framework featuring a Causal Growth Field (CGF) to model tumor invasion dynamics and a Causal Flow Attention (CFA) mechanism that embeds this field as an architectural prior. This causal engine is integrated within an iterative graph refinement loop that uses segmentation uncertainty to dynamically focus computational resources on the most ambiguous tissue regions. Our comprehensive experiments on multiple WSI datasets demonstrate that Topo-GraT establishes a new state-of-the-art, significantly outperforming existing methods and reducing the 95% Hausdorff Distance, a key boundary metric, by over 15%. Crucially, our framework yields the CGF as a rich, interpretable output whose structure correlates with tumor aggressiveness, positioning it as a novel biomarker for downstream prognostic tasks. By shifting the paradigm from static recognition to causal reasoning, Topo-GraT offers a more robust, efficient, and clinically insightful approach, setting a new direction for the causally-aware medical image analysis.

Adit Desai, Sudipta Roy, Mohna Chakraborty

Vision-Language Models (VLMs) are increasingly deployed across downstream tasks, yet their training data often encode social biases that surface in outputs. Unlike humans, who interpret images through contextual and social cues, VLMs process them through statistical associations, often leading to reasoning that diverges from human reasoning. By analyzing how a VLM reasons, we can understand how inherent biases are perpetuated and can adversely affect downstream performance. To examine this gap, we systematically analyze social biases in five open-source VLMs for an occupation prediction task, on the FairFace dataset. Across 32 occupations and three different prompting styles, we elicit both predictions and reasoning. Our findings show that the biased reasoning patterns systematically underlie intersectional disparities, highlighting the need to align VLM reasoning with human values before downstream deployment.

Abhilekh Borah, Niranjan Chebrolu, Kokil Jaidka

Refusals must be resilient, not brittle.” Yet guarding refusals against adversarial phrasing and shifting user contexts remains difficult: large language models (LLMs) still yield to jailbreak prompts that evade safety filters and surface harmful content. We propose Refusal Activation Steering (RAS), a training-free, inference-time method that uses contrastive activations to shift LLM responses, biasing generation trajectories toward refusals without altering model weights. The approach is modular and domain-targetable, avoiding collateral refusals on benign queries while strengthening activation- space boundaries for unsafe content. On adversarial evaluations with an 8B instruction-tuned model, we find that steering improves refusal rate by ∼ 52% and reduces attack success rate by ∼ 40%, establishing a lightweight and interpretable safety layer for robust refusal consistency. To foster further research in this domain, we have made our implementation publicly available.

Hyojun Ahn, Joongheon Kim

Large language models (LLMs) often generate hallucinations—fluent yet factually incorrect responses—that undermine reliability in knowledge-intensive tasks. Existing approaches for hallucination mitigation typically rely on external retrieval modules or probability heuristics, which either require additional resources or lack interpretability. In this work, we propose a diffusion-based hallucination detection framework (DHDF) that leverages U-Net denoising to reconstruct consensus answers from multiple LLM outputs. If the diffusion process exhibits spurious convergence away from factual ground truth, it provides a clear signal of hallucination. To quantify factual correctness, we incorporate TruthfulQA scores as a fact-grounded evaluation metric, distinguishing well-aligned models (high scores) from hallucination-prone models (low scores). Experimental results demonstrate that convergence dynamics under diffusion, combined with fact-grounded QA evaluation, offer an effective and interpretable pathway for hallucination detection without relying on external knowledge bases.

Yumin Zhou

Deep learning models offer state-of-the-art performance but their inherent opacity is a major barrier to adoption in high-stakes domains. In contrast, Takagi-Sugeno-Kang (TSK) fuzzy systems provide rule-based transparency but often lack the predictive power of deep networks. My PhD research addresses this critical trade-off by developing the Fuzzy-Modulated Linear Consequents (FMLC) framework, a novel hybrid architecture that synergizes these two paradigms. The core of FMLC is a deep neural network that processes fuzzified input features to generate context-dependent "modulators". These modulators dynamically parameterize a TSK-style linear consequent layer, creating a model that is both highly performant and inherently interpretable. My latest work, Learnable-FMLC (L-FMLC), advances this by introducing a regularized, adaptive fuzzification layer that autonomously learns the optimal fuzzy partitions from data, and a two-stage rule distillation framework to ensure interpretability remains scalable in high-dimensional problems. This research delivers a validated, theoretically-grounded, and scalable framework, contributing a significant step towards transparent and trustworthy AI.

Zijian Zhao

Recently, deep Reinforcement Learning (RL) methods have been widely used in labor management within transportation gig markets, such as ride-hailing, food delivery, and express delivery. Compared to traditional rule-based and optimization-based methods, RL can capture more information about long-term uncertainty and environmental dynamics, leading to better and non-myopic strategies. However, deep learning methods have long been criticized for their low interpretability, raising concerns about algorithmic discrimination in gig markets. Currently, most works focus on this issue from the perspective of statistical analysis and surveys. However, the underlying reasons related to the algorithms remain unclear, as most companies do not disclose their algorithms. This lack of transparency can hinder governments from designing efficient management policies to address these problems. To fill this research gap, this thesis proposal aims to develop appropriate RL methods to mimic the labor management behavior of transportation gig platforms and to propose effective policies that protect the rights of gig workers.

Fanmeng Wang

Molecular conformations, the stable three-dimensional structures corresponding to local minima on the potential energy surface, govern key molecular properties and consequently underpin a wide range of downstream tasks. However, contemporary learning-based methods often lack scalability, interpretability, and robustness, thereby significantly constraining their practical effectiveness and reliability. In this context, I will introduce my ongoing explorations and the proposed research plan to address these challenges, with the ultimate objective of developing conformation‑centric universal foundation models to accelerate scientific discovery.

Deepika SN Vemuri

This thesis aims to bridge the gap between data-driven models and symbolic learning through the lens of Concept-Based Learning, a paradigm that guides model learning through high-level, human-understandable concepts. Here, models first learn a set of concepts, subsequently using them to perform a task of interest. Prior work on concept-based models has largely focused on relatively simple classification settings, where classes are linear combinations of pre-specified concepts; treating concepts largely as tools to increase interpretability, rather than as fundamental building blocks of the learning process itself. In contrast, this thesis explores the broader potential of concepts, as the core units of representation and reasoning in neural network models, capable of shaping how models learn and generalize.

Sachin Vashistha

Large language models (LLMs) have rapidly become primary tools for searching and generating information given a carefully designed prompt (may contain few-shot examples). However, these models frequently produce factually incorrect statements that are not consistent with verifiable facts and reliable sources, raising fundamental questions about how these models store, update, and reason with facts. Improving factuality, therefore, requires more than surface-level mitigation strategies: it demands a deeper understanding of how LLMs construct and maintain world models, and how reasoning processes can be guided to remain faithful to the verifiable information. Existing strategies, such as retrieval-augmented generation, training-time alignment, post hoc verification, etc., partly address these challenges but do not provide a holistic account of how facts are internally stored, updated, or grounded in external knowledge sources. My research addresses this gap by studying factuality through the dual lens of reasoning and world modeling, asking how LLMs encode facts, how adversarial or linguistic perturbations compromise factual reasoning, and how interpretability tools can reveal and correct model vulnerabilities. In this work, I aim to develop a framework in which an LLM interacts with an explicit external knowledge source, thereby forming a robust world model for factual evaluation.