Bayesian Gated Non-Negative Contrastive Learning.
The Forty-Third International Conference on Machine Learning (ICML 2026).
Studies Bayesian Gated Non-Negative Contrastive Learning as part of PAI's work on usable XAI foundations.
Publications
The PAI publication archive.
Browse work through the PAI lens: provable foundations, actionable methods, and interpretability systems, ordered by year, area, venue, artifact, and topic.
* equal contribution † corresponding author __ / students mentored by the PI
Showing 63 of 63 publications.
Understanding Generalization and Forgetting in In-Context Continual Learning.
The Forty-Third International Conference on Machine Learning (ICML 2026).
Studies Understanding Generalization and Forgetting in In-Context Continual Learning as part of PAI's work on usable XAI foundations.
Beyond Scalars: Evaluating and Understanding LLM Reasoning via Geometric Progress and Stability.
The Forty-Third International Conference on Machine Learning (ICML 2026).
Studies Beyond Scalars: Evaluating and Understanding LLM Reasoning via Geometric Progress and Stability as part of PAI's work on large-model interpretability.
Multi-Adapter Representation Interventions via Energy Calibration.
The Forty-Third International Conference on Machine Learning (ICML 2026).
Studies Multi-Adapter Representation Interventions via Energy Calibration as part of PAI's work on large-model interpretability.
Algorithmic Recourse of In-Context Learning for Tabular Data.
The Forty-Third International Conference on Machine Learning (ICML 2026).
Studies Algorithmic Recourse of In-Context Learning for Tabular Data as part of PAI's work on interactive explanations.
Beyond Fixed Biases: Decoding the Role of Reasoning Uncertainty in MLLM Modality Conflicts.
The Forty-Third International Conference on Machine Learning (ICML 2026).
Studies Beyond Fixed Biases: Decoding the Role of Reasoning Uncertainty in MLLM Modality Conflicts as part of PAI's work on large-model interpretability.
Less is Enough: Synthesizing Diverse Data in Feature Space of LLMs.
The Forty-Third International Conference on Machine Learning (ICML 2026).
The diversity of post-training data is critical for effective downstream performance in large language models (LLMs).
Beyond External Monitors: Enhancing Transparency of Large Language Models for Easier Monitoring.
The Forty-Third International Conference on Machine Learning (ICML 2026).
Large language models (LLMs) are becoming increasingly capable, but the mechanisms of their thinking and decision-making process remain unclear.
Controlling Repetition in Protein Language Models.
The Fourteenth International Conference on Learning Representations (ICLR 2026).
Protein language models (PLMs) have enabled advances in structure prediction and de novo protein design, yet they frequently collapse into pathological repetition during generation.
The Price of Amortized Inference in Sparse Autoencoders.
The Fourteenth International Conference on Learning Representations (ICLR 2026).
Studies The Price of Amortized Inference in Sparse Autoencoders as part of PAI's work on large-model interpretability.
Dissecting Representation Misalignment in Contrastive Learning via Influence Function.
The Fourteenth International Conference on Learning Representations (ICLR 2026).
Studies Dissecting Representation Misalignment in Contrastive Learning via Influence Function as part of PAI's work on usable XAI foundations.
Evaluating Data Influence in Meta Learning.
The Fourteenth International Conference on Learning Representations (ICLR 2026).
As one of the most fundamental models, meta learning aims to effectively address few-shot learning challenges.
Predicting LLM Output Length via Entropy-Guided Representations.
The Fourteenth International Conference on Learning Representations (ICLR 2026).
The long-tailed distribution of sequence lengths in LLM serving and reinforcement learning (RL) sampling causes significant computational waste due to excessive padding in batched inference.
Dual-Kernel Adapter: Expanding Spatial Horizons for Data-Constrained Medical Image Analysis.
The Fourteenth International Conference on Learning Representations (ICLR 2026).
Adapters have become a widely adopted strategy for efficient fine-tuning of large pretrained models, particularly in resource-constrained settings.
Towards Reasoning-Preserving Unlearning in Multimodal Large Language Models.
The IEEE/CVF Conference on Computer Vision and Pattern Recognition 2026 (CVPR 2026).
Machine unlearning aims to erase requested data from trained models without full retraining.
Flattery in Motion: Benchmarking and Analyzing Sycophancy in Video-LLMs.
The 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026 Main).
As video large language models (Video-LLMs) become increasingly integrated into real-world applications that demand grounded multimodal reasoning, ensuring their factual consistency and reli
Visual Self-Fulfilling Alignment: Shaping Safety-Oriented Personas via Threat-Related Images.
The 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026 Main).
Studies Visual Self-Fulfilling Alignment: Shaping Safety-Oriented Personas via Threat-Related Images as part of PAI's work on interactive explanations.
Deciphering Cultural Representations in Large Language Models via Sparse Autoencoders.
The 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026 Findings).
Studies Deciphering Cultural Representations in Large Language Models via Sparse Autoencoders as part of PAI's work on large-model interpretability.
PIXEL: Adaptive Steering Via Position-wise Injection with eXact Estimated Levels under a Subspace Calibration.
The Web Conference (WWW 2026).
Reliable behavior control is central to deploying Large Language Models (LLMs) on the web.
Intelligent Stain-free Histology on Structural Colorimetric Nanocavities.
Advanced Science.
Optical microscopic imaging and analysis are indispensable in modern pathology for providing conclusive evidence in disease diagnosis.
Efficient Text-Attributed Graph Learning through Selective Annotation and Graph Alignment.
Transactions on Machine Learning Research (TMLR).
In the realm of Text-attributed Graphs (TAGs), traditional graph neural networks (GNNs) often fall short due to the complex textual information associated with each node.
Multimodal Deception in Explainable AI: Concept-Level Backdoor Attacks on Concept Bottleneck Models.
Transactions on Machine Learning Research (TMLR).
Studies Multimodal Deception in Explainable AI: Concept-Level Backdoor Attacks on Concept Bottleneck Models as part of PAI's work on large-model interpretability.
Towards Representation Backdoor on CLIP via Concept Confusion.
Transactions on Machine Learning Research (TMLR), Minor Revision.
Studies Towards Representation Backdoor on CLIP via Concept Confusion as part of PAI's work on large-model interpretability.
Usable XAI: 10 Strategies Towards Exploiting Explainability in the LLM Era.
The ACM Transactions on Knowledge Discovery from Data (TKDD).
Studies Usable XAI: 10 Strategies Towards Exploiting Explainability in the LLM Era as part of PAI's work on large-model interpretability.
EAP-GP: Mitigating Saturation Effect in Gradient-based Automated Circuit Identification.
Conference on Neural Information Processing Systems (NeurIPS 2025).
Understanding the internal mechanisms of transformer-based language models remains challenging.
Understanding How Value Neurons Shape the Generation of Specified Values in LLMs.
Conference on Empirical Methods in Natural Language Processing (EMNLP 2025 Findings).
Studies Understanding How Value Neurons Shape the Generation of Specified Values in LLMs as part of PAI's work on large-model interpretability.
Can Large Language Models Identify Implicit Suicidal Ideation? An Empirical Evaluation.
Conference on Empirical Methods in Natural Language Processing (EMNLP 2025 Findings).
We present a comprehensive evaluation framework for assessing Large Language Models'(LLMs) capabilities in suicide prevention, focusing on two critical aspects: the Identification of Implici
PgM: Partitioner Guided Modal Learning Framework.
The 33rd ACM International Conference on Multimedia (ACM MM 2025).
Studies PgM: Partitioner Guided Modal Learning Framework as part of PAI's work on usable XAI foundations.
From Guesswork to Guarantee: Towards Faithful Multimedia Web Forecasting with TimeSieve.
The 33rd ACM International Conference on Multimedia (ACM MM 2025).
Studies From Guesswork to Guarantee: Towards Faithful Multimedia Web Forecasting with TimeSieve as part of PAI's work on usable XAI foundations.
Semi-supervised Concept Bottleneck Models.
IEEE/CVF International Conference on Computer Vision (ICCV 2025).
Concept Bottleneck Models (CBMs) have garnered increasing attention due to their ability to provide concept-based explanations for black-box deep learning models while achieving high final p
Stable Vision Concept Transformers for Medical Diagnosis.
The 2025 European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML-PKDD 2025).
Studies Stable Vision Concept Transformers for Medical Diagnosis as part of PAI's work on scientific discovery.
COMPKE: Complex Question Answering under Knowledge Editing.
Annual Meeting of the Association for Computational Linguistics (ACL 2025 Findings).
Studies COMPKE: Complex Question Answering under Knowledge Editing as part of PAI's work on large-model interpretability.
CODEMENV: Benchmarking Large Language Models on Code Migration.
Annual Meeting of the Association for Computational Linguistics (ACL 2025 Findings).
Studies CODEMENV: Benchmarking Large Language Models on Code Migration as part of PAI's work on large-model interpretability.
Understanding the Repeat Curse in Large Language Models from a Feature Perspective.
Annual Meeting of the Association for Computational Linguistics (ACL 2025 Findings).
Studies Understanding the Repeat Curse in Large Language Models from a Feature Perspective as part of PAI's work on large-model interpretability.
Fraud-R1: A Multi-Round Benchmark for Assessing the Robustness of LLM Against Augmented Fraud and Phishing Inducements.
Annual Meeting of the Association for Computational Linguistics (ACL 2025 Findings).
Studies Fraud-R1: A Multi-Round Benchmark for Assessing the Robustness of LLM Against Augmented Fraud and Phishing Inducements as part of PAI's work on large-model interpretability.
Locate-then-edit for Multi-hop Factual Recall under Knowledge Editing.
The 42nd International Conference on Machine Learning (ICML 2025).
Studies Locate-then-edit for Multi-hop Factual Recall under Knowledge Editing as part of PAI's work on large-model interpretability.
Editable Concept Bottleneck Models.
The 42nd International Conference on Machine Learning (ICML 2025).
Studies Editable Concept Bottleneck Models as part of PAI's work on usable XAI foundations.
SMoA: Improving Multi-agent Large Language Models with Sparse Mixture-of-Agents.
The 29th Pacific-Asia Conference on Knowledge Discovery and Data Mining (PAKDD 2025).
While multi-agent systems have been shown to significantly enhance the performance of Large Language Models (LLMs) across various tasks and applications, the dense interaction between scalin
DRIVE: Dependable Robust Interpretable Visionary Ensemble Framework in Autonomous Driving.
IEEE International Conference on Robotics and Automation (ICRA 2025).
Recent advancements in autonomous driving have seen a paradigm shift towards end-to-end learning paradigms, which map sensory inputs directly to driving actions, thereby enhancing the robust
Backdooring Vision-Language Models with Out-Of-Distribution Data.
The 13th International Conference on Learning Representations (ICLR 2025).
Studies Backdooring Vision-Language Models with Out-Of-Distribution Data as part of PAI's work on large-model interpretability.
Mechanistic Unveiling of Transformer Circuits: Self-Influence as a Key to Model Reasoning.
Findings of the 2025 Annual Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics (NAACL 2025 Findings).
Transformer-based language models have achieved significant success; however, their internal mechanisms remain largely opaque due to the complexity of non-linear interactions and high-dimensional operations.
Fair Text-to-Image Diffusion via Fair Mapping.
The Thirty-Ninth AAAI Conference on Artificial Intelligence (AAAI 2025).
Studies Fair Text-to-Image Diffusion via Fair Mapping as part of PAI's work on large-model interpretability.
Tuning-Free Accountable Intervention for LLM Deployment - A Metacognitive Approach.
The Thirty-Ninth AAAI Conference on Artificial Intelligence (AAAI 2025).
Studies Tuning-Free Accountable Intervention for LLM Deployment - A Metacognitive Approach as part of PAI's work on large-model interpretability.
Faithful Interpretation for Graph Neural Networks.
Transactions on Machine Learning Research (TMLR).
Currently, attention mechanisms have garnered increasing attention in Graph Neural Networks (GNNs), such as Graph Attention Networks (GATs) and Graph Transformers (GTs).
Towards Stable and Explainable Attention Mechanisms.
IEEE Transactions on Knowledge and Data Engineering (TKDE).
Currently, attention mechanism has become a standard fixture in most state-of-the-art natural language processing (NLP) models, not only due to the outstanding performance it could gain but
Towards Multi-dimensional Explanation Alignment for Medical Classification.
The Conference on Neural Information Processing Systems (NeurIPS 2024).
The lack of interpretability in the field of medical image analysis has significant ethical and legal implications.
Private Language Models via Truncated Laplacian Mechanism.
Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP 2024 Main).
Recently it has been shown that deep learning models for NLP tasks are prone to attacks that can even reconstruct the verbatim training texts.
Dissecting Fine-Tuning Unlearning in Large Language Models.
Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP 2024 Main).
Studies Dissecting Fine-Tuning Unlearning in Large Language Models as part of PAI's work on large-model interpretability.
UniMEEC: Towards Unified Multimodal Emotion Recognition and Emotion Cause.
Findings of the 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP 2024 Findings).
Multimodal emotion recognition in conversation (MERC) and multimodal emotion-cause pair extraction (MECPE) have recently garnered significant attention.
SATO: Stable Text-to-Motion Framework.
The 32nd ACM Multimedia Conference (ACM MM 2024).
Is the Text to Motion model robust? Recent advancements in Text to Motion models primarily stem from more accurate predictions of specific actions.
Multi-hop Question Answering under Temporal Knowledge Editing.
The 1st Conference on Language Modeling (COLM 2024).
Studies Multi-hop Question Answering under Temporal Knowledge Editing as part of PAI's work on large-model interpretability.
MONAL: Model Autophagy Analysis for Modeling Human-AI Interactions.
The 1st Conference on Language Modeling (COLM 2024).
Studies MONAL: Model Autophagy Analysis for Modeling Human-AI Interactions as part of PAI's work on large-model interpretability.
Improving Interpretation Faithfulness for Vision Transformers.
The 41st International Conference on Machine Learning (ICML 2024).
Vision Transformers (ViTs) have achieved state-of-the-art performance for various vision tasks.
Differentially Private Natural Language Models: Recent Advances and Future Directions.
Findings of the 2024 European Chapter of the Association for Computational Linguistics (EACL 2024 Findings).
Recent developments in deep learning have led to great success in various natural language processing (NLP) tasks.
Faithful Vision-Language Interpretation via Concept Bottleneck Models.
The 12th International Conference on Learning Representations (ICLR 2024).
Studies Faithful Vision-Language Interpretation via Concept Bottleneck Models as part of PAI's work on usable XAI foundations.
Faster Rates of Private Stochastic Convex Optimization.
Journal of Machine Learning Research, Volume 25, 114 (2024), Pages 1−41 (JMLR).
Studies Faster Rates of Private Stochastic Convex Optimization as part of PAI's work on usable XAI foundations.
Privacy-preserving Sparse Generalized Eigenvalue Problem.
The 26th International Conference on Artificial Intelligence and Statistics (AISTATS 2023).
Studies Privacy-preserving Sparse Generalized Eigenvalue Problem as part of PAI's work on usable XAI foundations.
SEAT: Stable and Explainable Attention.
The 37th AAAI Conference on Artificial Intelligence (AAAI 2023).
Attention mechanism has become a standard fixture in many state-of-the-art natural language processing (NLP) models, not only due to its outstanding performance, but also because it provides
Finite Sample Guarantees of Differentially Private Expectation Maximization Algorithm.
The 26th European Conference on Artificial Intelligence (ECAI 2023).
Studies Finite Sample Guarantees of Differentially Private Expectation Maximization Algorithm as part of PAI's work on usable XAI foundations.
Generalized Linear Models in Non-interactive Local Differential Privacy with Public Data.
Journal of Machine Learning Research, Volume 24, 132 (2023), Pages 1-57 (JMLR).
Studies Generalized Linear Models in Non-interactive Local Differential Privacy with Public Data as part of PAI's work on usable XAI foundations.
Nearly Optimal Rates of Privacy-preserving Sparse Generalized Eigenvalue Problem.
IEEE Transactions on Knowledge and Data Engineering, 2023 (01): 1-14 (TKDE).
Studies Nearly Optimal Rates of Privacy-preserving Sparse Generalized Eigenvalue Problem as part of PAI's work on usable XAI foundations.
Faster Rates of Differentially Private Stochastic Convex Optimization.
Proceedings of The 33rd International Conference on Algorithmic Learning Theory (ALT 2022).
Studies Faster Rates of Differentially Private Stochastic Convex Optimization as part of PAI's work on usable XAI foundations.
High Dimensional Differentially Private Stochastic Optimization with Heavy-tailed Data.
Proceedings of the 41st ACM SIGMOD-SIGACT-SIGAI Symposium on Principles of Database Systems (PODS 2022).
Studies High Dimensional Differentially Private Stochastic Optimization with Heavy-tailed Data as part of PAI's work on usable XAI foundations.