arXiv subjects

Jiali Cheng

Publications and source records attributed to Jiali Cheng.

At least 19 recordsLinked to original sources

arXiv2026

Understanding Machine Unlearning Through the Lens of Mode Connectivity

Machine Unlearning aims to remove undesired information from trained models without full retraining from scratch. Despite recent progress, the loss landscape and optimization geometry of unlearning are poorly understood. In this paper, we study machine unlearning through the lens of mode connectivity--the phenomenon that independently trained models can often be connected by smooth low-loss paths in parameter space. We introduce {\em mode connectivity in unlearning} (MCU) and evaluate it across a range of settings, including curriculum learning, second-order optimization, and connectivity across different unlearning methods. We find that many unlearned models lie in connected basins with smooth retain/forget behavior, while changes in training dynamics can move solutions into different basins. MCU also reveals that models within the same basin can differ substantially on privacy metrics, and that unlearning progresses nonlinearly from the original model to the unlearned model. In addition, linear connectivity suggests that most approximate unlearning methods are mechanistically distinct from retraining. Finally, MCU-based ensembling can improve generalization and robustness to relearning attacks, and MCU smoothness correlates with unlearning difficulty. To our knowledge, this is the first study of machine unlearning through the lens of mode connectivity.

cs.LG↗

arXiv2026

Fair Cognitive Impairment Detection Through Unlearning

Mild Cognitive Impairment (MCI) is a medical condition characterized by a noticeable decline in memory, language, or thinking abilities. MCI detection from spontaneous speech is promising for scalable screening. However, learned models often exploit demographic cues correlated with labels, resulting in a large performance gap across subgroups. We present a multimodal framework that combines (i) cross-model fusion between modalities (speech, text, and image), and (ii) unlearning using gradient reversal that discourages the shared embedding from encoding task-irrelevant demographic attributes. Evaluated on the multilingual benchmarks TAUKADIAL and PREPARE, our method outperforms the state-of-the-art multilingual and multimodal baseline in MCI classification while substantially reducing the performance gap across patient subgroups (sex and language). We further analyze transfer across datasets, showing that demographic unlearning helps learn more robust representations for MCI detection.

cs.LG↗

arXiv2026

TRACER: Token ReAssignment for Concept ERasure in Generative Recommendation

Generative recommendation formulates next-item prediction as autoregressive generation over semantic ID (SID) sequences derived from users' historical interactions, making modern recommender systems structurally similar to large language models (LLMs). As privacy and safety concerns grow, these systems increasingly require concept unlearning to remove sensitive or harmful concepts associated with items. However, existing LLM unlearning methods cannot be directly applied to generative recommendation. Unlike word tokens with explicit semantics, SIDs are abstract identifiers that are often shared by both forget and retain items, leading to severe conflicts between concept removal and recommendation utility preservation. To address this challenge, we propose TRACER, an end-to-end concept unlearning framework based on token reassignment. Rather than directly suppressing shared SIDs, TRACER reassigns concept-related items to alternative tokens that better facilitate forgetting while minimizing side effects on retained items. We further introduce a coherence regularizer to preserve semantic consistency among retain items during unlearning. Experiments on real-world recommendation datasets demonstrate that TRACER effectively removes target concepts while substantially better preserving recommendation utility than existing unlearning baselines.

cs.IR↗

arXiv2026

CURE:Circuit-Aware Unlearning for LLM-based Recommendation

Recent advances in large language models (LLMs) have opened new opportunities for recommender systems by enabling rich semantic understanding and reasoning about user interests and item attributes. However, as privacy regulations tighten, incorporating user data into LLM-based recommendation (LLMRec) introduces significant privacy risks, making unlearning algorithms increasingly crucial for practical deployment. Despite growing interest in LLMRec unlearning, most existing approaches formulate unlearning as a weighted combination of forgetting and retaining objectives while updating model parameters in a uniform manner. Such formulations inevitably induce gradient conflicts between the two objectives, leading to unstable optimization and resulting in either ineffective unlearning or severe degradation of model utility. Moreover, the unlearning procedure remains largely black-box, undermining its transparency and trustworthiness. To tackle these challenges, we propose CURE, a circuit-aware unlearning framework that disentangles model components into functionally distinct subsets and selectively updates them. Here, a circuit refers to a computational subgraph that is causally responsible for task-specific behaviors. Specifically, we extract the core circuits underlying item recommendation and analyze how individual modules within these circuits contribute to the forget and retain objectives. Based on this analysis, these modules are categorized into forget-specific, retain-specific, and task-shared groups, each subject to function-specific update rules to mitigate gradient conflicts during unlearning. Experiments on real-world datasets show that our approach achieves more effective unlearning than existing baselines.

cs.IR↗

arXiv2026

Investigating Tool-Memory Conflicts in Tool-Augmented LLMs

Tool-augmented large language models (LLMs) have powered many applications. However, they are likely to suffer from knowledge conflict. In this paper, we propose a new type of knowledge conflict -- Tool-Memory Conflict (TMC), where the internal parametric knowledge contradicts with the external tool knowledge for tool-augmented LLMs. We find that existing LLMs, though powerful, suffer from TMC, especially on STEM-related tasks. We also uncover that under different conditions, tool knowledge and parametric knowledge may be prioritized differently. We then evaluate existing conflict resolving techniques, including prompting-based and RAG-based methods. Results show that none of these approaches can effectively resolve tool-memory conflicts.

cs.SE↗

arXiv2026

A Mechanistic Perspective and Circuit-Guided Difficulty Metric for Unlearning

Machine unlearning is becoming essential for building trustworthy and compliant language models. Yet unlearning success varies considerably across individual samples: some are reliably erased, while others persist despite the same procedure. We argue that this disparity is not only a data-side phenomenon, but also reflects model-internal mechanisms that encode and protect memorized information. We study this problem from a mechanistic perspective based on model circuits--structured interaction pathways that govern how predictions are formed. We propose Circuit-guided Unlearning Difficulty (CUD), a {\em pre-unlearning} metric that assigns each sample a continuous difficulty score using circuit-level signals. Extensive experiments demonstrate that CUD reliably separates intrinsically easy and hard samples, and remains stable across unlearning methods. We identify key circuit-level patterns that reveal a mechanistic signature of difficulty: easy-to-unlearn samples are associated with shorter, shallower interactions concentrated in earlier-to-intermediate parts of the original model, whereas hard samples rely on longer and deeper pathways closer to late-stage computation. Compared to existing qualitative studies, CUD takes a first step toward a principled, fine-grained, and interpretable analysis of unlearning difficulty; and motivates the development of unlearning methods grounded in model mechanisms.

cs.LG↗

arXiv2025

Do Students Debias Like Teachers? On the Distillability of Bias Mitigation Methods

Knowledge distillation (KD) is an effective method for model compression and transferring knowledge between models. However, its effect on model's robustness against spurious correlations that degrade performance on out-of-distribution data remains underexplored. This study investigates the effect of knowledge distillation on the transferability of ``debiasing'' capabilities from teacher models to student models on natural language inference (NLI) and image classification tasks. Through extensive experiments, we illustrate several key findings: (i) overall the debiasing capability of a model is undermined post-KD; (ii) training a debiased model does not benefit from injecting teacher knowledge; (iii) although the overall robustness of a model may remain stable post-distillation, significant variations can occur across different types of biases; and (iv) we pin-point the internal attention pattern and circuit that causes the distinct behavior post-KD. Given the above findings, we propose three effective solutions to improve the distillability of debiasing methods: developing high quality data for augmentation, implementing iterative knowledge distillation, and initializing student models with weights obtained from teacher models. To the best of our knowledge, this is the first study on the effect of KD on debiasing and its interenal mechanism at scale. Our findings provide understandings on how KD works and how to design better debiasing methods.

cs.LG↗

arXiv2025

WebATLAS: An LLM Agent with Experience-Driven Memory and Action Simulation

Large Language Model (LLM) web agents often struggle with long-horizon web navigation and web task completion in new websites, producing inefficient action sequences unless fine-tuned on environment-specific data. We show that experience-driven memory, combined with look-ahead action simulation, is sufficient for LLM agents to adapt to unseen web environments by remembering past failures and predicting the consequences of future actions. We introduce WebATLAS (Actor-Critic Task-completion with Look-ahead Action Simulation), a memory-augmented LLM web agent that learns a lightweight internal model of the environment from interaction experience and performs hypothetical action rollouts before acting in the real world. WebATLAS builds a persistent cognitive map via curiosity-driven exploration, stores interaction outcomes as experience-based memory, and evaluates candidate actions in cognitive space using a planner--simulator--critic loop. This enables the agent to reuse past experience, avoid previously unsuccessful behaviors, and generate more efficient plans. We evaluate WebATLAS on the WebArena-Lite benchmark for autonomous web navigation and demonstrate a success rate of 63%, outperforming the previous state-of-the-art at 53.9%. Unlike previous systems, our modular architecture requires no website-specific LLM fine-tuning. Ablation studies confirm that experience-driven memory, look-ahead action simulation, and hierarchical replanning play complementary roles in enabling robust, training-free web agents.

cs.LG↗

arXiv2025

FUTURE: Flexible Unlearning for Tree Ensemble

Tree ensembles are widely recognized for their effectiveness in classification tasks, achieving state-of-the-art performance across diverse domains, including bioinformatics, finance, and medical diagnosis. With increasing emphasis on data privacy and the \textit{right to be forgotten}, several unlearning algorithms have been proposed to enable tree ensembles to forget sensitive information. However, existing methods are often tailored to a particular model or rely on the discrete tree structure, making them difficult to generalize to complex ensembles and inefficient for large-scale datasets. To address these limitations, we propose FUTURE, a novel unlearning algorithm for tree ensembles. Specifically, we formulate the problem of forgetting samples as a gradient-based optimization task. In order to accommodate non-differentiability of tree ensembles, we adopt the probabilistic model approximations within the optimization framework. This enables end-to-end unlearning in an effective and efficient manner. Extensive experiments on real-world datasets show that FUTURE yields significant and successful unlearning performance.

cs.LG↗

arXiv2025

Speech Unlearning

We introduce machine unlearning for speech tasks, a novel and underexplored research problem that aims to efficiently and effectively remove the influence of specific data from trained speech models without full retraining. This has important applications in privacy preservation, removal of outdated or noisy data, and bias mitigation. While machine unlearning has been studied in computer vision and natural language processing, its application to speech is largely unexplored due to the high-dimensional, sequential, and speaker-dependent nature of speech data. We define two fundamental speech unlearning tasks: sample unlearning, which removes individual data points (e.g., a voice recording), and class unlearning, which removes an entire category (e.g., all data from a speaker), while preserving performance on the remaining data. Experiments on keyword spotting and speaker identification demonstrate that unlearning speech data is significantly more challenging than unlearning image or text data. We conclude with key future directions in this area, including structured training, robust evaluation, feature-level unlearning, broader applications, scalable methods, and adversarial robustness.

cs.LG↗

arXiv2025

Unveil Multi-Picture Descriptions for Multilingual Mild Cognitive Impairment Detection via Contrastive Learning

Detecting Mild Cognitive Impairment from picture descriptions is critical yet challenging, especially in multilingual and multiple picture settings. Prior work has primarily focused on English speakers describing a single picture (e.g., the 'Cookie Theft'). The TAUKDIAL-2024 challenge expands this scope by introducing multilingual speakers and multiple pictures, which presents new challenges in analyzing picture-dependent content. To address these challenges, we propose a framework with three components: (1) enhancing discriminative representation learning via supervised contrastive learning, (2) involving image modality rather than relying solely on speech and text modalities, and (3) applying a Product of Experts (PoE) strategy to mitigate spurious correlations and overfitting. Our framework improves MCI detection performance, achieving a +7.1% increase in Unweighted Average Recall (UAR) (from 68.1% to 75.2%) and a +2.9% increase in F1 score (from 80.6% to 83.5%) compared to the text unimodal baseline. Notably, the contrastive learning component yields greater gains for the text modality compared to speech. These results highlight our framework's effectiveness in multilingual and multi-picture MCI detection.

cs.CL↗

arXiv2025

Understanding Machine Unlearning Through the Lens of Mode Connectivity

cs.LG↗

arXiv2025

Linguistic Blind Spots of Large Language Models

Large language models (LLMs) are the foundation of many AI applications today. However, despite their remarkable proficiency in generating coherent text, questions linger regarding their ability to perform fine-grained linguistic annotation tasks, such as detecting nouns or verbs, or identifying more complex syntactic structures like clauses in input texts. These tasks require precise syntactic and semantic understanding of input text, and when LLMs underperform on specific linguistic structures, it raises concerns about their reliability for detailed linguistic analysis and whether their (even correct) outputs truly reflect an understanding of the inputs. In this paper, we empirically study the performance of recent LLMs on fine-grained linguistic annotation tasks. Through a series of experiments, we find that recent LLMs show limited efficacy in addressing linguistic queries and often struggle with linguistically complex inputs. We show that the most capable LLM (Llama3-70b) makes notable errors in detecting linguistic structures, such as misidentifying embedded clauses, failing to recognize verb phrases, and confusing complex nominals with clauses. Our results provide insights to inform future advancements in LLM design and development.

cs.CL↗

arXiv2025

FROG: Fair Removal on Graphs

With growing emphasis on privacy regulations, machine unlearning has become increasingly critical in real-world applications such as social networks and recommender systems, many of which are naturally represented as graphs. However, existing graph unlearning methods often modify nodes or edges indiscriminately, overlooking their impact on fairness. For instance, forgetting links between users of different genders may inadvertently exacerbate group disparities. To address this issue, we propose a novel framework that jointly optimizes both the graph structure and the model to achieve fair unlearning. Our method rewires the graph by removing redundant edges that hinder forgetting while preserving fairness through targeted edge augmentation. We further introduce a worst-case evaluation mechanism to assess robustness under challenging scenarios. Experiments on real-world datasets show that our approach achieves more effective and fair unlearning than existing baselines.

cs.LG↗

arXiv2025

FairFlow: Mitigating Dataset Biases through Undecided Learning

Language models are prone to dataset biases, known as shortcuts and spurious correlations in data, which often result in performance drop on new data. We present a new debiasing framework called ``FairFlow'' that mitigates dataset biases by learning to be undecided in its predictions for data samples or representations associated with known or unknown biases. The framework introduces two key components: a suite of data and model perturbation operations that generate different biased views of input samples, and a contrastive objective that learns debiased and robust representations from the resulting biased views of samples. Experiments show that FairFlow outperforms existing debiasing methods, particularly against out-of-domain and hard test samples without compromising the in-domain performance

cs.LG↗

arXiv2025

EqualizeIR: Mitigating Linguistic Biases in Retrieval Models

This study finds that existing information retrieval (IR) models show significant biases based on the linguistic complexity of input queries, performing well on linguistically simpler (or more complex) queries while underperforming on linguistically more complex (or simpler) queries. To address this issue, we propose EqualizeIR, a framework to mitigate linguistic biases in IR models. EqualizeIR uses a linguistically biased weak learner to capture linguistic biases in IR datasets and then trains a robust model by regularizing and refining its predictions using the biased weak learner. This approach effectively prevents the robust model from overfitting to specific linguistic patterns in data. We propose four approaches for developing linguistically-biased models. Extensive experiments on several datasets show that our method reduces performance disparities across linguistically simple and complex queries, while improving overall retrieval performance.

cs.CL↗

arXiv2025

Tool Unlearning for Tool-Augmented LLMs

Tool-augmented large language models (LLMs) are often trained on datasets of query-response pairs, which embed the ability to use tools or APIs directly into the parametric knowledge of LLMs. Tool-augmented LLMs need the ability to forget learned tools due to security vulnerabilities, privacy regulations, or tool deprecations. However, ``tool unlearning'' has not been investigated in unlearning literature. We introduce this novel task, which requires addressing distinct challenges compared to traditional unlearning: knowledge removal rather than forgetting individual samples, the high cost of optimizing LLMs, and the need for principled evaluation metrics. To bridge these gaps, we propose ToolDelete, the first approach for unlearning tools from tool-augmented LLMs. It implements three key properties to address the above challenges for effective tool unlearning and introduces a new membership inference attack (MIA) model for effective evaluation. Extensive experiments on multiple tool learning datasets and tool-augmented LLMs show that ToolDelete effectively unlearns randomly selected tools, while preserving the LLM's knowledge on non-deleted tools and maintaining performance on general tasks.

cs.LG↗

arXiv2024

MedDec: A Dataset for Extracting Medical Decisions from Discharge Summaries

Medical decisions directly impact individuals' health and well-being. Extracting decision spans from clinical notes plays a crucial role in understanding medical decision-making processes. In this paper, we develop a new dataset called "MedDec", which contains clinical notes of eleven different phenotypes (diseases) annotated by ten types of medical decisions. We introduce the task of medical decision extraction, aiming to jointly extract and classify different types of medical decisions within clinical notes. We provide a comprehensive analysis of the dataset, develop a span detection model as a baseline for this task, evaluate recent span detection approaches, and employ a few metrics to measure the complexity of data samples. Our findings shed light on the complexities inherent in clinical decision extraction and enable future work in this area of research. The dataset and code are available through https://github.com/CLU-UML/MedDec.

cs.CL↗