arXiv subjects

Xiaodan Zhu

Publications and source records attributed to Xiaodan Zhu.

At least 37 records · Page 2Linked to original sources

arXiv2025

Preemptive Detection and Correction of Misaligned Actions in LLM Agents

Deploying LLM-based agents in real-life applications often faces a critical challenge: the misalignment between agents' behavior and user intent. Such misalignment may lead agents to unintentionally execute critical actions that carry negative outcomes (e.g., accidentally triggering a "buy-now" in web shopping), resulting in undesirable or even irreversible consequences. Although addressing these issues is crucial, the preemptive detection and correction of misaligned actions remains relatively underexplored. To fill this gap, we introduce InferAct, a novel approach that leverages the belief reasoning ability of LLMs, grounded in Theory-of-Mind, to detect misaligned actions before execution. Once the misalignment is detected, InferAct alerts users for timely correction, preventing adverse outcomes and enhancing the reliability of LLM agents' decision-making processes. Experiments on three widely used tasks demonstrate that InferAct achieves up to 20% improvements on Marco-F1 against baselines in misaligned action detection. An in-depth evaluation of misalignment correction further highlights InferAct's effectiveness in improving agent alignment.

cs.CL↗

arXiv2025

Robust Utility-Preserving Text Anonymization Based on Large Language Models

Anonymizing text that contains sensitive information is crucial for a wide range of applications. Existing techniques face the emerging challenges of the re-identification ability of large language models (LLMs), which have shown advanced capability in memorizing detailed information and reasoning over dispersed pieces of patterns to draw conclusions. When defending against LLM-based re-identification, anonymization could jeopardize the utility of the resulting anonymized data in downstream tasks. In general, the interaction between anonymization and data utility requires a deeper understanding within the context of LLMs. In this paper, we propose a framework composed of three key LLM-based components: a privacy evaluator, a utility evaluator, and an optimization component, which work collaboratively to perform anonymization. Extensive experiments demonstrate that the proposed model outperforms existing baselines, showing robustness in reducing the risk of re-identification while preserving greater data utility in downstream tasks. We provide detailed studies on these core modules. To consider large-scale and real-time applications, we investigate the distillation of the anonymization capabilities into lightweight models. All of our code and datasets will be made publicly available at https://github.com/UKPLab/acl2025-rupta.

cs.CL↗

arXiv2025

Fine-Tuning on Diverse Reasoning Chains Drives Within-Inference CoT Refinement in LLMs

Requiring a large language model (LLM) to generate intermediary reasoning steps, known as Chain of Thought (CoT), has been shown to be an effective way of boosting performance. Previous approaches have focused on generating multiple independent CoTs, combining them through ensembling or other post-hoc strategies to enhance reasoning. In this work, we introduce a novel approach where LLMs are fine-tuned to generate a sequence of Diverse Chains of Thought (DCoT) within a single inference step, which is fundamentally different from prior work that primarily operate on parallel CoT generations. DCoT allows LLMs to gain the ability to perform within-inference refinement of reasoning chains without requiring external feedback. Through a rigorous set of experiments spanning a wide range of tasks that require various reasoning types, we show that fine-tuning on DCoT improves performance over the CoT baseline across model families and scales (1.3B to 70B). These improvements are particularly impactful for tasks with a large result state space, such as those involving numeric answers. Our work is also significant because both quantitative analyses and manual evaluations reveal the observed gains stem from the models' ability to refine an initial reasoning chain by generating a second, improved chain within the same inference step, demonstrating previously elusive self-improvement. Our code and data are publicly available at https://github.com/UKPLab/acl2025-diverse-cot.

cs.CL↗

arXiv2025

Red-Teaming for Inducing Societal Bias in Large Language Models

Ensuring the safe deployment of AI systems is critical in industry settings where biased outputs can lead to significant operational, reputational, and regulatory risks. Thorough evaluation before deployment is essential to prevent these hazards. Red-teaming addresses this need by employing adversarial attacks to develop guardrails that detect and reject biased or harmful queries, enabling models to be retrained or steered away from harmful outputs. However, most red-teaming efforts focus on harmful or unethical instructions rather than addressing social bias, leaving this critical area under-explored despite its significant real-world impact, especially in customer-facing systems. We propose two bias-specific red-teaming methods, Emotional Bias Probe (EBP) and BiasKG, to evaluate how standard safety measures for harmful content affect bias. For BiasKG, we refactor natural language stereotypes into a knowledge graph. We use these attacking strategies to induce biased responses from several open- and closed-source language models. Unlike prior work, these methods specifically target social bias. We find our method increases bias in all models, even those trained with safety guardrails. Our work emphasizes uncovering societal bias in LLMs through rigorous evaluation, and recommends measures ensure AI safety in high-stakes industry deployments.

cs.CL↗

arXiv2025

FAIIR: Building Toward A Conversational AI Agent Assistant for Youth Mental Health Service Provision

The world's healthcare systems and mental health agencies face both a growing demand for youth mental health services, alongside a simultaneous challenge of limited resources. Here, we focus on frontline crisis support, where Crisis Responders (CRs) engage in conversations for youth mental health support and assign an issue tag to each conversation. In this study, we develop FAIIR (Frontline Assistant: Issue Identification and Recommendation), an advanced tool leveraging an ensemble of domain-adapted and fine-tuned transformer models trained on a large conversational dataset comprising 780,000 conversations. The primary aim is to reduce the cognitive burden on CRs, enhance the accuracy of issue identification, and streamline post-conversation administrative tasks. We evaluate FAIIR on both retrospective and prospective conversations, emphasizing human-in-the-loop design with active CR engagement for model refinement, consensus-building, and overall assessment. Our results indicate that FAIIR achieves an average AUCROC of 94%, a sample average F1-score of 64%, and a sample average recall score of 81% on the retrospective test set. We also demonstrate the robustness and generalizability of the FAIIR tool during the silent testing phase, with less than a 2% drop in all performance metrics. Notably, CRs' responses exhibited an overall agreement of 90.9% with FAIIR's predictions. Furthermore, expert agreement with FAIIR surpassed their agreement with the original labels. To conclude, our findings indicate that assisting with the identification of issues of relevance helps reduce the burden on CRs, ensuring that appropriate resources can be provided and that active rescues and mandatory reporting can take place in critical situations requiring immediate de-escalation.

cs.AI↗

arXiv2025

Error Diversity Matters: An Error-Resistant Ensemble Method for Unsupervised Dependency Parsing

We address unsupervised dependency parsing by building an ensemble of diverse existing models through post hoc aggregation of their output dependency parse structures. We observe that these ensembles often suffer from low robustness against weak ensemble components due to error accumulation. To tackle this problem, we propose an efficient ensemble-selection approach that considers error diversity and avoids error accumulation. Results demonstrate that our approach outperforms each individual model as well as previous ensemble techniques. Additionally, our experiments show that the proposed ensemble-selection method significantly enhances the performance and robustness of our ensemble, surpassing previously proposed strategies, which have not accounted for error diversity.

cs.CL↗

arXiv2025

Exploring the Role of Reasoning Structures for Constructing Proofs in Multi-Step Natural Language Reasoning with Large Language Models

When performing complex multi-step reasoning tasks, the ability of Large Language Models (LLMs) to derive structured intermediate proof steps is important for ensuring that the models truly perform the desired reasoning and for improving models' explainability. This paper is centred around a focused study: whether the current state-of-the-art generalist LLMs can leverage the structures in a few examples to better construct the proof structures with \textit{in-context learning}. Our study specifically focuses on structure-aware demonstration and structure-aware pruning. We demonstrate that they both help improve performance. A detailed analysis is provided to help understand the results.

cs.CL↗

arXiv2024

SMARTCAL: An Approach to Self-Aware Tool-Use Evaluation and Calibration

The tool-use ability of Large Language Models (LLMs) has a profound impact on a wide range of industrial applications. However, LLMs' self-control and calibration capability in appropriately using tools remains understudied. The problem is consequential as it raises potential risks of degraded performance and poses a threat to the trustworthiness of the models. In this paper, we conduct a study on a family of state-of-the-art LLMs on three datasets with two mainstream tool-use frameworks. Our study reveals the tool-abuse behavior of LLMs, a tendency for models to misuse tools with overconfidence. We also find that this is a common issue regardless of model capability. Accordingly, we propose a novel approach, \textit{SMARTCAL}, to mitigate the observed issues, and our results show an average of 8.6 percent increase in the QA performance and a 21.6 percent decrease in Expected Calibration Error (ECE) compared to baseline models.

cs.LG↗

arXiv2024

Calibration Attacks: A Comprehensive Study of Adversarial Attacks on Model Confidence

In this work, we highlight and perform a comprehensive study on calibration attacks, a form of adversarial attacks that aim to trap victim models to be heavily miscalibrated without altering their predicted labels, hence endangering the trustworthiness of the models and follow-up decision making based on their confidence. We propose four typical forms of calibration attacks: underconfidence, overconfidence, maximum miscalibration, and random confidence attacks, conducted in both black-box and white-box setups. We demonstrate that the attacks are highly effective on both convolutional and attention-based models: with a small number of queries, they seriously skew confidence without changing the predictive performance. Given the potential danger, we further investigate the effectiveness of a wide range of adversarial defence and recalibration methods, including our proposed defences specifically designed for calibration attacks to mitigate the harm. From the ECE and KS scores, we observe that there are still significant limitations in handling calibration attacks. To the best of our knowledge, this is the first dedicated study that provides a comprehensive investigation on calibration-focused attacks. We hope this study helps attract more attention to these types of attacks and hence hamper their potential serious damages. To this end, this work also provides detailed analyses to understand the characteristics of the attacks. Our code is available at https://github.com/PhenetOs/CalibrationAttack

cs.LG↗

arXiv2024

NatLogAttack: A Framework for Attacking Natural Language Inference Models with Natural Logic

Reasoning has been a central topic in artificial intelligence from the beginning. The recent progress made on distributed representation and neural networks continues to improve the state-of-the-art performance of natural language inference. However, it remains an open question whether the models perform real reasoning to reach their conclusions or rely on spurious correlations. Adversarial attacks have proven to be an important tool to help evaluate the Achilles' heel of the victim models. In this study, we explore the fundamental problem of developing attack models based on logic formalism. We propose NatLogAttack to perform systematic attacks centring around natural logic, a classical logic formalism that is traceable back to Aristotle's syllogism and has been closely developed for natural language inference. The proposed framework renders both label-preserving and label-flipping attacks. We show that compared to the existing attack models, NatLogAttack generates better adversarial examples with fewer visits to the victim models. The victim models are found to be more vulnerable under the label-flipping setting. NatLogAttack provides a tool to probe the existing and future NLI models' capacity from a key viewpoint and we hope more logic-based attacks will be further explored for understanding the desired property of reasoning.

cs.CL↗

arXiv2024

Misinformation with Legal Consequences (MisLC): A New Task Towards Harnessing Societal Harm of Misinformation

Misinformation, defined as false or inaccurate information, can result in significant societal harm when it is spread with malicious or even innocuous intent. The rapid online information exchange necessitates advanced detection mechanisms to mitigate misinformation-induced harm. Existing research, however, has predominantly focused on assessing veracity, overlooking the legal implications and social consequences of misinformation. In this work, we take a novel angle to consolidate the definition of misinformation detection using legal issues as a measurement of societal ramifications, aiming to bring interdisciplinary efforts to tackle misinformation and its consequence. We introduce a new task: Misinformation with Legal Consequence (MisLC), which leverages definitions from a wide range of legal domains covering 4 broader legal topics and 11 fine-grained legal issues, including hate speech, election laws, and privacy regulations. For this task, we advocate a two-step dataset curation approach that utilizes crowd-sourced checkworthiness and expert evaluations of misinformation. We provide insights about the MisLC task through empirical evidence, from the problem definition to experiments and expert involvement. While the latest large language models and retrieval-augmented generation are effective baselines for the task, we find they are still far from replicating expert performance.

cs.CL↗

arXiv2024

Fine-Tuning Language Models with Differential Privacy through Adaptive Noise Allocation

Language models are capable of memorizing detailed patterns and information, leading to a double-edged effect: they achieve impressive modeling performance on downstream tasks with the stored knowledge but also raise significant privacy concerns. Traditional differential privacy based training approaches offer robust safeguards by employing a uniform noise distribution across all parameters. However, this overlooks the distinct sensitivities and contributions of individual parameters in privacy protection and often results in suboptimal models. To address these limitations, we propose ANADP, a novel algorithm that adaptively allocates additive noise based on the importance of model parameters. We demonstrate that ANADP narrows the performance gap between regular fine-tuning and traditional DP fine-tuning on a series of datasets while maintaining the required privacy constraints.

cs.AI↗

arXiv2024

Code Prompting Elicits Conditional Reasoning Abilities in Text+Code LLMs

Reasoning is a fundamental component of language understanding. Recent prompting techniques, such as chain of thought, have consistently improved LLMs' performance on various reasoning tasks. Nevertheless, there is still little understanding of what triggers reasoning abilities in LLMs in the inference stage. In this paper, we introduce code prompting, a chain of prompts that transforms a natural language problem into code and directly prompts the LLM using the generated code without resorting to external code execution. We hypothesize that code prompts can elicit certain reasoning capabilities of LLMs trained on text and code and utilize the proposed method to improve conditional reasoning, the ability to infer different conclusions depending on the fulfillment of certain conditions. We find that code prompting exhibits a high-performance boost for multiple LLMs (up to 22.52 percentage points on GPT 3.5, 7.75 on Mixtral, and 16.78 on Mistral) across multiple conditional reasoning datasets. We then conduct comprehensive experiments to understand how code prompts trigger reasoning abilities and which capabilities are elicited in the underlying models. Our analysis of GPT 3.5 reveals that the code formatting of the input problem is essential for performance improvement. Furthermore, code prompts improve sample efficiency of in-context learning and facilitate state tracking of variables or entities.

cs.CL↗

arXiv2024

Experimenting with Legal AI Solutions: The Case of Question-Answering for Access to Justice

Generative AI models, such as the GPT and Llama series, have significant potential to assist laypeople in answering legal questions. However, little prior work focuses on the data sourcing, inference, and evaluation of these models in the context of laypersons. To this end, we propose a human-centric legal NLP pipeline, covering data sourcing, inference, and evaluation. We introduce and release a dataset, LegalQA, with real and specific legal questions spanning from employment law to criminal law, corresponding answers written by legal experts, and citations for each answer. We develop an automatic evaluation protocol for this dataset, then show that retrieval-augmented generation from only 850 citations in the train set can match or outperform internet-wide retrieval, despite containing 9 orders of magnitude less data. Finally, we propose future directions for open-sourced efforts, which fall behind closed-sourced models.

cs.CL↗

arXiv2024

HGOT: Hierarchical Graph of Thoughts for Retrieval-Augmented In-Context Learning in Factuality Evaluation

With the widespread adoption of large language models (LLMs) in numerous applications, the challenge of factuality and the propensity for hallucinations has emerged as a significant concern. To address this issue, particularly in retrieval-augmented in-context learning, we introduce the hierarchical graph of thoughts (HGOT), a structured, multi-layered graph approach designed to enhance the retrieval of pertinent passages during in-context learning. The framework utilizes the emergent planning capabilities of LLMs, employing the divide-and-conquer strategy to break down complex queries into manageable sub-queries. It refines self-consistency majority voting for answer selection, which incorporates the recently proposed citation recall and precision metrics to assess the quality of thoughts, linking an answer's credibility intrinsically to the thought's quality. This methodology introduces a weighted system in majority voting, prioritizing answers based on the citation quality of their thoughts. Additionally, we propose a scoring mechanism for evaluating retrieved passages, considering factors such as citation frequency and quality, self-consistency confidence, and the retrieval module's ranking. Experiments indicate that HGOT excels as a versatile approach, outperforming competing models in FEVER by up to $7\%$ and matching leading models such as Retrieve-then-Read in Open-SQuAD, and DSP in HotPotQA, demonstrating its efficacy in enhancing LLMs' factuality.

cs.AI↗

arXiv2024

Mitigating Social Biases in Language Models through Unlearning

Mitigating bias in language models (LMs) has become a critical problem due to the widespread deployment of LMs. Numerous approaches revolve around data pre-processing and fine-tuning of language models, tasks that can be both time-consuming and computationally demanding. Consequently, there is a growing interest in machine unlearning techniques given their capacity to induce the forgetting of undesired behaviors of the existing pre-trained or fine-tuned models with lower computational cost. In this work, we explore two unlearning methods, (1) Partitioned Contrastive Gradient Unlearning (PCGU) applied on decoder models and (2) Negation via Task Vector, to reduce social biases in state-of-the-art and open-source LMs such as LLaMA-2 and OPT. We also implement distributed PCGU for large models. It is empirically shown, through quantitative and qualitative analyses, that negation via Task Vector method outperforms PCGU in debiasing with minimum deterioration in performance and perplexity of the models. On LLaMA-27B, negation via Task Vector reduces the bias score by 11.8%

cs.CL↗

arXiv2024

DARA: Decomposition-Alignment-Reasoning Autonomous Language Agent for Question Answering over Knowledge Graphs

Answering Questions over Knowledge Graphs (KGQA) is key to well-functioning autonomous language agents in various real-life applications. To improve the neural-symbolic reasoning capabilities of language agents powered by Large Language Models (LLMs) in KGQA, we propose the DecompositionAlignment-Reasoning Agent (DARA) framework. DARA effectively parses questions into formal queries through a dual mechanism: high-level iterative task decomposition and low-level task grounding. Importantly, DARA can be efficiently trained with a small number of high-quality reasoning trajectories. Our experimental results demonstrate that DARA fine-tuned on LLMs (e.g. Llama-2-7B, Mistral) outperforms both in-context learning-based agents with GPT-4 and alternative fine-tuned agents, across different benchmarks in zero-shot evaluation, making such models more accessible for real-life applications. We also show that DARA attains performance comparable to state-of-the-art enumerating-and-ranking-based methods for KGQA.

cs.CL↗

arXiv2024

SpaRC and SpaRP: Spatial Reasoning Characterization and Path Generation for Understanding Spatial Reasoning Capability of Large Language Models

Spatial reasoning is a crucial component of both biological and artificial intelligence. In this work, we present a comprehensive study of the capability of current state-of-the-art large language models (LLMs) on spatial reasoning. To support our study, we created and contribute a novel Spatial Reasoning Characterization (SpaRC) framework and Spatial Reasoning Paths (SpaRP) datasets, to enable an in-depth understanding of the spatial relations and compositions as well as the usefulness of spatial reasoning chains. We found that all the state-of-the-art LLMs do not perform well on the datasets -- their performances are consistently low across different setups. The spatial reasoning capability improves substantially as model sizes scale up. Finetuning both large language models (e.g., Llama-2-70B) and smaller ones (e.g., Llama-2-13B) can significantly improve their F1-scores by 7--32 absolute points. We also found that the top proprietary LLMs still significantly outperform their open-source counterparts in topological spatial understanding and reasoning.

cs.CL↗