arXiv subjects

Zhixuan Chu

Publications and source records attributed to Zhixuan Chu.

At least 55 records · Page 3Linked to original sources

arXiv2025

Perplexity-Aware Data Scaling Law: Perplexity Landscapes Predict Performance for Continual Pre-training

Continual Pre-training (CPT) serves as a fundamental approach for adapting foundation models to domain-specific applications. Scaling laws for pre-training define a power-law relationship between dataset size and the test loss of an LLM. However, the marginal gains from simply increasing data for CPT diminish rapidly, yielding suboptimal data utilization and inefficient training. To address this challenge, we propose a novel perplexity-aware data scaling law to establish a predictive relationship between the perplexity landscape of domain-specific data and the test loss. Our approach leverages the perplexity derived from the pre-trained model on domain data as a proxy for estimating the knowledge gap, effectively quantifying the informational perplexity landscape of candidate training samples. By fitting this scaling law across diverse perplexity regimes, we enable adaptive selection of high-utility data subsets, prioritizing content that maximizes knowledge absorption while minimizing redundancy and noise. Extensive experiments demonstrate that our method consistently identifies near-optimal training subsets and achieves superior performance on both medical and general-domain benchmarks.

cs.LG↗

arXiv2025

Interpretable LLM Guardrails via Sparse Representation Steering

Large language models (LLMs) exhibit impressive capabilities in generation tasks but are prone to producing harmful, misleading, or biased content, posing significant ethical and safety concerns. To mitigate such risks, representation engineering, which steer model behavior toward desired attributes by injecting carefully designed steering vectors into LLM's representations at inference time, has emerged as a promising alternative to fine-tuning approaches. However, due to the semantically entangled nature of LLM's representation, existing representation engineering methods still suffer from several limitations: limited fine-grained controllability, content quality degradation, and conflict in multi-attribute control. To overcome these challenges, we propose Sparse Representation Steering (SRS), a novel framework that achieves fine-grained and interpretable control over LLM behavior by first disentangling internal activations into a sparse, semantically meaningful representation space, and then selectively steering relevant dimensions. Specifically, SRS leverages a pretrained Sparse Autoencoder (SAE) to transform dense, entangled activation patterns into a sparse monosemantic feature space. To identify relevant features, SRS contrasts sparse activations from positive and negative prompt pairs and measures their bidirectional KL divergence to locate dimensions most associated with the target attribute. We conduct comprehensive experiments on Gemma-2 series model across three alignment dimensions, i.e., safety, fairness, and truthfulness, to evaluate the effectiveness of SRS. Results show that SRS consistently outperforms existing steering methods, which achieves significantly improved controllability across both single and multiple attribute settings, while preserving high linguistic quality and general ability.

cs.CR↗

arXiv2025

Understanding and Mitigating Overrefusal in LLMs from an Unveiling Perspective of Safety Decision Boundary

Large language models (LLMs) have demonstrated remarkable capabilities across a wide range of tasks, yet they often refuse to answer legitimate queries--a phenomenon known as overrefusal. Overrefusal typically stems from over-conservative safety alignment, causing models to treat many reasonable prompts as potentially risky. To systematically understand this issue, we probe and leverage the models' safety decision boundaries to analyze and mitigate overrefusal. Our findings reveal that overrefusal is closely tied to misalignment at these boundary regions, where models struggle to distinguish subtle differences between benign and harmful content. Building on these insights, we present RASS, an automated framework for prompt generation and selection that strategically targets overrefusal prompts near the safety boundary. By harnessing steering vectors in the representation space, RASS efficiently identifies and curates boundary-aligned prompts, enabling more effective and targeted mitigation of overrefusal. This approach not only provides a more precise and interpretable view of model safety decisions but also seamlessly extends to multilingual scenarios. We have explored the safety decision boundaries of various LLMs and construct the MORBench evaluation set to facilitate robust assessment of model safety and helpfulness across multiple languages. Code and datasets are available at https://github.com/Master-PLC/RASS.

cs.AI↗

arXiv2025

Probe before You Talk: Towards Black-box Defense against Backdoor Unalignment for Large Language Models

Backdoor unalignment attacks against Large Language Models (LLMs) enable the stealthy compromise of safety alignment using a hidden trigger while evading normal safety auditing. These attacks pose significant threats to the applications of LLMs in the real-world Large Language Model as a Service (LLMaaS) setting, where the deployed model is a fully black-box system that can only interact through text. Furthermore, the sample-dependent nature of the attack target exacerbates the threat. Instead of outputting a fixed label, the backdoored LLM follows the semantics of any malicious command with the hidden trigger, significantly expanding the target space. In this paper, we introduce BEAT, a black-box defense that detects triggered samples during inference to deactivate the backdoor. It is motivated by an intriguing observation (dubbed the probe concatenate effect), where concatenated triggered samples significantly reduce the refusal rate of the backdoored LLM towards a malicious probe, while non-triggered samples have little effect. Specifically, BEAT identifies whether an input is triggered by measuring the degree of distortion in the output distribution of the probe before and after concatenation with the input. Our method addresses the challenges of sample-dependent targets from an opposite perspective. It captures the impact of the trigger on the refusal signal (which is sample-independent) instead of sample-specific successful attack behaviors. It overcomes black-box access limitations by using multiple sampling to approximate the output distribution. Extensive experiments are conducted on various backdoor attacks and LLMs (including the closed-source GPT-3.5-turbo), verifying the effectiveness and efficiency of our defense. Besides, we also preliminarily verify that BEAT can effectively defend against popular jailbreak attacks, as they can be regarded as 'natural backdoors'.

cs.CR↗

arXiv2025

Towards Real-world Debiasing: Rethinking Evaluation, Challenge, and Solution

Spurious correlations in training data significantly hinder the generalization capability of machine learning models when faced with distribution shifts, leading to the proposition of numberous debiasing methods. However, it remains to be asked: \textit{Do existing benchmarks for debiasing really represent biases in the real world?} Recent works attempt to address such concerns by sampling from real-world data (instead of synthesizing) according to some predefined biased distributions to ensure the realism of individual samples. However, the realism of the biased distribution is more critical yet challenging and underexplored due to the complexity of real-world bias distributions. To tackle the problem, we propose a fine-grained framework for analyzing biased distributions, based on which we empirically and theoretically identify key characteristics of biased distributions in the real world that are poorly represented by existing benchmarks. Towards applicable debiasing in the real world, we further introduce two novel real-world-inspired biases to bridge this gap and build a systematic evaluation framework for real-world debiasing, RDBench\footnote{RDBench: Code to be released. Preliminary version in supplementary material for anonimized review.}. Furthermore, focusing on the practical setting of debiasing w/o bias label, we find real-world biases pose a novel \textit{Sparse bias capturing} challenge to the existing paradigm. We propose a simple yet effective approach named Debias in Destruction (DiD), to address the challenge, whose effectiveness is validated with extensive experiments on 8 datasets of various biased distributions.

cs.LG↗

arXiv2025

TimeMixer++: A General Time Series Pattern Machine for Universal Predictive Analysis

Time series analysis plays a critical role in numerous applications, supporting tasks such as forecasting, classification, anomaly detection, and imputation. In this work, we present the time series pattern machine (TSPM), a model designed to excel in a broad range of time series tasks through powerful representation and pattern extraction capabilities. Traditional time series models often struggle to capture universal patterns, limiting their effectiveness across diverse tasks. To address this, we define multiple scales in the time domain and various resolutions in the frequency domain, employing various mixing strategies to extract intricate, task-adaptive time series patterns. Specifically, we introduce a general-purpose TSPM that processes multi-scale time series using (1) multi-resolution time imaging (MRTI), (2) time image decomposition (TID), (3) multi-scale mixing (MCM), and (4) multi-resolution mixing (MRM) to extract comprehensive temporal patterns. MRTI transforms multi-scale time series into multi-resolution time images, capturing patterns across both temporal and frequency domains. TID leverages dual-axis attention to extract seasonal and trend patterns, while MCM hierarchically aggregates these patterns across scales. MRM adaptively integrates all representations across resolutions. This method achieves state-of-the-art performance across 8 time series analytical tasks, consistently surpassing both general-purpose and task-specific models. Our work marks a promising step toward the next generation of TSPMs, paving the way for further advancements in time series analysis.

cs.LG↗

arXiv2025

FAMMA: A Benchmark for Financial Domain Multilingual Multimodal Question Answering

In this paper, we introduce FAMMA, an open-source benchmark for \underline{f}in\underline{a}ncial \underline{m}ultilingual \underline{m}ultimodal question \underline{a}nswering (QA). Our benchmark aims to evaluate the abilities of large language models (LLMs) in answering complex reasoning questions that require advanced financial knowledge. The benchmark has two versions: FAMMA-Basic consists of 1,945 questions extracted from university textbooks and exams, along with human-annotated answers and rationales; FAMMA-LivePro consists of 103 novel questions created by human domain experts, with answers and rationales held out from the public for a contamination-free evaluation. These questions cover advanced knowledge of 8 major subfields in finance (e.g., corporate finance, derivatives, and portfolio management). Some are in Chinese or French, while a majority of them are in English. Each question has some non-text data such as charts, diagrams, or tables. Our experiments reveal that FAMMA poses a significant challenge on LLMs, including reasoning models such as GPT-o1 and DeepSeek-R1. Additionally, we curated 1,270 reasoning trajectories of DeepSeek-R1 on the FAMMA-Basic data, and fine-tuned a series of open-source Qwen models using this reasoning data. We found that training a model on these reasoning trajectories can significantly improve its performance on FAMMA-LivePro. We released our leaderboard, data, code, and trained models at https://famma-bench.github.io/famma/.

cs.CL↗

arXiv2025

JailbreakLens: Interpreting Jailbreak Mechanism in the Lens of Representation and Circuit

Despite the outstanding performance of Large language Models (LLMs) in diverse tasks, they are vulnerable to jailbreak attacks, wherein adversarial prompts are crafted to bypass their security mechanisms and elicit unexpected responses. Although jailbreak attacks are prevalent, the understanding of their underlying mechanisms remains limited. Recent studies have explained typical jailbreaking behavior (e.g., the degree to which the model refuses to respond) of LLMs by analyzing representation shifts in their latent space caused by jailbreak prompts or identifying key neurons that contribute to the success of jailbreak attacks. However, these studies neither explore diverse jailbreak patterns nor provide a fine-grained explanation from the failure of circuit to the changes of representational, leaving significant gaps in uncovering the jailbreak mechanism. In this paper, we propose JailbreakLens, an interpretation framework that analyzes jailbreak mechanisms from both representation (which reveals how jailbreaks alter the model's harmfulness perception) and circuit perspectives~(which uncovers the causes of these deceptions by identifying key circuits contributing to the vulnerability), tracking their evolution throughout the entire response generation process. We then conduct an in-depth evaluation of jailbreak behavior on five mainstream LLMs under seven jailbreak strategies. Our evaluation reveals that jailbreak prompts amplify components that reinforce affirmative responses while suppressing those that produce refusal. This manipulation shifts model representations toward safe clusters to deceive the LLM, leading it to provide detailed responses instead of refusals. Notably, we find a strong and consistent correlation between representation deception and activation shift of key circuits across diverse jailbreak methods and multiple LLMs.

cs.CR↗

arXiv2025

Mitigating Social Bias in Large Language Models: A Multi-Objective Approach within a Multi-Agent Framework

Natural language processing (NLP) has seen remarkable advancements with the development of large language models (LLMs). Despite these advancements, LLMs often produce socially biased outputs. Recent studies have mainly addressed this problem by prompting LLMs to behave ethically, but this approach results in unacceptable performance degradation. In this paper, we propose a multi-objective approach within a multi-agent framework (MOMA) to mitigate social bias in LLMs without significantly compromising their performance. The key idea of MOMA involves deploying multiple agents to perform causal interventions on bias-related contents of the input questions, breaking the shortcut connection between these contents and the corresponding answers. Unlike traditional debiasing techniques leading to performance degradation, MOMA substantially reduces bias while maintaining accuracy in downstream tasks. Our experiments conducted on two datasets and two models demonstrate that MOMA reduces bias scores by up to 87.7%, with only a marginal performance degradation of up to 6.8% in the BBQ dataset. Additionally, it significantly enhances the multi-objective metric icat in the StereoSet dataset by up to 58.1%. Code will be made available at https://github.com/Cortantse/MOMA.

cs.CL↗

arXiv2024

A Survey on Medical Large Language Models: Technology, Application, Trustworthiness, and Future Directions

With the advent of Large Language Models (LLMs), medical artificial intelligence (AI) has experienced substantial technological progress and paradigm shifts, highlighting the potential of LLMs to streamline healthcare delivery and improve patient outcomes. Considering this rapid technical progress, in this survey, we trace the recent advances of Medical Large Language Models (Med-LLMs), including the background, key findings, and mainstream techniques, especially for the evolution from general-purpose models to medical-specialized applications. Firstly, we delve into the foundational technology of Med-LLMs, indicating how general models can be progressively adapted and refined for the complicated medical tasks. Secondly, the wide-ranging applications of Med-LLMs are investigated across various healthcare domains, as well as an up-to-date review of existing Med-LLMs. The transformative impact of these models on daily medical practice is evident through their ability to assist clinicians, educators, and patients. Recognizing the importance of responsible innovation, we discuss the challenges associated with ensuring fairness, accountability, privacy, and robustness. Ethical considerations, rigorous evaluation methodologies, and the establishment of regulatory frameworks are crucial for building trustworthiness in the real-world system. We emphasize the need for ongoing scrutiny and development to maintain high standards of safety and reliability. Finally, we anticipate possible future trajectories for Med-LLMs, identifying key avenues for prudent expansion. By consolidating these insights, our review aims to provide professionals and researchers with a thorough understanding of the strengths and limitations of Med-LLMs, fostering a balanced and ethical approach to their integration into the healthcare ecosystem.

cs.CL↗

arXiv2024

Explainable Behavior Cloning: Teaching Large Language Model Agents through Learning by Demonstration

Autonomous mobile app interaction has become increasingly important with growing complexity of mobile applications. Developing intelligent agents that can effectively navigate and interact with mobile apps remains a significant challenge. In this paper, we propose an Explainable Behavior Cloning LLM Agent (EBC-LLMAgent), a novel approach that combines large language models (LLMs) with behavior cloning by learning demonstrations to create intelligent and explainable agents for autonomous mobile app interaction. EBC-LLMAgent consists of three core modules: Demonstration Encoding, Code Generation, and UI Mapping, which work synergistically to capture user demonstrations, generate executable codes, and establish accurate correspondence between code and UI elements. We introduce the Behavior Cloning Chain Fusion technique to enhance the generalization capabilities of the agent. Extensive experiments on five popular mobile applications from diverse domains demonstrate the superior performance of EBC-LLMAgent, achieving high success rates in task completion, efficient generalization to unseen scenarios, and the generation of meaningful explanations.

cs.CL↗

arXiv2024

A Causal Explainable Guardrails for Large Language Models

Large Language Models (LLMs) have shown impressive performance in natural language tasks, but their outputs can exhibit undesirable attributes or biases. Existing methods for steering LLMs toward desired attributes often assume unbiased representations and rely solely on steering prompts. However, the representations learned from pre-training can introduce semantic biases that influence the steering process, leading to suboptimal results. We propose LLMGuardrail, a novel framework that incorporates causal analysis and adversarial learning to obtain unbiased steering representations in LLMs. LLMGuardrail systematically identifies and blocks the confounding effects of biases, enabling the extraction of unbiased steering representations. Additionally, it includes an explainable component that provides insights into the alignment between the generated output and the desired direction. Experiments demonstrate LLMGuardrail's effectiveness in steering LLMs toward desired attributes while mitigating biases. Our work contributes to the development of safe and reliable LLMs that align with desired attributes.

cs.CL↗

arXiv2024

Multiscale Representation Enhanced Temporal Flow Fusion Model for Long-Term Workload Forecasting

Accurate workload forecasting is critical for efficient resource management in cloud computing systems, enabling effective scheduling and autoscaling. Despite recent advances with transformer-based forecasting models, challenges remain due to the non-stationary, nonlinear characteristics of workload time series and the long-term dependencies. In particular, inconsistent performance between long-term history and near-term forecasts hinders long-range predictions. This paper proposes a novel framework leveraging self-supervised multiscale representation learning to capture both long-term and near-term workload patterns. The long-term history is encoded through multiscale representations while the near-term observations are modeled via temporal flow fusion. These representations of different scales are fused using an attention mechanism and characterized with normalizing flows to handle non-Gaussian/non-linear distributions of time series. Extensive experiments on 9 benchmarks demonstrate superiority over existing methods.

cs.LG↗

arXiv2024

Causal Interventional Prediction System for Robust and Explainable Effect Forecasting

Although the widespread use of AI systems in today's world is growing, many current AI systems are found vulnerable due to hidden bias and missing information, especially in the most commonly used forecasting system. In this work, we explore the robustness and explainability of AI-based forecasting systems. We provide an in-depth analysis of the underlying causality involved in the effect prediction task and further establish a causal graph based on treatment, adjustment variable, confounder, and outcome. Correspondingly, we design a causal interventional prediction system (CIPS) based on a variational autoencoder and fully conditional specification of multiple imputations. Extensive results demonstrate the superiority of our system over state-of-the-art methods and show remarkable versatility and extensibility in practice.

cs.LG↗

arXiv2024

Prompt-Consistency Image Generation (PCIG): A Unified Framework Integrating LLMs, Knowledge Graphs, and Controllable Diffusion Models

The rapid advancement of Text-to-Image(T2I) generative models has enabled the synthesis of high-quality images guided by textual descriptions. Despite this significant progress, these models are often susceptible in generating contents that contradict the input text, which poses a challenge to their reliability and practical deployment. To address this problem, we introduce a novel diffusion-based framework to significantly enhance the alignment of generated images with their corresponding descriptions, addressing the inconsistency between visual output and textual input. Our framework is built upon a comprehensive analysis of inconsistency phenomena, categorizing them based on their manifestation in the image. Leveraging a state-of-the-art large language module, we first extract objects and construct a knowledge graph to predict the locations of these objects in potentially generated images. We then integrate a state-of-the-art controllable image generation model with a visual text generation module to generate an image that is consistent with the original prompt, guided by the predicted object locations. Through extensive experiments on an advanced multimodal hallucination benchmark, we demonstrate the efficacy of our approach in accurately generating the images without the inconsistency with the original prompt. The code can be accessed via https://github.com/TruthAI-Lab/PCIG.

cs.CV↗

arXiv2024

DB-GPT-Hub: Towards Open Benchmarking Text-to-SQL Empowered by Large Language Models

Large language models (LLMs) becomes the dominant paradigm for the challenging task of text-to-SQL. LLM-empowered text-to-SQL methods are typically categorized into prompting-based and tuning approaches. Compared to prompting-based methods, benchmarking fine-tuned LLMs for text-to-SQL is important yet under-explored, partially attributed to the prohibitively high computational cost. In this paper, we present DB-GPT-Hub, an open benchmark suite for LLM-empowered text-to-SQL, which primarily focuses on tuning LLMs at large scales. The proposed benchmark consists of: 1. a standardized and comprehensive evaluation of text-to-SQL tasks by fine-tuning medium to large-sized open LLMs; 2. a modularized and easy-to-extend codebase with mainstream LLMs and experimental scenarios supported, which prioritizes fine-tuning methods but can be easily extended to prompt-based setting. Our work investigates the potential gains and the performance boundaries of tuning approaches, compared to prompting approaches and explores optimal solutions tailored to specific scenarios. We hope DB-GPT-Hub, along with these findings, enables further research and broad applications that would otherwise be difficult owing to the absence of a dedicated open benchmark. The project code has been released at https://github.com/eosphoros-ai/DB-GPT-Hub.

cs.DB↗

arXiv2024

Sora Detector: A Unified Hallucination Detection for Large Text-to-Video Models

The rapid advancement in text-to-video (T2V) generative models has enabled the synthesis of high-fidelity video content guided by textual descriptions. Despite this significant progress, these models are often susceptible to hallucination, generating contents that contradict the input text, which poses a challenge to their reliability and practical deployment. To address this critical issue, we introduce the SoraDetector, a novel unified framework designed to detect hallucinations across diverse large T2V models, including the cutting-edge Sora model. Our framework is built upon a comprehensive analysis of hallucination phenomena, categorizing them based on their manifestation in the video content. Leveraging the state-of-the-art keyframe extraction techniques and multimodal large language models, SoraDetector first evaluates the consistency between extracted video content summary and textual prompts, then constructs static and dynamic knowledge graphs (KGs) from frames to detect hallucination both in single frames and across frames. Sora Detector provides a robust and quantifiable measure of consistency, static and dynamic hallucination. In addition, we have developed the Sora Detector Agent to automate the hallucination detection process and generate a complete video quality report for each input video. Lastly, we present a novel meta-evaluation benchmark, T2VHaluBench, meticulously crafted to facilitate the evaluation of advancements in T2V hallucination detection. Through extensive experiments on videos generated by Sora and other large T2V models, we demonstrate the efficacy of our approach in accurately detecting hallucinations. The code and dataset can be accessed via GitHub.

cs.LG↗

arXiv2024

LLM-Guided Multi-View Hypergraph Learning for Human-Centric Explainable Recommendation

As personalized recommendation systems become vital in the age of information overload, traditional methods relying solely on historical user interactions often fail to fully capture the multifaceted nature of human interests. To enable more human-centric modeling of user preferences, this work proposes a novel explainable recommendation framework, i.e., LLMHG, synergizing the reasoning capabilities of large language models (LLMs) and the structural advantages of hypergraph neural networks. By effectively profiling and interpreting the nuances of individual user interests, our framework pioneers enhancements to recommendation systems with increased explainability. We validate that explicitly accounting for the intricacies of human preferences allows our human-centric and explainable LLMHG approach to consistently outperform conventional models across diverse real-world datasets. The proposed plug-and-play enhancement framework delivers immediate gains in recommendation performance while offering a pathway to apply advanced LLMs for better capturing the complexity of human interests across machine learning applications.

cs.IR↗