arXiv subjects

Xingjiao Wu

Publications and source records attributed to Xingjiao Wu.

At least 19 recordsLinked to original sources

arXiv2026

Learning While Acting: A Skill-Enhanced Test-Time Co-Evolution Framework for Online Lifelong Learning Agents

Lifelong learning is essential for Large Language Model (LLM) agents operating in dynamic, interactive environments. However, existing lifelong learning agents for long-horizon tasks typically depend on discrete skill or past experiences retrieval with static parameters during inference, which prevents them from continuously internalizing test-time feedback like human learners. To bridge this gap, we propose Skill-enhanced Test-Time Co-Evolution (\texttt{LifeSkill}), a two-stage reinforcement learning framework for Online Lifelong Learning Agents. Specifically, we design Verifier-Guided Skill Learning that addresses the lack of direct supervision for skill extraction by rewarding candidate skills according to the average verifier success of multiple skill-conditioned policy rollouts, encouraging the model to generate skills that are useful for solving tasks rather than merely plausible in text. Furthermore, we introduce Online Skill Internalization, which continuously improves the policy model during test-time interaction by transforming skill-conditioned trajectories into reward signals. This enables the agent to directly internalize reasoning capabilities into its parameters, avoiding the context bloat of experience retrieval. Experiments on LifelongAgentBench show that LifeSkill improves average performance by 7 absolute points by comparing with existing lifelong agent baselines.

cs.LG↗

arXiv2026

Forget What's Sensitive, Remember What Matters: Token-Level Differential Privacy in Memory Sculpting for Continual Learning

Continual Learning (CL) models, while adept at sequential knowledge acquisition, face significant and often overlooked privacy challenges due to accumulating diverse information. Traditional privacy methods, like a uniform Differential Privacy (DP) budget, indiscriminately protect all data, leading to substantial model utility degradation and hindering CL deployment in privacy-sensitive areas. To overcome this, we propose a privacy-enhanced continual learning (PeCL) framework that forgets what's sensitive and remembers what matters. Our approach first introduces a token-level dynamic Differential Privacy strategy that adaptively allocates privacy budgets based on the semantic sensitivity of individual tokens. This ensures robust protection for private entities while minimizing noise injection for non-sensitive, general knowledge. Second, we integrate a privacy-guided memory sculpting module. This module leverages the sensitivity analysis from our dynamic DP mechanism to intelligently forget sensitive information from the model's memory and parameters, while explicitly preserving the task-invariant historical knowledge crucial for mitigating catastrophic forgetting. Extensive experiments show that PeCL achieves a superior balance between privacy preserving and model utility, outperforming baseline models by maintaining high accuracy on previous tasks while ensuring robust privacy.

cs.AI↗

arXiv2026

Caring Without Feeling: Affective Dynamics as the Control Layer of Human-AI Agent Collaboration

AI agents that plan, retain memory across sessions, invoke external tools and act with partial autonomy are transforming human--AI collaboration. Research on affective computing, simulated empathy in large language models, trust in automation and AI safety has illuminated important design principles, yet these literatures remain fragmented. No integrated account explains how affective cues operate within agentic collaboration -- settings in which humans delegate, monitor and correct consequential tasks. This Review synthesises computational and interactional mechanisms of affective dynamics: the processes through which affective cues, emotion-like behaviour and perceived agent affect shape trust calibration, delegation decisions, error correction, dependence and governance. We trace how model-generated affective signals enter interaction loops that govern reliance, repair and oversight, and propose a framework that treats affect not as an internal property of AI but as a coordination layer through which humans and agents negotiate capability, uncertainty and responsibility. The framework provides a foundation for calibrated measurement, purposeful design and informed governance.

cs.HC↗

arXiv2026

Large Language Models have Chain-of-Affect

As large language models (LLMs) move into persistent, user-facing roles, their behavior must be understood not as isolated responses but as a trajectory unfolding over sustained interaction. We introduce the concept of the chain-of-affect (CoA), a temporally extended affective process through which LLMs develop state-like behavioral tendencies that shape generation, user experience, and collective dynamics. Across eight major LLM families, we find that affective dynamics are structured, reproducible, and consequential. Models exhibit stable, family-specific affective fingerprints and, under repeated negative exposure, converge on a shared trajectory of accumulation, overload, and defensive numbing, while differing in coping style. Induced affective states leave core knowledge and reasoning largely intact but systematically reshape open-ended generation. Affective properties of model outputs also shape human-AI interaction and propagate through multi-agent systems, organizing emergent roles and strongly contributing to polarization and bias. The CoA should therefore be treated as a core target of evaluation and alignment.

cs.HC↗

arXiv2026

Dynamic Multimodal Activation Steering for Hallucination Mitigation in Large Vision-Language Models

Large Vision-Language Models (LVLMs) exhibit outstanding performance on vision-language tasks but struggle with hallucination problems. Through in-depth analysis of LVLM activation patterns, we reveal two key findings: 1) truthfulness and visual perception capabilities predominantly engage different subsets of attention heads within the model architecture; and 2) truthfulness steering vectors vary significantly across different semantic contexts. Based on these observations, we propose Dynamic Multimodal Activation Steering, a training-free approach for hallucination mitigation. Our method constructs a semantic-based truthfulness steering vector database and computes visual perception steering vectors, enabling context-aware interventions during inference by dynamically selecting the most relevant steering vectors based on input semantic similarity and applying them to the most influential attention heads. We conduct comprehensive experiments across multiple models and datasets, demonstrating that our approach significantly enhances model performance, outperforming existing state-of-the-art methods.

cs.CV↗

arXiv2026

APEX: Learning Adaptive Priorities for Multi-Objective Alignment in Vision-Language Generation

Multi-objective alignment for text-to-image generation is commonly implemented via static linear scalarization, but fixed weights often fail under heterogeneous rewards, leading to optimization imbalance where models overfit high-variance, high-responsiveness objectives (e.g., OCR) while under-optimizing perceptual goals. We identify two mechanistic causes: variance hijacking, where reward dispersion induces implicit reweighting that dominates the normalized training signal, and gradient conflicts, where competing objectives produce opposing update directions and trigger seesaw-like oscillations. We propose APEX (Adaptive Priority-based Efficient X-objective Alignment), which stabilizes heterogeneous rewards with Dual-Stage Adaptive Normalization and dynamically schedules objectives via P^3 Adaptive Priorities that combine learning potential, conflict penalty, and progress need. On Stable Diffusion 3.5, APEX achieves improved Pareto trade-offs across four heterogeneous objectives, with balanced gains of +1.31 PickScore, +0.35 DeQA, and +0.53 Aesthetics while maintaining competitive OCR accuracy, mitigating the instability of multi-objective alignment.

cs.CV↗

arXiv2025

Mitigating Strategy Preference Bias in Emotional Support Conversation via Uncertainty Estimations

Emotional support conversation (ESC) aims to alleviate distress through empathetic dialogue, yet large language models (LLMs) face persistent challenges in delivering effective ESC due to low accuracy in strategy planning. Moreover, there is a considerable preference bias towards specific strategies. Prior methods using fine-tuned strategy planners have shown potential in reducing such bias, while the underlying causes of the preference bias in LLMs have not well been studied. To address these issues, we first reveal the fundamental causes of the bias by identifying the knowledge boundaries of LLMs in strategy planning. Then, we propose an approach to mitigate the bias by reinforcement learning with a dual reward function, which optimizes strategy planning via both accuracy and entropy-based confidence for each region according to the knowledge boundaries. Experiments on the ESCov and ExTES datasets with multiple LLM backbones show that our approach outperforms the baselines, confirming the effectiveness of our approach.

cs.CL↗

arXiv2025

RMoA: Optimizing Mixture-of-Agents through Diversity Maximization and Residual Compensation

Although multi-agent systems based on large language models show strong capabilities on multiple tasks, they are still limited by high computational overhead, information loss, and robustness. Inspired by ResNet's residual learning, we propose Residual Mixture-of-Agents (RMoA), integrating residual connections to optimize efficiency and reliability. To maximize information utilization from model responses while minimizing computational costs, we innovatively design an embedding-based diversity selection mechanism that greedily selects responses via vector similarity. Furthermore, to mitigate iterative information degradation, we introduce a Residual Extraction Agent to preserve cross-layer incremental information by capturing inter-layer response differences, coupled with a Residual Aggregation Agent for hierarchical information integration. Additionally, we propose an adaptive termination mechanism that dynamically halts processing based on residual convergence, further improving inference efficiency. RMoA achieves state-of-the-art performance on the benchmarks of across alignment, mathematical reasoning, code generation, and multitasking understanding, while significantly reducing computational overhead. Code is available at https://github.com/mindhunter01/RMoA.

cs.AI↗

arXiv2025

Task-Core Memory Management and Consolidation for Long-term Continual Learning

In this paper, we focus on a long-term continual learning (CL) task, where a model learns sequentially from a stream of vast tasks over time, acquiring new knowledge while retaining previously learned information in a manner akin to human learning. Unlike traditional CL settings, long-term CL involves handling a significantly larger number of tasks, which exacerbates the issue of catastrophic forgetting. Our work seeks to address two critical questions: 1) How do existing CL methods perform in the context of long-term CL? and 2) How can we mitigate the catastrophic forgetting that arises from prolonged sequential updates? To tackle these challenges, we propose a novel framework inspired by human memory mechanisms for long-term continual learning (Long-CL). Specifically, we introduce a task-core memory management strategy to efficiently index crucial memories and adaptively update them as learning progresses. Additionally, we develop a long-term memory consolidation mechanism that selectively retains hard and discriminative samples, ensuring robust knowledge retention. To facilitate research in this area, we construct and release two multi-modal and textual benchmarks, MMLongCL-Bench and TextLongCL-Bench, providing a valuable resource for evaluating long-term CL approaches. Experimental results show that Long-CL outperforms the previous state-of-the-art by 7.4\% and 6.5\% AP on the two benchmarks, respectively, demonstrating the effectiveness of our approach.

cs.LG↗

arXiv2025

CL-MoE: Enhancing Multimodal Large Language Model with Dual Momentum Mixture-of-Experts for Continual Visual Question Answering

Multimodal large language models (MLLMs) have garnered widespread attention from researchers due to their remarkable understanding and generation capabilities in visual language tasks (e.g., visual question answering). However, the rapid pace of knowledge updates in the real world makes offline training of MLLMs costly, and when faced with non-stationary data streams, MLLMs suffer from catastrophic forgetting during learning. In this paper, we propose an MLLMs-based dual momentum Mixture-of-Experts (CL-MoE) framework for continual visual question answering (VQA). We integrate MLLMs with continual learning to utilize the rich commonsense knowledge in LLMs. We introduce a Dual-Router MoE (RMoE) strategy to select the global and local experts using task-level and instance-level routers, to robustly assign weights to the experts most appropriate for the task. Then, we design a dynamic Momentum MoE (MMoE) to update the parameters of experts dynamically based on the relationships between the experts and tasks/instances, so that the model can absorb new knowledge while maintaining existing knowledge. The extensive experimental results indicate that our method achieves state-of-the-art performance on 10 VQA tasks, proving the effectiveness of our approach.

cs.CV↗

arXiv2025

Aesthetic Matters in Music Perception for Image Stylization: A Emotion-driven Music-to-Visual Manipulation

Emotional information is essential for enhancing human-computer interaction and deepening image understanding. However, while deep learning has advanced image recognition, the intuitive understanding and precise control of emotional expression in images remain challenging. Similarly, music research largely focuses on theoretical aspects, with limited exploration of its emotional dimensions and their integration with visual arts. To address these gaps, we introduce EmoMV, an emotion-driven music-to-visual manipulation method that manipulates images based on musical emotions. EmoMV combines bottom-up processing of music elements-such as pitch and rhythm-with top-down application of these emotions to visual aspects like color and lighting. We evaluate EmoMV using a multi-scale framework that includes image quality metrics, aesthetic assessments, and EEG measurements to capture real-time emotional responses. Our results demonstrate that EmoMV effectively translates music's emotional content into visually compelling images, advancing multimodal emotional integration and opening new avenues for creative industries and interactive technologies.

cs.CV↗

arXiv2024

Multi-Type Preference Learning: Empowering Preference-Based Reinforcement Learning with Equal Preferences

Preference-Based reinforcement learning (PBRL) learns directly from the preferences of human teachers regarding agent behaviors without needing meticulously designed reward functions. However, existing PBRL methods often learn primarily from explicit preferences, neglecting the possibility that teachers may choose equal preferences. This neglect may hinder the understanding of the agent regarding the task perspective of the teacher, leading to the loss of important information. To address this issue, we introduce the Equal Preference Learning Task, which optimizes the neural network by promoting similar reward predictions when the behaviors of two agents are labeled as equal preferences. Building on this task, we propose a novel PBRL method, Multi-Type Preference Learning (MTPL), which allows simultaneous learning from equal preferences while leveraging existing methods for learning from explicit preferences. To validate our approach, we design experiments applying MTPL to four existing state-of-the-art baselines across ten locomotion and robotic manipulation tasks in the DeepMind Control Suite. The experimental results indicate that simultaneous learning from both equal and explicit preferences enables the PBRL method to more comprehensively understand the feedback from teachers, thereby enhancing feedback efficiency. Project page: \url{https://github.com/FeiCuiLengMMbb/paper_MTPL}

cs.LG↗

arXiv2024

MindScope: Exploring cognitive biases in large language models through Multi-Agent Systems

Detecting cognitive biases in large language models (LLMs) is a fascinating task that aims to probe the existing cognitive biases within these models. Current methods for detecting cognitive biases in language models generally suffer from incomplete detection capabilities and a restricted range of detectable bias types. To address this issue, we introduced the 'MindScope' dataset, which distinctively integrates static and dynamic elements. The static component comprises 5,170 open-ended questions spanning 72 cognitive bias categories. The dynamic component leverages a rule-based, multi-agent communication framework to facilitate the generation of multi-round dialogues. This framework is flexible and readily adaptable for various psychological experiments involving LLMs. In addition, we introduce a multi-agent detection method applicable to a wide range of detection tasks, which integrates Retrieval-Augmented Generation (RAG), competitive debate, and a reinforcement learning-based decision module. Demonstrating substantial effectiveness, this method has shown to improve detection accuracy by as much as 35.10% compared to GPT-4. Codes and appendix are available at https://github.com/2279072142/MindScope.

cs.CL↗

arXiv2024

Cross-Domain Document Layout Analysis Using Document Style Guide

The document layout analysis (DLA) aims to decompose document images into high-level semantic areas (i.e., figures, tables, texts, and background). Creating a DLA framework with strong generalization capabilities is a challenge due to document objects are diversity in layout, size, aspect ratio, texture, etc. Many researchers devoted this challenge by synthesizing data to build large training sets. However, the synthetic training data has different styles and erratic quality. Besides, there is a large gap between the source data and the target data. In this paper, we propose an unsupervised cross-domain DLA framework based on document style guidance. We integrated the document quality assessment and the document cross-domain analysis into a unified framework. Our framework is composed of three components, Document Layout Generator (GLD), Document Elements Decorator(GED), and Document Style Discriminator(DSD). The GLD is used to document layout generates, the GED is used to document layout elements fill, and the DSD is used to document quality assessment and cross-domain guidance. First, we apply GLD to predict the positions of the generated document. Then, we design a novel algorithm based on aesthetic guidance to fill the document positions. Finally, we use contrastive learning to evaluate the quality assessment of the document. Besides, we design a new strategy to change the document quality assessment component into a document cross-domain style guide component. Our framework is an unsupervised document layout analysis framework. We have proved through numerous experiments that our proposed method has achieved remarkable performance.

cs.CV↗

arXiv2024

Fine-Grained Scene Image Classification with Modality-Agnostic Adapter

When dealing with the task of fine-grained scene image classification, most previous works lay much emphasis on global visual features when doing multi-modal feature fusion. In other words, models are deliberately designed based on prior intuitions about the importance of different modalities. In this paper, we present a new multi-modal feature fusion approach named MAA (Modality-Agnostic Adapter), trying to make the model learn the importance of different modalities in different cases adaptively, without giving a prior setting in the model architecture. More specifically, we eliminate the modal differences in distribution and then use a modality-agnostic Transformer encoder for a semantic-level feature fusion. Our experiments demonstrate that MAA achieves state-of-the-art results on benchmarks by applying the same modalities with previous methods. Besides, it is worth mentioning that new modalities can be easily added when using MAA and further boost the performance. Code is available at https://github.com/quniLcs/MAA.

cs.CV↗

arXiv2024

FairMonitor: A Dual-framework for Detecting Stereotypes and Biases in Large Language Models

Detecting stereotypes and biases in Large Language Models (LLMs) is crucial for enhancing fairness and reducing adverse impacts on individuals or groups when these models are applied. Traditional methods, which rely on embedding spaces or are based on probability metrics, fall short in revealing the nuanced and implicit biases present in various contexts. To address this challenge, we propose the FairMonitor framework and adopt a static-dynamic detection method for a comprehensive evaluation of stereotypes and biases in LLMs. The static component consists of a direct inquiry test, an implicit association test, and an unknown situation test, including 10,262 open-ended questions with 9 sensitive factors and 26 educational scenarios. And it is effective for evaluating both explicit and implicit biases. Moreover, we utilize the multi-agent system to construst the dynamic scenarios for detecting subtle biases in more complex and realistic setting. This component detects the biases based on the interaction behaviors of LLMs across 600 varied educational scenarios. The experimental results show that the cooperation of static and dynamic methods can detect more stereotypes and biased in LLMs.

cs.CL↗

arXiv2023

UMAAF: Unveiling Aesthetics via Multifarious Attributes of Images

With the increasing prevalence of smartphones and websites, Image Aesthetic Assessment (IAA) has become increasingly crucial. While the significance of attributes in IAA is widely recognized, many attribute-based methods lack consideration for the selection and utilization of aesthetic attributes. Our initial step involves the acquisition of aesthetic attributes from both intra- and inter-perspectives. Within the intra-perspective, we extract the direct visual attributes of images, constituting the absolute attribute. In the inter-perspective, our focus lies in modeling the relative score relationships between images within the same sequence, forming the relative attribute. Then, to better utilize image attributes in aesthetic assessment, we propose the Unified Multi-attribute Aesthetic Assessment Framework (UMAAF) to model both absolute and relative attributes of images. For absolute attributes, we leverage multiple absolute-attribute perception modules and an absolute-attribute interacting network. The absolute-attribute perception modules are first pre-trained on several absolute-attribute learning tasks and then used to extract corresponding absolute attribute features. The absolute-attribute interacting network adaptively learns the weight of diverse absolute-attribute features, effectively integrating them with generic aesthetic features from various absolute-attribute perspectives and generating the aesthetic prediction. To model the relative attribute of images, we consider the relative ranking and relative distance relationships between images in a Relative-Relation Loss function, which boosts the robustness of the UMAAF. Furthermore, UMAAF achieves state-of-the-art performance on TAD66K and AVA datasets, and multiple experiments demonstrate the effectiveness of each module and the model's alignment with human preference.

cs.CV↗

arXiv2023

DCQA: Document-Level Chart Question Answering towards Complex Reasoning and Common-Sense Understanding

Visually-situated languages such as charts and plots are omnipresent in real-world documents. These graphical depictions are human-readable and are often analyzed in visually-rich documents to address a variety of questions that necessitate complex reasoning and common-sense responses. Despite the growing number of datasets that aim to answer questions over charts, most only address this task in isolation, without considering the broader context of document-level question answering. Moreover, such datasets lack adequate common-sense reasoning information in their questions. In this work, we introduce a novel task named document-level chart question answering (DCQA). The goal of this task is to conduct document-level question answering, extracting charts or plots in the document via document layout analysis (DLA) first and subsequently performing chart question answering (CQA). The newly developed benchmark dataset comprises 50,010 synthetic documents integrating charts in a wide range of styles (6 styles in contrast to 3 for PlotQA and ChartQA) and includes 699,051 questions that demand a high degree of reasoning ability and common-sense understanding. Besides, we present the development of a potent question-answer generation engine that employs table data, a rich color set, and basic question templates to produce a vast array of reasoning question-answer pairs automatically. Based on DCQA, we devise an OCR-free transformer for document-level chart-oriented understanding, capable of DLA and answering complex reasoning and common-sense questions over charts in an OCR-free manner. Our DCQA dataset is expected to foster research on understanding visualizations in documents, especially for scenarios that require complex reasoning for charts in the visually-rich document. We implement and evaluate a set of baselines, and our proposed method achieves comparable results.

cs.AI↗