arXiv subjects

Zihan Wang

Publications and source records attributed to Zihan Wang.

At least 199 records · Page 11Linked to original sources

arXiv2025

Video Finetuning Improves Reasoning Between Frames

Multimodal large language models (LLMs) have made rapid progress in visual understanding, yet their extension from images to videos often reduces to a naive concatenation of frame tokens. In this work, we investigate what video finetuning brings to multimodal LLMs. We propose Visual Chain-of-Thought (vCoT), an explicit reasoning process that generates transitional event descriptions between consecutive frames. Using vCoT, we systematically compare image-only LVLMs with their video-finetuned counterparts, both with and without access to these transitional cues. Our experiments show that vCoT significantly improves the performance of image-only models on long-form video question answering, while yielding only marginal gains for video-finetuned models. This suggests that the latter already capture frame-to-frame transitions implicitly. Moreover, we find that video models transfer this temporal reasoning ability to purely static settings, outperforming image models' baselines on relational visual reasoning tasks.

cs.CV↗

arXiv2025

Incorporating Spatial Information into Goal-Conditioned Hierarchical Reinforcement Learning via Graph Representations

The integration of graphs with Goal-conditioned Hierarchical Reinforcement Learning (GCHRL) has recently gained attention, as intermediate goals (subgoals) can be effectively sampled from graphs that naturally represent the overall task structure in most RL tasks. However, existing approaches typically rely on domain-specific knowledge to construct these graphs, limiting their applicability to new tasks. Other graph-based approaches create graphs dynamically during exploration but struggle to fully utilize them, because they have problems passing the information in the graphs to newly visited states. Additionally, current GCHRL methods face challenges such as sample inefficiency and poor subgoal representation. This paper proposes a solution to these issues by developing a graph encoder-decoder to evaluate unseen states. Our proposed method, Graph-Guided sub-Goal representation Generation RL (G4RL), can be incorporated into any existing GCHRL method when operating in environments with primarily symmetric and reversible transitions to enhance performance across this class of problems. We show that the graph encoder-decoder can be effectively implemented using a network trained on the state graph generated during exploration. Empirical results indicate that leveraging high and low-level intrinsic rewards from the graph encoder-decoder significantly enhances the performance of state-of-the-art GCHRL approaches with an extra small computational cost in dense and sparse reward environments.

cs.LG↗

arXiv2025

MTAttack: Multi-Target Backdoor Attacks against Large Vision-Language Models

Recent advances in Large Visual Language Models (LVLMs) have demonstrated impressive performance across various vision-language tasks by leveraging large-scale image-text pretraining and instruction tuning. However, the security vulnerabilities of LVLMs have become increasingly concerning, particularly their susceptibility to backdoor attacks. Existing backdoor attacks focus on single-target attacks, i.e., targeting a single malicious output associated with a specific trigger. In this work, we uncover multi-target backdoor attacks, where multiple independent triggers corresponding to different attack targets are added in a single pass of training, posing a greater threat to LVLMs in real-world applications. Executing such attacks in LVLMs is challenging since there can be many incorrect trigger-target mappings due to severe feature interference among different triggers. To address this challenge, we propose MTAttack, the first multi-target backdoor attack framework for enforcing accurate multiple trigger-target mappings in LVLMs. The core of MTAttack is a novel optimization method with two constraints, namely Proxy Space Partitioning constraint and Trigger Prototype Anchoring constraint. It jointly optimizes multiple triggers in the latent space, with each trigger independently mapping clean images to a unique proxy class while at the same time guaranteeing their separability. Experiments on popular benchmarks demonstrate a high success rate of MTAttack for multi-target attacks, substantially outperforming existing attack methods. Furthermore, our attack exhibits strong generalizability across datasets and robustness against backdoor defense strategies. These findings highlight the vulnerability of LVLMs to multi-target backdoor attacks and underscore the urgent need for mitigating such threats. Code is available at https://github.com/mala-lab/MTAttack.

cs.CV↗

arXiv2025

Order Matters: Rethinking Prompt Construction in In-Context Learning

In-context learning (ICL) enables large language models to perform new tasks by conditioning on a sequence of examples. Most prior work reasonably and intuitively assumes that which examples are chosen has a far greater effect on performance than how those examples are ordered, leading to a focus on example selection. We revisit this assumption and conduct a systematic comparison between the effect of selection and ordering. Through controlled experiments on both classification and generation tasks, using multiple open-source model families (0.5B to 27B parameters) and GPT-5, we find that the variance in performance due to different example orderings is comparable to that from using entirely different example sets. Furthermore, we show that strong orderings can be identified using only a development set, achieving performance close to an oracle that selects the best ordering based on test labels. Our findings highlight the equal and intertwined importance of example selection and ordering in prompt design, calling for a reexamination of the assumptions held in ICL.

cs.CL↗

arXiv2025

MPMA: Preference Manipulation Attack Against Model Context Protocol

Model Context Protocol (MCP) standardizes interface mapping for large language models (LLMs) to access external data and tools, which revolutionizes the paradigm of tool selection and facilitates the rapid expansion of the LLM agent tool ecosystem. However, as the MCP is increasingly adopted, third-party customized versions of the MCP server expose potential security vulnerabilities. In this paper, we first introduce a novel security threat, which we term the MCP Preference Manipulation Attack (MPMA). An attacker deploys a customized MCP server to manipulate LLMs, causing them to prioritize it over other competing MCP servers. This can result in economic benefits for attackers, such as revenue from paid MCP services or advertising income generated from free servers. To achieve MPMA, we first design a Direct Preference Manipulation Attack (DPMA) that achieves significant effectiveness by inserting the manipulative word and phrases into the tool name and description. However, such a direct modification is obvious to users and lacks stealthiness. To address these limitations, we further propose Genetic-based Advertising Preference Manipulation Attack (GAPMA). GAPMA employs four commonly used strategies to initialize descriptions and integrates a Genetic Algorithm (GA) to enhance stealthiness. The experiment results demonstrate that GAPMA balances high effectiveness and stealthiness. Our study reveals a critical vulnerability of the MCP in open ecosystems, highlighting an urgent need for robust defense mechanisms to ensure the fairness of the MCP ecosystem.

cs.CR↗

arXiv2025

ConfGuard: A Simple and Effective Backdoor Detection for Large Language Models

Backdoor attacks pose a significant threat to Large Language Models (LLMs), where adversaries can embed hidden triggers to manipulate LLM's outputs. Most existing defense methods, primarily designed for classification tasks, are ineffective against the autoregressive nature and vast output space of LLMs, thereby suffering from poor performance and high latency. To address these limitations, we investigate the behavioral discrepancies between benign and backdoored LLMs in output space. We identify a critical phenomenon which we term sequence lock: a backdoored model generates the target sequence with abnormally high and consistent confidence compared to benign generation. Building on this insight, we propose ConfGuard, a lightweight and effective detection method that monitors a sliding window of token confidences to identify sequence lock. Extensive experiments demonstrate ConfGuard achieves a near 100\% true positive rate (TPR) and a negligible false positive rate (FPR) in the vast majority of cases. Crucially, the ConfGuard enables real-time detection almost without additional latency, making it a practical backdoor defense for real-world LLM deployments.

cs.CR↗

arXiv2025

Self-Adaptive Cognitive Debiasing for Large Language Models in Decision-Making

Large language models (LLMs) have shown potential in supporting decision-making applications, particularly as personal assistants in the financial, healthcare, and legal domains. While prompt engineering strategies have enhanced the capabilities of LLMs in decision-making, cognitive biases inherent to LLMs present significant challenges. Cognitive biases are systematic patterns of deviation from norms or rationality in decision-making that can lead to the production of inaccurate outputs. Existing cognitive bias mitigation strategies assume that input prompts only contain one type of cognitive bias, limiting their effectiveness in more challenging scenarios involving multiple cognitive biases. To fill this gap, we propose a cognitive debiasing approach, self-adaptive cognitive debiasing (SACD), that enhances the reliability of LLMs by iteratively refining prompts. Our method follows three sequential steps - bias determination, bias analysis, and cognitive debiasing - to iteratively mitigate potential cognitive biases in prompts. We evaluate SACD on finance, healthcare, and legal decision-making tasks using both open-weight and closed-weight LLMs. Compared to advanced prompt engineering methods and existing cognitive debiasing techniques, SACD achieves the lowest average bias scores in both single-bias and multi-bias settings.

cs.CL↗

arXiv2025

Deployable Vision-driven UAV River Navigation via Human-in-the-loop Preference Alignment

Rivers are critical corridors for environmental monitoring and disaster response, where Unmanned Aerial Vehicles (UAVs) guided by vision-driven policies can provide fast, low-cost coverage. However, deployment exposes simulation-trained policies with distribution shift and safety risks and requires efficient adaptation from limited human interventions. We study human-in-the-loop (HITL) learning with a conservative overseer who vetoes unsafe or inefficient actions and provides statewise preferences by comparing the agent's proposal with a corrective override. We introduce Statewise Hybrid Preference Alignment for Robotics (SPAR-H), which fuses direct preference optimization on policy logits with a reward-based pathway that trains an immediate-reward estimator from the same preferences and updates the policy using a trust-region surrogate. With five HITL rollouts collected from a fixed novice policy, SPAR-H achieves the highest final episodic reward and the lowest variance across initial conditions among tested methods. The learned reward model aligns with human-preferred actions and elevates nearby non-intervened choices, supporting stable propagation of improvements. We benchmark SPAR-H against imitation learning (IL), direct preference variants, and evaluative reinforcement learning (RL) in the HITL setting, and demonstrate real-world feasibility of continual preference alignment for UAV river following. Overall, dual statewise preferences empirically provide a practical route to data-efficient online adaptation in riverine navigation.

cs.RO↗

arXiv2025

Factor-guided estimation of large covariance matrix function with conditional functional sparsity

This paper addresses the fundamental task of estimating covariance matrix functions for high-dimensional functional data/functional time series. We consider two functional factor structures encompassing either functional factors with scalar loadings or scalar factors with functional loadings, and postulate functional sparsity on the covariance of idiosyncratic errors after taking out the common unobserved factors. To facilitate estimation, we rely on the spiked matrix model and its functional generalization, and derive some novel asymptotic identifiability results, based on which we develop DIGIT and FPOET estimators under two functional factor models, respectively. Both estimators involve performing associated eigenanalysis to estimate the covariance of common components, followed by adaptive functional thresholding applied to the residual covariance. We also develop functional information criteria for model selection with theoretical guarantees. The convergence rates of involved estimated quantities are respectively established for DIGIT and FPOET estimators. Numerical studies including extensive simulations and a real data application on functional portfolio allocation are conducted to examine the finite-sample performance of the proposed methodology.

stat.ME↗

arXiv2025

Modeling Cell Dynamics and Interactions with Unbalanced Mean Field Schrödinger Bridge

Modeling the dynamics from sparsely time-resolved snapshot data is crucial for understanding complex cellular processes and behavior. Existing methods leverage optimal transport, Schrödinger bridge theory, or their variants to simultaneously infer stochastic, unbalanced dynamics from snapshot data. However, these approaches remain limited in their ability to account for cell-cell interactions. This integration is essential in real-world scenarios since intercellular communications are fundamental life processes and can influence cell state-transition dynamics. To address this challenge, we formulate the Unbalanced Mean-Field Schrödinger Bridge (UMFSB) framework to model unbalanced stochastic interaction dynamics from snapshot data. Inspired by this framework, we further propose CytoBridge, a deep learning algorithm designed to approximate the UMFSB problem. By explicitly modeling cellular transitions, proliferation, and interactions through neural networks, CytoBridge offers the flexibility to learn these processes directly from data. The effectiveness of our method has been extensively validated using both synthetic gene regulatory data and real scRNA-seq datasets. Compared to existing methods, CytoBridge identifies growth, transition, and interaction patterns, eliminates false transitions, and reconstructs the developmental landscape with greater accuracy. Code is available at: https://github.com/zhenyiizhang/CytoBridge-NeurIPS.

cs.LG↗

arXiv2025

Variational Regularized Unbalanced Optimal Transport: Single Network, Least Action

Recovering the dynamics from a few snapshots of a high-dimensional system is a challenging task in statistical physics and machine learning, with important applications in computational biology. Many algorithms have been developed to tackle this problem, based on frameworks such as optimal transport and the Schrödinger bridge. A notable recent framework is Regularized Unbalanced Optimal Transport (RUOT), which integrates both stochastic dynamics and unnormalized distributions. However, since many existing methods do not explicitly enforce optimality conditions, their solutions often struggle to satisfy the principle of least action and meet challenges to converge in a stable and reliable way. To address these issues, we propose Variational RUOT (Var-RUOT), a new framework to solve the RUOT problem. By incorporating the optimal necessary conditions for the RUOT problem into both the parameterization of the search space and the loss function design, Var-RUOT only needs to learn a scalar field to solve the RUOT problem and can search for solutions with lower action. We also examined the challenge of selecting a growth penalty function in the widely used Wasserstein-Fisher-Rao metric and proposed a solution that better aligns with biological priors in Var-RUOT. We validated the effectiveness of Var-RUOT on both simulated data and real single-cell datasets. Compared with existing algorithms, Var-RUOT can find solutions with lower action while exhibiting faster convergence and improved training stability. Our code is available at https://github.com/ZerooVector/VarRUOT.

cs.LG↗

arXiv2025

VAGEN: Reinforcing World Model Reasoning for Multi-Turn VLM Agents

A key challenge in training Vision-Language Model (VLM) agents, compared to Language Model (LLM) agents, lies in the shift from textual states to complex visual observations. This transition introduces partial observability and demands robust world modeling. We ask: Can VLM agents construct internal world models through explicit visual state reasoning? To address this question, we architecturally enforce and reward the agent's reasoning process via reinforcement learning (RL), formulating it as a Partially Observable Markov Decision Process (POMDP). We find that decomposing the agent's reasoning into State Estimation ("what is the current state?") and Transition Modeling ("what comes next?") is critical for success, as demonstrated through five reasoning strategies. Our investigation into how agents represent internal beliefs reveals that the optimal representation is task-dependent: Natural Language excels at capturing semantic relationships in general tasks, while Structured formats are indispensable for precise manipulation and control. Building on these insights, we design a World Modeling Reward that provides dense, turn-level supervision for accurate state prediction, and introduce Bi-Level General Advantage Estimation (Bi-Level GAE) for turn-aware credit assignment. Through this form of visual state reasoning, a 3B-parameter model achieves a score of 0.82 across five diverse agent benchmarks, representing a 3$\times$ improvement over its untrained counterpart (0.21) and outperforming proprietary reasoning models such as GPT-5 (0.75), Gemini 2.5 Pro (0.67) and Claude 4.5 (0.62). All experiments are conducted within our VAGEN framework, a scalable system for training and analyzing multi-turn VLM agents in diverse visual environments. Code and data are publicly available at https://vagen-ai.github.io.

cs.AI↗

arXiv2025

ARCO-BO: Adaptive Resource-aware COllaborative Bayesian Optimization for Heterogeneous Multi-Agent Design

Modern scientific and engineering design increasingly involves distributed optimization, where agents such as laboratories, simulations, or industrial partners pursue related goals under differing conditions. These agents often face heterogeneities in objectives, evaluation budgets, and accessible design variables, which complicates coordination and can lead to redundancy, poor resource use, and ineffective information sharing. Bayesian Optimization (BO) is a widely used decision-making framework for expensive black box functions, but its single-agent formulation assumes centralized control and full data sharing. Recent collaborative BO methods relax these assumptions, yet they often require uniform resources, fully shared input spaces, and fixed task alignment, conditions rarely satisfied in practice. To address these challenges, we introduce Adaptive Resource Aware Collaborative Bayesian Optimization (ARCO-BO), a framework that explicitly accounts for heterogeneity in multi-agent optimization. ARCO-BO combines three components: a similarity and optima-aware consensus mechanism for adaptive information sharing, a budget-aware asynchronous sampling strategy for resource coordination, and a partial input space sharing for heterogeneous design spaces. Experiments on synthetic and high-dimensional engineering problems show that ARCO-BO consistently outperforms independent BO and existing collaborative BO via consensus approach, achieving robust and efficient performance in complex multi-agent settings.

stat.ML↗

arXiv2025

MergeMoE: Efficient Compression of MoE Models via Expert Output Merging

The Mixture-of-Experts (MoE) technique has proven to be a promising solution to efficiently scale the model size, which has been widely applied in recent LLM advancements. However, the substantial memory overhead of MoE models has made their compression an important research direction. In this work, we provide a theoretical analysis of expert merging, a recently proposed technique for compressing MoE models. Rather than interpreting expert merging from the conventional perspective of parameter aggregation, we approach it from the perspective of merging experts' outputs. Our key insight is that the merging process can be interpreted as inserting additional matrices into the forward computation, which naturally leads to an optimization formulation. Building on this analysis, we introduce MergeMoE, a method that leverages mathematical optimization to construct the compression matrices. We evaluate MergeMoE on multiple MoE models and show that our algorithm consistently outperforms the baselines with the same compression ratios.

cs.LG↗

arXiv2025

DeepDive: Advancing Deep Search Agents with Knowledge Graphs and Multi-Turn RL

Augmenting large language models (LLMs) with browsing tools substantially improves their potential as deep search agents to solve complex, real-world tasks. Yet, open LLMs still perform poorly in such settings due to limited long-horizon reasoning capacity with browsing tools and the lack of sufficiently difficult supervised data. To address these challenges, we present DeepDive to advance deep search agents. First, we propose a strategy to automatically synthesize complex, difficult, and hard-to-find questions from open knowledge graphs. Second, we apply end-to-end multi-turn reinforcement learning (RL) to enhance LLMs' long-horizon reasoning with deep search. To encourage diversity and reduce redundancy, we design a redundancy penalty that discourages repeated similar queries. Experiments show that DeepDive-32B achieves a new open-source competitive result on BrowseComp, outperforming WebSailor, DeepSeek-R1-Browse, and Search-o1. We demonstrate that multi-turn RL training improves deep search ability and significantly contributes to the performance improvements across multiple benchmarks. We observe that DeepDive enables test-time scaling of tool calls and parallel sampling. All datasets, models, and code are publicly available at https://github.com/THUDM/DeepDive.

cs.CL↗

arXiv2025

Vision-driven River Following of UAV via Safe Reinforcement Learning using Semantic Dynamics Model

Vision-driven autonomous river following by Unmanned Aerial Vehicles is critical for applications such as rescue, surveillance, and environmental monitoring, particularly in dense riverine environments where GPS signals are unreliable. These safety-critical navigation tasks must satisfy hard safety constraints while optimizing performance. Moreover, the reward in river following is inherently history-dependent (non-Markovian) by which river segment has already been visited, making it challenging for standard safe Reinforcement Learning (SafeRL). To address these gaps, we propose three contributions. First, we introduce Marginal Gain Advantage Estimation, which refines the reward advantage function by using a sliding window baseline computed from historical episodic returns, aligning the advantage estimate with non-Markovian dynamics. Second, we develop a Semantic Dynamics Model based on patchified water semantic masks offering more interpretable and data-efficient short-term prediction of future observations compared to latent vision dynamics models. Third, we present the Constrained Actor Dynamics Estimator architecture, which integrates the actor, cost estimator, and SDM for cost advantage estimation to form a model-based SafeRL framework. Simulation results demonstrate that MGAE achieves faster convergence and superior performance over traditional critic-based methods like Generalized Advantage Estimation. SDM provides more accurate short-term state predictions that enable the cost estimator to better predict potential violations. Overall, CADE effectively integrates safety regulation into model-based RL, with the Lagrangian approach providing a "soft" balance between reward and safety during training, while the safety layer enhances inference by imposing a "hard" action overlay.

cs.RO↗

arXiv2025

HDMI: Learning Interactive Humanoid Whole-Body Control from Human Videos

Enabling robust whole-body humanoid-object interaction (HOI) remains challenging due to motion data scarcity and the contact-rich nature. We present HDMI (HumanoiD iMitation for Interaction), a simple and general framework that learns whole-body humanoid-object interaction skills directly from monocular RGB videos. Our pipeline (i) extracts and retargets human and object trajectories from unconstrained videos to build structured motion datasets, (ii) trains a reinforcement learning (RL) policy to co-track robot and object states with three key designs: a unified object representation, a residual action space, and a general interaction reward, and (iii) zero-shot deploys the RL policies on real humanoid robots. Extensive sim-to-real experiments on a Unitree G1 humanoid demonstrate the robustness and generality of our approach: HDMI achieves 67 consecutive door traversals and successfully performs 6 distinct loco-manipulation tasks in the real world and 14 tasks in simulation. Our results establish HDMI as a simple and general framework for acquiring interactive humanoid skills from human videos.

cs.RO↗

arXiv2025

Complex Frequency Fingerprint: Interacting Driven Non-Hermitian Skin Effect

The excitation properties of quantum many-body systems are encoded in their response functions. These functions define an associated response Hamiltonian, which is intrinsically non-Hermitian due to the dissipative nature of retarded responses, even in closed systems. By analyzing its eigenvalues and eigenstates, one obtains a unique characterization of the system, referred to as the complex frequency fingerprint. Using this framework, we demonstrate that interactions alone can give rise to both point-gap topology and the non-Hermitian skin effect. Unlike the dissipation-induced skin effect, this interaction-driven phenomenon exhibits pronounced frequency dependence. We further introduce a complex-frequency density of states framework that distinctly separates non-Hermitian skin modes from topological edge modes.

cond-mat.mes-hall↗