论文检索

Zhibin Ni, Hai Wan, Xibin Zhao

To identify the root causes of attacks, behavior abstraction (BA) converts audit logs into multiple behavior graphs and finds similar ones, which has proven effective in bridging the semantic gap and reducing manual workload. Existing works fail to achieve both interpretability and generalization, while also exhibiting limited robustness when facing adversarial attacks. In this paper, we give the first attempt at interpretable and robust behavior abstraction and propose a novel method called Environment-Disentangled Heterogeneous Graph Neural Network (EDHGNN). Motivated by Information Bottleneck (IB) principle, we propose a Heterogeneous Subgraph Disentanglement (HSD) module to disentangle label-relevant and environmental subgraphs through single optimization. We also introduce an Adapted Graph-Level Attention (AGLA) module to extract minimal sufficient representations from label-relevant subgraphs, a Label-Guided Graph Reconstructor (LGGR) to maximize environmental information coverage via reconstruction, and a Relevance Discriminator (RD) to enhance disentanglement quality. Additionally, we construct a new dataset contains ground-truth explanations and 4,160 behavior graphs. Extensive experiments demonstrate that EDHGNN outperforms the state-of-the-art methods in terms of interpretability and robustness against adversarial attacks.

Ziqiang Liu, Bowen Li, Zhenyu Xu, Yantao Li, Junwei Zhang, Chulin Sha, Xiaolin Li

Inspired by the success of large language models (LLMs) in natural language processing, cell language models (CLMs) have emerged as a promising paradigm to learn cell representations from high-dimensional single-cell data—particularly transcriptomic profiles from scRNA-seq. These foundation models have shown remarkable potential across a variety of downstream applications. However, there remains a lack of foundation models for scATAC-seq data, which measures chromatin accessibility at single-cell level and is critical for decoding epigenetic regulation. Developing such model is considerably more challenging due to the unique characteristics of scATAC-seq data, including the vast number of chromatin regions, lack of standardized annotations, extreme sparsity, and near-binary distributions. To address these challenges, we systematically explore various strategies and propose CLM-Access, a specialized foundation model for scATAC-seq data. CLM-Access incorporates three main innovations: (1) an unified data processing pipeline that maps 2.8 million cells onto an unified reference of over 1 million chromatin regions; (2) a specialized patching and embedding strategy to effectively manage high-dimensional inputs; and (3) a tailored masking and loss function design that preserves fine-grained regional information while enhancing training efficiency and representation quality. With comprehensive benchmarks, we show that CLM-Access significantly outperforms existing methods in key downstream tasks, including batch effect correction, cell type annotation, RNA expression prediction, and multi-modal integration. This work establishes a scalable and interpretable foundation model for single-cell epigenomic analysis and expands the application of CLMs in single-cell research.

Xiangyu Liu, Haodi Lei, Yi Liu, Yang Liu, Wei Hu

Sparse Autoencoder (SAE) has emerged as a powerful tool for mechanistic interpretability of large language models. Recent works apply SAE to protein language models (PLMs), aiming to extract and analyze biologically meaningful features from their latent spaces. However, SAE suffers from semantic entanglement, where individual neurons often mix multiple nonlinear concepts, making it difficult to reliably interpret or manipulate model behaviors. In this paper, we propose a semantically-guided SAE, called ProtSAE. Unlike existing SAE which requires annotation datasets to filter and interpret activations, we guide semantic disentanglement during training using both annotation datasets and domain knowledge to mitigate the effects of entangled attributes. We design interpretability experiments showing that ProtSAE learns more biologically relevant and interpretable hidden features compared to previous methods. Performance analyses further demonstrate that ProtSAE maintains high reconstruction fidelity while achieving better results in interpretable probing. We also show the potential of ProtSAE in steering PLMs for downstream generation tasks.

Maoqi Liu, Quan Fang, Yuhao Wu, Can Zhao, Yang Yang, Kaiquan Cai

Accurate interpretation of Notices To Airmen (NOTAMs) is critical for aviation safety, yet their condensed and cryptic language poses significant challenges to both manual and automated processing. Existing automated systems are typically limited to "Shallow Parsing," failing to extract the actionable intelligence needed for operational decisions. We formalize the complete interpretation task as "Deep Parsing," a dual-reasoning challenge requiring both dynamic knowledge grounding (linking the NOTAM to evolving real-world aeronautical data) and schema-based inference (applying static domain rules to deduce operational status). To tackle this challenge, we propose NOTAM-Evolve, a self-evolving framework that enables a Large Language Model (LLM) to autonomously master complex NOTAM interpretation. Leveraging a knowledge graph-enhanced retrieval module for data grounding, the framework introduces a crucial closed-loop learning process where the LLM progressively improves from its own outputs, minimizing the need for extensive human-annotated reasoning traces. In conjunction with this framework, we introduce a new benchmark dataset of 10,000 expert-annotated NOTAMs. Our experiments demonstrate that NOTAM-Evolve achieves a 30.4% absolute accuracy improvement over the base LLM, establishing a new state-of-the-art on the task of structured NOTAM interpretation.

Shunfan Li, Jiangkai Long, Xin Zou, Chang Tang, Yuanyuan Liu, Xiao He, Xuesong Yan

Protein-Ligand Affinity (PLA) prediction quantifies the interaction strength to guide rational drug design. Existing approaches typically analyze interaction at a single granularity and overlook tightly coupled relationships between protein and ligand in both structure and functionality, consequently yielding suboptimal representations, leading to significant performance drops in real-world scenarios. To address this problem, we propose PLA-MGRA, a minimalist and effective PLA prediction framework. Specifically, PLA-MGRA captures both fine-grained atomic details and coarse grained functional semantics within the 3D structure of protein–ligand complexes, through multi-granularity learning. To further parse the coupled protein–ligand relationships, we design relation-aware learning to enhance the binding nature of representations. Extensive experiments demonstrate that our method achieves state-of-the-art performance on multiple protein–ligand affinity prediction benchmarks, while also offering generalizability and interpretability.

Yingxin Lai, Yufei Liu, Guoqing Yang, Jiaxing Chai, Zhiming Luo, Shaozi Li

Despite recent advancements in font generation, practitioners still grapple with a laborious trial-and-error workflow. To streamline this, we propose OneFont, an end-to-end framework that interprets user intents via free-form dialogue, seamlessly integrating both glyph synthesis and refinement modules. We introduce the Font with Thought (FwT) paradigm, reframing font design as a reasoning task where the model plans actions and articulates design rationales. OneFont’s core planner is trained via a two-stage regimen to master this paradigm. First, we instill reasoning abilities via Supervised Fine-Tuning (SFT) on a new, comprehensive benchmark of 1,500 font families we built. Second, we refine the model's policy with a novel reinforcement learning algorithm, Group Relative Policy Optimization (GRPO), guided by a hybrid reward that assesses visual fidelity, rationale coherence, and transformation correctness. Extensive experiments show OneFont significantly surpasses existing methods in design quality and stroke precision across diverse scripts, validated on our new benchmark. We will release our dataset, code, and models.

Yitian Kou, Yihe Gu, Chen Zhou, Dandan Zhu, Shu-Guang Kuai

Navigating human-populated environments without causing discomfort is a critical capability for socially-aware agents. While rule-based approaches offer interpretability through predefined psychological principles, they often lack generalizability and flexibility. Conversely, data-driven methods can learn complex behaviors from large-scale datasets, but are typically inefficient, opaque, and difficult to align with human intuitions. To bridge this gap, we propose RLSLM, a hybrid Reinforcement Learning framework that integrates a rule-based Social Locomotion Model, grounded in empirical behavioral experiments, into the reward function of a reinforcement learning framework. The social locomotion model generates an orientation-sensitive social comfort field that quantifies human comfort across space, enabling socially aligned navigation policies with minimal training. RLSLM then jointly optimizes mechanical energy and social comfort, allowing agents to avoid intrusions into personal or group space. A human-agent interaction experiment using an immersive VR-based setup demonstrates that RLSLM outperforms state-of-the-art rule-based models in user experience. Ablation and sensitivity analyses further show the model’s significantly improved interpretability over conventional data-driven methods. This work presents a scalable, human-centered methodology that effectively integrates cognitive science and machine learning for real-world social navigation.

Victor Shea-Jay Huang, Le Zhuo, Yi Xin, Zhaokai Wang, Fu-Yun Wang, Yuchi Wang, Renrui Zhang, Peng Gao, Hongsheng Li

Diffusion Transformers (DiTs) are a powerful yet underexplored class of generative models compared to U-Net-based diffusion architectures. We propose TIDE—Temporal-aware sparse autoencoders for Interpretable Diffusion transformErs—a framework designed to extract sparse, interpretable activation features across timesteps in DiTs. TIDE effectively captures temporally-varying representations and reveals that DiTs naturally learn hierarchical semantics (e.g., 3D structure, object class, and fine-grained concepts) during large-scale pretraining. Experiments show that TIDE enhances interpretability and controllability while maintaining reasonable generation quality, enabling applications such as safe image editing and style transfer.

Mingyang Geng, Shanzhi Gu, Zhipeng Liu, Chuanfu Xu, Zhaoyang Qu, Haotian Wang

Recent advances in multi-agent Large Language Model-based code generation enable collaborative software development through role-specialized agents. However, failure localization of code generation remains challenging due to inter-agent dependencies and solution-path multiplicity. Consequently, existing prompting-based localization methods exhibit vulnerability towards semantically valid but non-canonical strategies. To address this, we propose FLKR (Failure Localization via Knowledge-guided Reasoning), an self-supervised framework that combines behavior encoding, knowledge-strategy alignment, and consistency scoring for solution-path invariant localization. To evaluate, we also introduce COFL (Code Oriented Failure Localization), the first expert-annotated benchmark for fine-grained failure localization. Experiments show FLKR outperforms state-of-the-art prompting-based baselines by up to 14 points in Fault Localization Accuracy and 45 points in Top-1 accuracy, with strong performance in divergent, real-world, and refinement-critical cases. Such results demonstrate that our proposed FLKR generalizes well to real-world software development scenarios and opens up a new direction for failure-aware refinement recommendation by providing precise and interpretable responsibility signals.

Yufei Gao, Zhengong Cai, Bowei Yang

As microservice architectures become increasingly complex and system events become more frequent, Root Cause Analysis (RCA) has emerged as a critical task to ensure system reliability. However, existing deep learning-based methods often struggle with limited flexibility and a lack of interpretability when addressing complex system failures. Recent efforts to integrate large language models (LLMs) have shown promise in enhancing diagnostic transparency and reasoning capability. However, expansive search spaces, intricate workflows, and entangled constraints constrain practical adoption. We propose RCAFlow, a multi-agent framework that integrates structured workflow knowledge with hierarchical planning to address these challenges. RCAFlow transforms semi-structured documents into behavior tree-style workflows to support interpretable plan generation, employs a Git-inspired branching mechanism for modular and hierarchical task execution with path isolation, and leverages state-aware task execution with semantic analysis to improve result understanding and feedback. We evaluate RCAFlow on three benchmark datasets provided by OpenRCA. Experimental results demonstrate that RCAFlow consistently outperforms existing methods across all datasets. Further ablation studies confirm the effectiveness of each core module, highlighting the reliability, extensibility, and interpretability of RCAFlow to support complex RCA tasks within intelligent IT operations.

Yichao Feng, Haoran Luo, Lang Feng, Shuai Zhao, Anh Tuan Luu

Large Language Models show promise in emotion understanding, social reasoning, and empathy, yet struggle with psychologically grounded tasks requiring inference of implicit mental states in complex, socially and contextually ambiguous settings. These limitations stem from lacking theory-aligned supervision and difficulty capturing nuanced mental processes in real-world narratives. To bridge this gap, we leverage expert-labeled scenarios and propose a trajectory-aware reinforcement learning framework imitating expert psychological reasoning. By integrating real-world stimuli with structured reasoning guidance, our approach enables compact models to internalize social-cognitive principles, perform nuanced inference, and support continual self-improvement. Experiments across benchmarks show expert-level interpretive capability across psychological tasks.

Ke Ding, Brian Parker, Jiayu Wen

Pre-training large language models on genomic sequences has become a powerful approach for learning biologically meaningful representations. While masked language modeling (MLM)-based approaches, such as DNABERT and Nucleotide Transformer (NT), achieve strong performance, they are hindered by inefficiencies due to partial token supervision, pre-training/fine-tuning mismatches, and high computational costs. We introduce NucEL, the first ELECTRA-style pre-training framework for genomic foundation models, which overcomes these challenges. Through a discriminator network identifying tokens modified by a generator, NucEL achieves comprehensive token-level supervision across all sequence positions, thereby markedly improving training efficiency relative to the partial supervision of masked positions inherent in MLM frameworks. By integrating ModernBERT’s architectural advancements, including hybrid local-global attention and flash attention mechanisms, NucEL establishes an optimized BERT architecture for genomic sequence modeling. Unlike traditional methods that tokenize genomic sequences into 6-mers, NucEL implements single-nucleotide tokenization, enabling fine-grained resolution and improving both efficiency and interpretability. Pre-trained on the human genome only, NucEL achieves state-of-the-art performance on benchmark datasets across diverse downstream tasks in both human and non-human species, including regulatory element identification (e.g., promoters, enhancers), transcription factor binding prediction in human and mouse, open chromatin region classification, and histone modification profiles, surpassing MLM-based models of similar size and rivaling models 25 times larger, such as NT. Ablation studies provide critical insights into tokenization and masking strategies, optimizing ELECTRA-style pretraining for DNA sequences. Attention analyses reveal NucEL’s superior ability to capture biologically relevant sequence motifs compared to NT, offering valuable insights into its hierarchical learning process and regulatory element modeling capabilities. This work highlights the potential of ELECTRA-style pretraining as an efficient and effective strategy for advancing genomic representation learning with broad implications for future genomic research.

Shengqi Dang, Fu Chai, Jiaxin Li, Chao Yuan, Wei Ye, Nan Cao

The rise of 3D generative models has enabled automatic 3D geometry and texture synthesis from multimodal inputs (e.g., text or images). However, these methods often ignore physical constraints and manufacturability considerations. In this work, we address the challenge of producing 3D designs that are both lightweight and self-supporting. We present DensiCrafter, a framework for generating lightweight, self-supporting 3D hollow structures by optimizing the density field. Starting from coarse voxel grids produced by Trellis, we interpret these as continuous density fields to optimize and introduce three differentiable, physically constrained, and simulation-free loss terms. Additionally, a mass regularization penalizes unnecessary material, while a restricted optimization domain preserves the outer surface. Our method seamlessly integrates with pretrained Trellis-based models (e.g., Trellis, DSO) without any architectural changes. In extensive evaluations, we achieve up to 43% reduction in material mass on the text-to-3D task. Compared to state-of-the-art baselines, our method could improve the stability and maintain high geometric fidelity. Real-world 3D-printing experiments confirm that our hollow designs can be reliably fabricated and could be self-supporting.

Yihong Tang, Haicheng Liao, Tong Nie, Junlin He, Ao Qu, Kehua Chen, Wei Ma, Zhenning Li, Lijun Sun, Chengzhong Xu

End-to-end autonomous driving (AD) systems increasingly adopt vision-language-action (VLA) models, yet they ignore the passenger's emotional state, which is central to comfort and AD acceptance. We introduce Open-Domain End-to-End (OD-E2E) AD, where an autonomous vehicle must interpret free-form natural-language commands, infer the emotion, and plan a physically feasible trajectory. We propose E3AD, an emotion-aware VLA framework that augments semantic understanding with two cognitively inspired components: a continuous Valence-Arousal-Dominance (VAD) emotion model that captures tone and urgency from language, and a dual-pathway spatial reasoning module that fuses egocentric and allocentric views for human-like spatial cognition. A consistency-oriented training scheme, combining modality pretraining with preference-based alignment, further enforces coherence between emotional intent and driving actions. Across real-world datasets, E3AD improves visual grounding and waypoint planning and achieves state-of-the-art (SOTA) VAD correlation for emotion estimation. These results show that injecting emotion into VLA-style driving yields more human-aligned grounding, planning, and feedback.

Peter Yongho Kim, Juhyeon Park, Jungwoo Park, Jubin Choi, Jungwoo Seo, Jiook Cha, Taesup Moon

Modeling long-range spatiotemporal dynamics in functional Magnetic Resonance Imaging (fMRI) remains a key challenge due to the high dimensionality of the four-dimensional signals. Prior voxel-based models, although demonstrating excellent performance and interpretation capabilities, are constrained by prohibitive memory demands and thus can only capture limited temporal windows. To address this, we propose TABLeT (Two-dimensionally Autoencoded Brain Latent Transformer), a novel approach that tokenizes fMRI volumes using a pre-trained 2D natural image autoencoder. Each 3D fMRI volume is compressed into a compact set of continuous tokens, enabling long-sequence modeling with a simple Transformer encoder with limited VRAM. Across large-scale benchmarks including the UK-Biobank (UKB), Human Connectome Project (HCP), and ADHD-200 datasets, TABLeT outperforms existing models in multiple tasks, while demonstrating substantial gains in computational and memory efficiency over the state-of-the-art voxel-based method given the same input. Furthermore, we develop a self-supervised masked token modeling approach to pre-train TABLeT, which improves the model's performance for various downstream tasks. Our findings suggest a promising approach for scalable and interpretable spatiotemporal modeling of brain activity. Our code is available at https://github.com/beotborry/TABLeT.

Xingyu Zhu, Beier Zhu, Shuo Wang, Junfeng Fang, Kesen Zhao, Hanwang Zhang, Xiangnan He

As vision-language models (VLMs) are increasingly deployed in open-world scenarios, they can be easily induced by visual jailbreak attacks to generate harmful content, posing serious risks to model safety and trustworthy usage.Recent activation steering methods inject directional vectors into model activations during inference to induce refusal behaviors and have demonstrated effectiveness.However, a steering vector may both enhance refusal ability and cause over-refusal, thereby degrading model performance on benign inputs.Moreover, due to the lack of theoretical interpretability, these methods still suffer from limited robustness and effectiveness.To better balance safety and utility, we propose \texttt NullSteer , a null-space projected activation defense framework.Our method constructs refusal directions within model activations through a linear transformation: it maintains zero perturbation within the benign subspace while dynamically inducing refusal along potentially harmful directions, thereby theoretically achieving safety enhancement without impairing the model's general capabilities.Extensive experiments show that \texttt NullSteer significantly reduces harmful outputs under various jailbreak attacks (average ASR reduction over 15% on MiniGPT-4) while maintaining comparable performance to the original model on general benchmarks.

Hao Zhou, Tiru Wu, Yan Jiang, Wanqi Zhou, Junxing Hu, Ai Han

Multi-modal multi-agent systems (MM-MAS) have gained increasing attention for their capacity to enable complex reasoning and coordination across diverse modalities. As these systems continue to expand in scale and functionality, investigating their potential vulnerabilities has become increasingly important.However, existing studies on adversarial attacks in multi-agent systems primarily focus on isolated agents or unimodal settings, leaving the vulnerabilities of MM-MAS largely underexplored. To bridge this gap, we introduce HAM\textsuperscript 3 , a Hierarchical Attack framework for multi-modal multi-agent systems that decomposes attacks into three interconnected layers. Specifically, at the perception layer, HAM\textsuperscript 3 mounts attacks by perturbing visual inputs, textual inputs, and their fused visual-textual representations. At the communication layer, it performs communication-level attacks that corrupt message content and interaction topology, such as manipulating shared context or communication links to distort collective information flow. At the reasoning layer, it conducts reasoning-level attacks that interfere with each agent's cognitive pipeline, biasing reasoning trajectories and ultimately compromising final decisions. We evaluate HAM\textsuperscript 3 on the GQA benchmark through multi-agent systems built on distinct reasoning paradigms including ReAct, Plan-and-Solve, and Reflexion. Experiments demonstrate that our framework achieves an Attack Success Rate of up to 78.3%, with reasoning-layer attacks being the most effective. More than half of the successful attacks lead multiple agents to produce consistent errors. These findings offer valuable insights for building more robust and interpretable multi-agent intelligence.

Qi Song, Honglin Li, Yingchen Yu, Haoyi Zhou, Lin Yang, Song Bai, Qi She, Zilong Huang, Yunqing Zhao

Recent releases such as o3 highlight human-like "thinking with images" reasoning that combines tool use with stepwise verification, yet most open-source approaches still rely on text-only chains, rigid visual schemas, or single-step pipelines, limiting flexibility, interpretability, and transferability on complex tasks. We introduce CodeDance, which explores executable code as a general solver for visual reasoning. Unlike fixed-schema calls (e.g., only predicting bounding-box coordinates), CodeDance defines, composes, and executes code to orchestrate multiple tools, compute intermediate results, and render visual artifacts (e.g., boxes, lines, plots) that support transparent, self-checkable reasoning. To guide this process, we introduce a reward for balanced and adaptive tool calling, which balances exploration with efficiency and mitigates tool overuse. Interestingly, beyond the expected capabilities taught by atomic supervision, we empirically observe novel emergent behaviors during RL training: CodeDance demonstrates novel tool invocations, unseen compositions, and cross-task transfer. These behaviors arise without task-specific fine-tuning, suggesting a general and scalable mechanism for executable visual reasoning. Extensive experiments across reasoning benchmarks (e.g., visual search, math, chart QA) show that CodeDance not only consistently outperforms schema-driven and text-only baselines, but also surpasses closed models such as GPT-4o and larger open-source models.

Zheng Wang, Haoran Chen, Haoxuan Qin, Zhipeng Wei, Tianwen Qian, Cong Bai

Long video understanding is challenging due to dense visual redundancy, long-range temporal dependencies, and the tendency of chain-of-thought and retrieval-based agents to accumulate semantic drift and correlation-driven errors. We argue that long-video reasoning should begin not with reactive retrieval, but with deliberate task formulation: the model must first articulate what must be true in the video for each candidate answer to hold. This thinking-before-finding principle motivates VideoHV-Agent, a framework that reformulates video question answering as a structured hypothesis-verification process.Based on video summaries, a Thinker rewrites answer candidates into testable hypotheses, a Judge derives a discriminative clue specifying what evidence must be checked, a Verifier grounds and tests the clue using localized, fine-grained video content, and an Answer agent integrates validated evidence to produce the final answer.Experiments on three long-video understanding benchmarks show that VideoHV-Agent achieves state-of-the-art accuracy while providing enhanced interpretability, improved logical soundness, and lower computational cost.

Guangchen Shi, Yirui Wu, Wei Zhu, Tao Wang, Hao Zhang, Bo Li, Tong Lu

Few-shot Semantic Segmentation (FSS) aims to segment objects of novel categories given only a handful of labeled examples. However, existing methods often rely on complex category-specific modeling, resulting in high computational cost and limited generalization under low-data regimes. To address these challenges, we propose a Bayesian Probabilistic Network (BPNet) that reformulates FSS as a composition of three interpretable components: a prior, a likelihood, and a class-consistency term. Specifically, an efficient Segment Anything Model (SAM) is employed to generate fragmented prior regions for the query image, while both the likelihood and the consistency terms are estimated by a lightweight Class-Agnostic Localization Model (CALM). CALM simultaneously predicts the class consistency between support-query pairs through a binary classification head and estimates the likelihood by localizing the target region in the support image. By evaluating SAM-generated regions in parallel, CALM can efficiently identify the core region, thereby transforming the segmentation problem into a simple binary classification task. Furthermore, to mitigate the semantic incompleteness of SAM proposals, we introduce an attention-based Semantic Completion Module (SCM), which leverages local and global context cues to integrate fragmented regions into semantically complete masks. Extensive experiments demonstrate that BPNet achieves state-of-the-art performance while maintaining high efficiency.