# Research Papers — 2025-10 · bundle page 139 of 217

> Full-text markdown of papers #6901–6950 of 10843 published in 2025-10, newest first. AI-readable mirrors of the full 10843-paper 2025-10 archive (50 papers per file).

Canonical page: https://aipod.dev/papers/archive/2025-10/page/139
[← Previous bundle](https://aipod.dev/papers/archive/2025-10/page/138.md)
[Next bundle →](https://aipod.dev/papers/archive/2025-10/page/140.md)

### A Unified Frequency Domain Decomposition Framework for Interpretable and Robust Time Series Forecasting

Authors: Cheng He, Xijie Liang, Zengrong Zheng, Patrick P. C. Lee, Xu Huang, Zhaoyi Li, Hong Xie, Defu Lian, Enhong Chen
Published: 2025-10-11
Categories: cs.LG, cs.AI
arXiv: https://arxiv.org/abs/2510.10145v1

Current approaches for time series forecasting, whether in the time or frequency domain, predominantly use deep learning models based on linear layers or transformers. They often encode time series data in a black-box manner and rely on trial-and-error optimization solely based on forecasting performance, leading to limited interpretability and theoretical understanding. Furthermore, the dynamics in data distribution over time and frequency domains pose a critical challenge to accurate forecasting. We propose FIRE, a unified frequency domain decomposition framework that provides a mathematical abstraction for diverse types of time series, so as to achieve interpretable and robust time series forecasting. FIRE introduces several key innovations: (i) independent modeling of amplitude and phase components, (ii) adaptive learning of weights of frequency basis components, (iii) a targeted loss function, and (iv) a novel training paradigm for sparse data. Extensive experiments demonstrate that FIRE consistently outperforms state-of-the-art models on long-term forecasting benchmarks, achieving superior predictive performance and significantly enhancing interpretability of time series

---

### Debiasing LLMs by Masking Unfairness-Driving Attention Heads

Authors: Tingxu Han, Wei Song, Ziqi Ding, Ziming Li, Chunrong Fang, Yuekang Li, Dongfang Liu, Zhenyu Chen, Zhenting Wang
Published: 2025-10-11
Categories: cs.CL, cs.AI
arXiv: https://arxiv.org/abs/2510.10142v3

Large language models (LLMs) increasingly mediate decisions in domains where unfair treatment of demographic groups is unacceptable. Existing work probes when biased outputs appear, but gives little insight into the mechanisms that generate them, leaving existing mitigations largely fragile. In this paper, we conduct a systematic investigation LLM unfairness and propose DiffHeads, a lightweight debiasing framework for LLMs. We first compare Direct-Answer (DA) prompting to Chain-of-Thought (CoT) prompting across eight representative open- and closed-source LLMs. DA will trigger the nature bias part of LLM and improve measured unfairness by 534.5%-391.9% in both one-turn and two-turn dialogues. Next, we define a token-to-head contribution score that traces each token's influence back to individual attention heads. This reveals a small cluster of bias heads that activate under DA but stay largely dormant with CoT, providing the first causal link between prompting strategy and bias emergence. Finally, building on this insight, we propose DiffHeads that identifies bias heads through differential activation analysis between DA and CoT, and selectively masks only those heads. DiffHeads reduces unfairness by 49.4%, and 40.3% under DA and CoT, respectively, without harming model utility.

---

### YOLOv11-Litchi: Efficient Litchi Fruit Detection based on UAV-Captured Agricultural Imagery in Complex Orchard Environments

Authors: Hongxing Peng, Haopei Xie, Weijia Lia, Huanai Liuc, Ximing Li
Published: 2025-10-11
Categories: cs.CV, cs.LG, eess.IV
arXiv: https://arxiv.org/abs/2510.10141v1

Litchi is a high-value fruit, yet traditional manual selection methods are increasingly inadequate for modern production demands. Integrating UAV-based aerial imagery with deep learning offers a promising solution to enhance efficiency and reduce costs. This paper introduces YOLOv11-Litchi, a lightweight and robust detection model specifically designed for UAV-based litchi detection. Built upon the YOLOv11 framework, the proposed model addresses key challenges such as small target size, large model parameters hindering deployment, and frequent target occlusion. To tackle these issues, three major innovations are incorporated: a multi-scale residual module to improve contextual feature extraction across scales, a lightweight feature fusion method to reduce model size and computational costs while maintaining high accuracy, and a litchi occlusion detection head to mitigate occlusion effects by emphasizing target regions and suppressing background interference. Experimental results validate the model's effectiveness. YOLOv11-Litchi achieves a parameter size of 6.35 MB - 32.5% smaller than the YOLOv11 baseline - while improving mAP by 2.5% to 90.1% and F1-Score by 1.4% to 85.5%. Additionally, the model achieves a frame rate of 57.2 FPS, meeting real-time detection requirements. These findings demonstrate the suitability of YOLOv11-Litchi for UAV-based litchi detection in complex orchard environments, showcasing its potential for broader applications in precision agriculture.

---

### Adversarial Attacks on Downstream Weather Forecasting Models: Application to Tropical Cyclone Trajectory Prediction

Authors: Yue Deng, Francisco Santos, Pang-Ning Tan, Lifeng Luo
Published: 2025-10-11
Categories: cs.LG, cs.CR, stat.ML
arXiv: https://arxiv.org/abs/2510.10140v2

Deep learning-based weather forecasting (DLWF) models leverage past weather observations to generate future forecasts, supporting a wide range of downstream applications, including tropical cyclone (TC) prediction. In this paper, we investigate their vulnerability to adversarial attacks, where subtle perturbations to the upstream forecasts can alter the downstream TC trajectory predictions. Although research into adversarial attacks on DLWF models has grown recently, it remains challenging to craft perturbed upstream forecasts that steer the downstream outputs toward attacker-specified trajectories. First, conventional TC detection systems are opaque, non-differentiable black boxes, making standard gradient-based attacks infeasible. Second, the extreme rarity of TC events leads to severe class imbalance problem, making it difficult to develop attack methods for perturbing upstream forecasts that produce realistic-looking cyclone paths aligned with attacker's target trajectories. To overcome these limitations, we propose Cyc-Attack, a novel method for perturbing the upstream forecasts of DLWF models to generate adversarial trajectories. The proposed method uses a differentiable surrogate model to approximate the TC detector's output, enabling the application of gradient-based attacks. Cyc-Attack also employs a skewness-aware loss function with kernel dilation strategy to address the imbalance problem. Finally, a distance-based gradient weighting scheme and regularization are used to constrain the perturbations and eliminate unrealistic-looking trajectories, thereby making the adversarial upstream forecasts less easily detectable. Our experiments show that Cyc-Attack achieves a higher true positive rate in matching the attacker's target trajectories, along with lower false alarm rates and stealthier perturbations than conventional attack methods.

---

### Hybrid OCR-LLM Framework for Enterprise-Scale Document Information Extraction Under Copy-heavy Task

Authors: Zilong Wang, Xiaoyu Shen
Published: 2025-10-11
Categories: cs.CL, cs.AI
arXiv: https://arxiv.org/abs/2510.10138v1

Information extraction from copy-heavy documents, characterized by massive volumes of structurally similar content, represents a critical yet understudied challenge in enterprise document processing. We present a systematic framework that strategically combines OCR engines with Large Language Models (LLMs) to optimize the accuracy-efficiency trade-off inherent in repetitive document extraction tasks. Unlike existing approaches that pursue universal solutions, our method exploits document-specific characteristics through intelligent strategy selection. We implement and evaluate 25 configurations across three extraction paradigms (direct, replacement, and table-based) on identity documents spanning four formats (PNG, DOCX, XLSX, PDF). Through table-based extraction methods, our adaptive framework delivers outstanding results: F1=1.0 accuracy with 0.97s latency for structured documents, and F1=0.997 accuracy with 0.6 s for challenging image inputs when integrated with PaddleOCR, all while maintaining sub-second processing speeds. The 54 times performance improvement compared with multimodal methods over naive approaches, coupled with format-aware routing, enables processing of heterogeneous document streams at production scale. Beyond the specific application to identity extraction, this work establishes a general principle: the repetitive nature of copy-heavy tasks can be transformed from a computational burden into an optimization opportunity through structure-aware method selection.

---

### PermLLM: Learnable Channel Permutation for N:M Sparse Large Language Models

Authors: Lancheng Zou, Shuo Yin, Zehua Pei, Tsung-Yi Ho, Farzan Farnia, Bei Yu
Published: 2025-10-11
Categories: cs.LG, cs.AI
arXiv: https://arxiv.org/abs/2510.10136v1

Channel permutation is a powerful technique for enhancing the accuracy of N:M sparse models by reordering the channels of weight matrices to prioritize the retention of important weights. However, traditional channel permutation methods rely on handcrafted quality metrics, which often fail to accurately capture the true impact of pruning on model performance. To address this limitation, we propose PermLLM, a novel post-training pruning framework that introduces learnable channel permutation (LCP) for N:M sparsity. LCP leverages Sinkhorn normalization to transform discrete permutation matrices into differentiable soft permutation matrices, enabling end-to-end optimization. Additionally, PermLLM incorporates an efficient block-wise channel permutation strategy, which significantly reduces the number of learnable parameters and computational complexity. PermLLM seamlessly integrates with existing one-shot pruning methods to adaptively optimize channel permutations, effectively mitigating pruning-induced errors. Extensive experiments on the LLaMA series, Qwen, and OPT models demonstrate that PermLLM achieves superior performance in optimizing N:M sparse models. The code is available at https://github.com/lanchengzou/PermLLM.

---

### CharCom: Composable Identity Control for Multi-Character Story Illustration

Authors: Zhongsheng Wang, Ming Lin, Zhedong Lin, Yaser Shakib, Qian Liu, Jiamou Liu
Published: 2025-10-11
Categories: cs.AI
arXiv: https://arxiv.org/abs/2510.10135v2

Ensuring character identity consistency across varying prompts remains a fundamental limitation in diffusion-based text-to-image generation. We propose CharCom, a modular and parameter-efficient framework that achieves character-consistent story illustration through composable LoRA adapters, enabling efficient per-character customization without retraining the base model. Built on a frozen diffusion backbone, CharCom dynamically composes adapters at inference using prompt-aware control. Experiments on multi-scene narratives demonstrate that CharCom significantly enhances character fidelity, semantic alignment, and temporal coherence. It remains robust in crowded scenes and enables scalable multi-character generation with minimal overhead, making it well-suited for real-world applications such as story illustration and animation.

---

### CacheClip: Accelerating RAG with Effective KV Cache Reuse

Authors: Bin Yang, Qiuyu Leng, Jun Zeng, Zhenhua Wu
Published: 2025-10-11
Categories: cs.LG, cs.AI
arXiv: https://arxiv.org/abs/2510.10129v2

Retrieval-Augmented Generation (RAG) systems suffer from severe time-to-first-token (TTFT) bottlenecks due to long input sequences. Existing KV cache reuse methods face a fundamental trade-off: prefix caching requires identical prefixes that rarely occur in RAG scenarios, while direct precomputation sacrifices quality due to missing inter-chunk attention and repeated attention sinks. Recent methods like APE and CacheBlend partially address these issues but remain inadequate for robust RAG applications. This paper presents CacheClip, a novel framework that achieves both fast TTFT and high generation quality. Our key insight is that small auxiliary LLMs exhibit similar last-layer attention distributions to primary LLMs (the target model for generation), enabling efficient identification of tokens critical for restoring inter-chunk attention, thereby significantly improving response quality on cross-chunk reasoning tasks. CacheClip integrates four techniques: (1) auxiliary-model-guided token selection for selective KV cache recomputation, (2) shared prefixes to eliminate redundant attention sinks, (3) a sliding-window grouping strategy to maintain local coherence during partial KV cache updates, and (4) a CPU-GPU hybrid design that offloads auxiliary model inference to idle CPU resources, avoiding additional GPU overhead. The recomputation ratio is adjustable, allowing users to flexibly balance efficiency and quality for different deployment requirements. Experiments show CacheClip retains up to 85.2% and 91.1% of full-attention performance on NIAH and LongBench, outperforming CacheBlend and APE by 16.1 and 12.8 points on NIAH, and by 4.5 and 4.2 points on LongBench (with recomp% = 20%). Meanwhile, CacheClip accelerates LLM inference by up to 3.33$\times$ in prefill time (with recomp% = 20%), providing a practical solution to the efficiency-quality trade-off in RAG systems.

---

### Ctrl-World: A Controllable Generative World Model for Robot Manipulation

Authors: Yanjiang Guo, Lucy Xiaoyang Shi, Jianyu Chen, Chelsea Finn
Published: 2025-10-11
Categories: cs.RO, cs.AI
arXiv: https://arxiv.org/abs/2510.10125v3

Generalist robot policies can now perform a wide range of manipulation skills, but evaluating and improving their ability with unfamiliar objects and instructions remains a significant challenge. Rigorous evaluation requires a large number of real-world rollouts, while systematic improvement demands additional corrective data with expert labels. Both of these processes are slow, costly, and difficult to scale. World models offer a promising, scalable alternative by enabling policies to rollout within imagination space. However, a key challenge is building a controllable world model that can handle multi-step interactions with generalist robot policies. This requires a world model compatible with modern generalist policies by supporting multi-view prediction, fine-grained action control, and consistent long-horizon interactions, which is not achieved by previous works. In this paper, we make a step forward by introducing a controllable multi-view world model that can be used to evaluate and improve the instruction-following ability of generalist robot policies. Our model maintains long-horizon consistency with a pose-conditioned memory retrieval mechanism and achieves precise action control through frame-level action conditioning. Trained on the DROID dataset (95k trajectories, 564 scenes), our model generates spatially and temporally consistent trajectories under novel scenarios and new camera placements for over 20 seconds. We show that our method can accurately rank policy performance without real-world robot rollouts. Moreover, by synthesizing successful trajectories in imagination and using them for supervised fine-tuning, our approach can improve policy success by 44.7\%.

---

### The Hybrid Multimodal Graph Index (HMGI): A Comprehensive Framework for Integrated Relational and Vector Search

Authors: Joydeep Chandra, Satyam Kumar Navneet, Yong Zhang
Published: 2025-10-11
Categories: cs.DB, cs.LG
arXiv: https://arxiv.org/abs/2510.10123v1

The proliferation of complex, multimodal datasets has exposed a critical gap between the capabilities of specialized vector databases and traditional graph databases. While vector databases excel at semantic similarity search, they lack the capacity for deep relational querying. Conversely, graph databases master complex traversals but are not natively optimized for high-dimensional vector search. This paper introduces the Hybrid Multimodal Graph Index (HMGI), a novel framework designed to bridge this gap by creating a unified system for efficient, hybrid queries on multimodal data. HMGI leverages the native graph database architecture and integrated vector search capabilities, exemplified by platforms like Neo4j, to combine Approximate Nearest Neighbor Search (ANNS) with expressive graph traversal queries. Key innovations of the HMGI framework include modality-aware partitioning of embeddings to optimize index structure and query performance, and a system for adaptive, low-overhead index updates to support dynamic data ingestion, drawing inspiration from the architectural principles of systems like TigerVector. By integrating semantic similarity search directly with relational context, HMGI aims to outperform pure vector databases like Milvus in complex, relationship-heavy query scenarios and achieve sub-linear query times for hybrid tasks.

---

### DeepFusionNet: Autoencoder-Based Low-Light Image Enhancement and Super-Resolution

Authors: Halil Hüseyin Çalışkan, Talha Koruk
Published: 2025-10-11
Categories: cs.CV, cs.AI
arXiv: https://arxiv.org/abs/2510.10122v1

Computer vision and image processing applications suffer from dark and low-light images, particularly during real-time image transmission. Currently, low light and dark images are converted to bright and colored forms using autoencoders; however, these methods often achieve low SSIM and PSNR scores and require high computational power due to their large number of parameters. To address these challenges, the DeepFusionNet architecture has been developed. According to the results obtained with the LOL-v1 dataset, DeepFusionNet achieved an SSIM of 92.8% and a PSNR score of 26.30, while containing only approximately 2.5 million parameters. On the other hand, conversion of blurry and low-resolution images into high-resolution and blur-free images has gained importance in image processing applications. Unlike GAN-based super-resolution methods, an autoencoder-based super resolution model has been developed that contains approximately 100 thousand parameters and uses the DeepFusionNet architecture. According to the results of the tests, the DeepFusionNet based super-resolution method achieved a PSNR of 25.30 and a SSIM score of 80.7 percent according to the validation set.

---

### Multi Class Parkinson Disease Detection Based on Finger Tapping Using Attention Enhanced CNN BiLSTM

Authors: Abu Saleh Musa Miah, Najmul Hassan, Md Maruf Al Hossain, Yuichi Okuyama, Jungpil Shin
Published: 2025-10-11
Categories: cs.CV
arXiv: https://arxiv.org/abs/2510.10121v2

Accurate evaluation of Parkinsons disease (PD) severity is essential for effective clinical management and intervention development. Despite the proposal of several gesture based PD recognition systems, including those using the finger tapping task to assess Parkinsonian symptoms, their performance remains unsatisfactory. In this study, we present a multi class PD detection system based on finger-tapping, using an attention-enhanced CNN BiLSTM framework combined with handcrafted feature extraction and deep learning techniques. In the procedure, we used an existing dataset of finger tapping videos to extract temporal, frequency, and amplitude-based features from wrist and hand movements using their formulas. These handcrafted features were then processed through our attention enhanced CNN BiLSTM model, a hybrid deep learning framework that integrates CNN, BiLSTM, and attention mechanisms to classify PD severity into multiple levels. The features first pass through a Conv1D MaxPooling block to capture local spatial dependencies, followed by processing through a BiLSTM layer to model the temporal dynamics of the motion. An attention mechanism is applied to emphasize the most informative temporal features, which are then refined by a second BiLSTM layer. The CNN derived features and attention enhanced BiLSTM outputs are concatenated, followed by dense and dropout layers, before being passed through a softmax classifier to predict the PD severity level. Our model demonstrated strong performance in distinguishing between the five severity classes, showcasing the effectiveness of combining spatial temporal representations with attention mechanisms for automated PD severity detection. This approach offers a promising non invasive tool to assist clinicians in monitoring PD progression and making informed treatment decisions.

---

### DixitWorld: Evaluating Multimodal Abductive Reasoning in Vision-Language Models with Multi-Agent Dixit Gameplay

Authors: Yunxiang Mo, Tianshi Zheng, Qing Zong, Jiayu Liu, Baixuan Xu, Yauwai Yim, Chunkit Chan, Jiaxin Bai, Yangqiu Song
Published: 2025-10-11
Categories: cs.AI
arXiv: https://arxiv.org/abs/2510.10117v1

Multimodal abductive reasoning--the generation and selection of explanatory hypotheses from partial observations--is a cornerstone of intelligence. Current evaluations of this ability in vision-language models (VLMs) are largely confined to static, single-agent tasks. Inspired by Dixit, we introduce DixitWorld, a comprehensive evaluation suite designed to deconstruct this challenge. DIXITWORLD features two core components: DixitArena, a dynamic, multi-agent environment that evaluates both hypothesis generation (a "storyteller" crafting cryptic clues) and hypothesis selection ("listeners" choosing the target image from decoys) under imperfect information; and DixitBench, a static QA benchmark that isolates the listener's task for efficient, controlled evaluation. Results from DixitArena reveal distinct, role-dependent behaviors: smaller open-source models often excel as creative storytellers, producing imaginative yet less discriminative clues, whereas larger proprietary models demonstrate superior overall performance, particularly as listeners. Performance on DixitBench strongly correlates with listener results in DixitArena, validating it as a reliable proxy for hypothesis selection. Our findings reveal a key trade-off between generative creativity and discriminative understanding in multimodal abductive reasoning, a central challenge for developing more balanced and capable vision-language agents.

---

### Preference-driven Knowledge Distillation for Few-shot Node Classification

Authors: Xing Wei, Chunchun Chen, Rui Fan, Xiaofeng Cao, Sourav Medya, Wei Ye
Published: 2025-10-11
Categories: cs.LG, cs.SI
arXiv: https://arxiv.org/abs/2510.10116v3

Graph neural networks (GNNs) can efficiently process text-attributed graphs (TAGs) due to their message-passing mechanisms, but their training heavily relies on the human-annotated labels. Moreover, the complex and diverse local topologies of nodes of real-world TAGs make it challenging for a single mechanism to handle. Large language models (LLMs) perform well in zero-/few-shot learning on TAGs but suffer from a scalability challenge. Therefore, we propose a preference-driven knowledge distillation (PKD) framework to synergize the complementary strengths of LLMs and various GNNs for few-shot node classification. Specifically, we develop a GNN-preference-driven node selector that effectively promotes prediction distillation from LLMs to teacher GNNs. To further tackle nodes' intricate local topologies, we develop a node-preference-driven GNN selector that identifies the most suitable teacher GNN for each node, thereby facilitating tailored knowledge distillation from teacher GNNs to the student GNN. Extensive experiments validate the efficacy of our proposed framework in few-shot node classification on real-world TAGs. Our code is be available.

---

### LinearRAG: Linear Graph Retrieval Augmented Generation on Large-scale Corpora

Authors: Luyao Zhuang, Shengyuan Chen, Yilin Xiao, Huachi Zhou, Yujing Zhang, Hao Chen, Qinggang Zhang, Xiao Huang
Published: 2025-10-11
Categories: cs.CL
arXiv: https://arxiv.org/abs/2510.10114v4

Retrieval-Augmented Generation (RAG) is widely used to mitigate hallucinations of Large Language Models (LLMs) by leveraging external knowledge. While effective for simple queries, traditional RAG systems struggle with large-scale, unstructured corpora where information is fragmented. Recent advances incorporate knowledge graphs to capture relational structures, enabling more comprehensive retrieval for complex, multi-hop reasoning tasks. However, existing graph-based RAG (GraphRAG) methods rely on unstable and costly relation extraction for graph construction, often producing noisy graphs with incorrect or inconsistent relations that degrade retrieval quality. In this paper, we revisit the pipeline of existing GraphRAG systems and propose LinearRAG (Linear Graph-based Retrieval-Augmented Generation), an efficient framework that enables reliable graph construction and precise passage retrieval. Specifically, LinearRAG constructs a relation-free hierarchical graph, termed Tri-Graph, using only lightweight entity extraction and semantic linking, avoiding unstable relation modeling. This new paradigm of graph construction scales linearly with corpus size and incurs no extra token consumption, providing an economical and reliable indexing of the original passages. For retrieval, LinearRAG adopts a two-stage strategy: (i) relevant entity activation via local semantic bridging, followed by (ii) passage retrieval through global importance aggregation. Extensive experiments on four datasets demonstrate that LinearRAG significantly outperforms baseline models. Our code and datasets are available at https://github.com/DEEP-PolyU/LinearRAG.

---

### ImmerIris: A Large-Scale Dataset and Benchmark for Off-Axis and Unconstrained Iris Recognition in Immersive Applications

Authors: Yuxi Mi, Qiuyang Yuan, Zhizhou Zhong, Xuan Zhao, Jiaogen Zhou, Fubao Zhu, Jihong Guan, Shuigeng Zhou
Published: 2025-10-11
Categories: cs.CV
arXiv: https://arxiv.org/abs/2510.10113v3

Recently, iris recognition is regaining prominence in immersive applications such as extended reality as a means of seamless user identification. This application scenario introduces unique challenges compared to traditional iris recognition under controlled setups, as the ocular images are primarily captured off-axis and less constrained, causing perspective distortion, intra-subject variation, and quality degradation in iris textures. Datasets capturing these challenges remain limited. This paper fills this gap by presenting a large-scale iris dataset collected via head-mounted displays, termed ImmerIris. It contains 499,791 ocular images from 546 subjects, and is, to our knowledge, the largest public iris dataset to date and among the first dedicated to immersive applications. It is accompanied by a comprehensive set of evaluation protocols that benchmark recognition systems under various challenging conditions. This paper also draws attention to a shared obstacle of current recognition methods, the reliance on a pre-processing, normalization stage, which is fallible in off-axis and unconstrained setups. To this end, this paper further proposes a normalization-free paradigm that directly learns from minimally adjusted ocular images. Despite its simplicity, it outperforms normalization-based prior arts, indicating a promising direction for robust iris recognition.

---

### Training-Free In-Context Forensic Chain for Image Manipulation Detection and Localization

Authors: Rui Chen, Bin Liu, Changtao Miao, Xinghao Wang, Yi Li, Tao Gong, Qi Chu, Nenghai Yu
Published: 2025-10-11
Categories: cs.CV, cs.AI, cs.CR
arXiv: https://arxiv.org/abs/2510.10111v3

Advances in image tampering pose serious security threats, underscoring the need for effective image manipulation localization (IML). While supervised IML achieves strong performance, it depends on costly pixel-level annotations. Existing weakly supervised or training-free alternatives often underperform and lack interpretability. We propose the In-Context Forensic Chain (ICFC), a training-free framework that leverages multi-modal large language models (MLLMs) for interpretable IML tasks. ICFC integrates an objectified rule construction with adaptive filtering to build a reliable knowledge base and a multi-step progressive reasoning pipeline that mirrors expert forensic workflows from coarse proposals to fine-grained forensics results. This design enables systematic exploitation of MLLM reasoning for image-level classification, pixel-level localization, and text-level interpretability. Across multiple benchmarks, ICFC not only surpasses state-of-the-art training-free methods but also achieves competitive or superior performance compared to weakly and fully supervised approaches.

---

### Uncertainty-Aware Post-Detection Framework for Enhanced Fire and Smoke Detection in Compact Deep Learning Models

Authors: Aniruddha Srinivas Joshi, Godwyn James William, Shreyas Srinivas Joshi
Published: 2025-10-11
Categories: cs.CV, cs.AI, cs.LG, eess.IV
arXiv: https://arxiv.org/abs/2510.10108v1

Accurate fire and smoke detection is critical for safety and disaster response, yet existing vision-based methods face challenges in balancing efficiency and reliability. Compact deep learning models such as YOLOv5n and YOLOv8n are widely adopted for deployment on UAVs, CCTV systems, and IoT devices, but their reduced capacity often results in false positives and missed detections. Conventional post-detection methods such as Non-Maximum Suppression and Soft-NMS rely only on spatial overlap, which can suppress true positives or retain false alarms in cluttered or ambiguous fire scenes. To address these limitations, we propose an uncertainty aware post-detection framework that rescales detection confidences using both statistical uncertainty and domain relevant visual cues. A lightweight Confidence Refinement Network integrates uncertainty estimates with color, edge, and texture features to adjust detection scores without modifying the base model. Experiments on the D-Fire dataset demonstrate improved precision, recall, and mean average precision compared to existing baselines, with only modest computational overhead. These results highlight the effectiveness of post-detection rescoring in enhancing the robustness of compact deep learning models for real-world fire and smoke detection.

---

### Lighter-X: An Efficient and Plug-and-play Strategy for Graph-based Recommendation through Decoupled Propagation

Authors: Yanping Zheng, Zhewei Wei, Frank de Hoog, Xu Chen, Hongteng Xu, Yuhang Ye, Jiadeng Huang
Published: 2025-10-11
Categories: cs.LG
arXiv: https://arxiv.org/abs/2510.10105v1

Graph Neural Networks (GNNs) have demonstrated remarkable effectiveness in recommendation systems. However, conventional graph-based recommenders, such as LightGCN, require maintaining embeddings of size $d$ for each node, resulting in a parameter complexity of $\mathcal{O}(n \times d)$, where $n$ represents the total number of users and items. This scaling pattern poses significant challenges for deployment on large-scale graphs encountered in real-world applications. To address this scalability limitation, we propose \textbf{Lighter-X}, an efficient and modular framework that can be seamlessly integrated with existing GNN-based recommender architectures. Our approach substantially reduces both parameter size and computational complexity while preserving the theoretical guarantees and empirical performance of the base models, thereby enabling practical deployment at scale. Specifically, we analyze the original structure and inherent redundancy in their parameters, identifying opportunities for optimization. Based on this insight, we propose an efficient compression scheme for the sparse adjacency structure and high-dimensional embedding matrices, achieving a parameter complexity of $\mathcal{O}(h \times d)$, where $h \ll n$. Furthermore, the model is optimized through a decoupled framework, reducing computational complexity during the training process and enhancing scalability. Extensive experiments demonstrate that Lighter-X achieves comparable performance to baseline models with significantly fewer parameters. In particular, on large-scale interaction graphs with millions of edges, we are able to attain even better results with only 1\% of the parameter over LightGCN.

---

### Answer-Consistent Chain-of-thought Reinforcement Learning For Multi-modal Large Langauge Models

Authors: Minbin Huang, Runhui Huang, Chuanyang Zheng, Jingyao Li, Guoxuan Chen, Han Shi, Hong Cheng
Published: 2025-10-11
Categories: cs.CV
arXiv: https://arxiv.org/abs/2510.10104v1

Recent advances in large language models (LLMs) have demonstrated that reinforcement learning with verifiable rewards (RLVR) can significantly enhance reasoning abilities by directly optimizing correctness, rather than relying solely on supervised imitation. This paradigm has been extended to multimodal LLMs for complex video and image understanding tasks. However, while outcome-driven RL improves answer accuracy, it can inadvertently decouple the reasoning chain from the final answer, leading to situations where models produce inconsistency between the reasoning trace and final answer. In our experiments on multiple-choice visual question-answering tasks, the standard GRPO method yields only 79.7\% consistency on MMVU between the reasoning steps and the chosen answers, indicating frequent mismatches between answers and reasoning. To this end, we propose Answer-Consistent Reinforcement Learning (ACRE) that modifies the GRPO algorithm with an auxiliary consistency check. After the model generates a chain of thought and an initial answer for a given question, we shuffle the answer options and prompt the model again with the same reasoning trace to predict a second answer. We design a consistency-verification reward that grants a high reward only if both the original and the post-shuffle answers agree and are correct; otherwise, a lower reward is assigned accordingly. This mechanism penalizes reasoning-answer misalignment and discourages the model from relying on spurious patterns, such as option ordering biases. We evaluate ACRE on challenging Video Reasoning benchmarks and multimodal math reasoning benchmarks, achieving an average 2.2\% and 1.5\% improvement for Video Reasoning and Math Reasoning tasks over the GRPO baseline.

---

### Stop When Enough: Adaptive Early-Stopping for Chain-of-Thought Reasoning

Authors: Renliang Sun, Wei Cheng, Dawei Li, Haifeng Chen, Wei Wang
Published: 2025-10-11
Categories: cs.CL
arXiv: https://arxiv.org/abs/2510.10103v1

Chain-of-Thought (CoT) reasoning has driven recent gains of large language models (LLMs) on reasoning-intensive tasks by externalizing intermediate steps. However, excessive or redundant reasoning -- so-called overthinking -- can increase inference costs and lead LLMs toward incorrect conclusions. In this paper, we present REFRAIN ($\underline{REF}$lective-$\underline{R}$edundancy for $\underline{A}$daptive $\underline{IN}$ference), a training-free framework that adaptively determines when to stop reasoning to mitigate overthinking. REFRAIN integrates a two-stage stop discriminator to identify reflective yet redundant reasoning and a sliding-window Upper Confidence Bound (SW-UCB) multi-armed bandit controller to dynamically adjust stopping thresholds according to problem difficulty without supervision or fine-tuning. Across four representative benchmarks and two model families, REFRAIN reduces token usage by 20-55% while maintaining or improving accuracy compared to standard CoT prompting. Extensive ablation and robustness analyses demonstrate its stability across models, scorers, and prompt variations. In summary, our findings highlight when-to-stop as a new and practical axis of test-time scaling -- enabling models to reason not just more, but just enough.

---

### PANTHER: Generative Pretraining Beyond Language for Sequential User Behavior Modeling

Authors: Guilin Li, Yun Zhang, Xiuyuan Chen, Chengqi Li, Bo Wang, Linghe Kong, Wenjia Wang, Weiran Huang, Matthias Hwai Yong Tan
Published: 2025-10-11
Categories: cs.LG
arXiv: https://arxiv.org/abs/2510.10102v2

Large language models (LLMs) have shown that generative pretraining can distill vast world knowledge into compact token representations. While LLMs encapsulate extensive world knowledge, they remain limited in modeling the behavioral knowledge contained within user interaction histories. User behavior forms a distinct modality, where each action, defined by multi-dimensional attributes such as time, context, and transaction type, constitutes a behavioral token. Modeling these high-cardinality sequences is challenging, and discriminative models often falter under limited supervision. To bridge this gap, we extend generative pretraining to user behavior, learning transferable representations from unlabeled behavioral data analogous to how LLMs learn from text. We present PANTHER, a hybrid generative-discriminative framework that unifies user behavior pretraining and downstream adaptation, enabling large-scale sequential user representation learning and real-time inference. PANTHER introduces: (1) Structured Tokenization to compress multi-dimensional transaction attributes into an interpretable vocabulary; (2) Sequence Pattern Recognition Module (SPRM) for modeling periodic transaction motifs; (3) a Unified User-Profile Embedding that fuses static demographics with dynamic transaction histories; and (4) Real-time scalability enabled by offline caching of pretrained embeddings for millisecond-level inference. Fully deployed and operational online at WeChat Pay, PANTHER delivers a 25.6 percent boost in next-transaction prediction HitRate@1 and a 38.6 percent relative improvement in fraud detection recall over baselines. Cross-domain evaluations on public benchmarks show strong generalization, achieving up to 21 percent HitRate@1 gains over transformer baselines, establishing PANTHER as a scalable, high-performance framework for industrial sequential user behavior modeling.

---

### On the Rademacher Complexity of Graph Neural Networks: Unifying Expressivity and Geometry

Authors: Martin Carrasco, Caio F. Deberaldini Netto, Vahan A. Martirosyan, Ehimare Okoyomon, Caterina Graziani
Published: 2025-10-11
Categories: cs.LG
arXiv: https://arxiv.org/abs/2510.10101v4

Understanding the interplay between generalization, expressivity, and the geometry of the input space is a central challenge in graph learning. The expressivity of Graph Neural Networks (GNNs) is typically characterized through their correspondence with graph invariants, such as those from the Weisfeiler-Leman (WL) hierarchy. While more expressive GNNs can distinguish a richer set of graphs, they are also associated with weaker generalization guarantees. Previous works have addressed this trade-off using the VC dimension, a purely combinatorial measure, independent of the training data. In this work, we adopt a data-dependent measure of generalization, the empirical Rademacher complexity, and derive tight generalization bounds that jointly consider the expressive power of GNNs and the geometry of the underlying input space. Specifically, any graph invariant that upper-bounds a GNN's expressive power partitions the input space into equivalence classes, and we show that the empirical Rademacher complexity is controlled by the distribution of training samples across these classes. Moving beyond discrete partitions, we incorporate the geometry of the input space and derive covering-number bounds under Lipschitz continuity, showing that the complexity cost can be mitigated when the hypothesis class remains smooth over the data geometry. In addition, we prove that the empirical Rademacher complexity is Lipschitz continuous with respect to the Wasserstein distance between empirical measures supported on different datasets. This yields robustness and generalization guarantees under sampling variability. Importantly, our framework is not restricted to message-passing GNNs or WL, but extends to arbitrary GNN architectures and their associated invariants, providing a step toward a unified theory of GNN generalization.

---

### Cooperative Pseudo Labeling for Unsupervised Federated Classification

Authors: Kuangpu Guo, Lijun Sheng, Yongcan Yu, Jian Liang, Zilei Wang, Ran He
Published: 2025-10-11
Categories: cs.CV, cs.LG
arXiv: https://arxiv.org/abs/2510.10100v1

Unsupervised Federated Learning (UFL) aims to collaboratively train a global model across distributed clients without sharing data or accessing label information. Previous UFL works have predominantly focused on representation learning and clustering tasks. Recently, vision language models (e.g., CLIP) have gained significant attention for their powerful zero-shot prediction capabilities. Leveraging this advancement, classification problems that were previously infeasible under the UFL paradigm now present promising new opportunities, yet remain largely unexplored. In this paper, we extend UFL to the classification problem with CLIP for the first time and propose a novel method, \underline{\textbf{Fed}}erated \underline{\textbf{Co}}operative \underline{\textbf{P}}seudo \underline{\textbf{L}}abeling (\textbf{FedCoPL}). Specifically, clients estimate and upload their pseudo label distribution, and the server adjusts and redistributes them to avoid global imbalance among classes. Moreover, we introduce a partial prompt aggregation protocol for effective collaboration and personalization. In particular, visual prompts containing general image features are aggregated at the server, while text prompts encoding personalized knowledge are retained locally. Extensive experiments demonstrate the superior performance of our FedCoPL compared to baseline methods. Our code is available at \href{https://github.com/krumpguo/FedCoPL}{https://github.com/krumpguo/FedCoPL}.

---

### Uncovering Singularities in Feynman Integrals via Machine Learning

Authors: Yuanche Liu, Yingxuan Xu, Yang Zhang
Published: 2025-10-11
Categories: hep-ph, cs.AI, cs.LG, hep-th
arXiv: https://arxiv.org/abs/2510.10099v2

We introduce a machine-learning framework based on symbolic regression to extract the full symbol alphabet of multi-loop Feynman integrals. By targeting the analytic structure rather than reduction, the method is broadly applicable and interpretable across different families of integrals. It successfully reconstructs complete symbol alphabets in nontrivial examples, demonstrating both robustness and generality. Beyond accelerating computations case by case, it uncovers the analytic structure universally. This framework opens new avenues for multi-loop amplitude analysis and provides a versatile tool for exploring scattering amplitudes.

---

### Gesplat: Robust Pose-Free 3D Reconstruction via Geometry-Guided Gaussian Splatting

Authors: Jiahui Lu, Haihong Xiao, Xueyan Zhao, Wenxiong Kang
Published: 2025-10-11
Categories: cs.CV
arXiv: https://arxiv.org/abs/2510.10097v2

Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS) have advanced 3D reconstruction and novel view synthesis, but remain heavily dependent on accurate camera poses and dense viewpoint coverage. These requirements limit their applicability in sparse-view settings, where pose estimation becomes unreliable and supervision is insufficient. To overcome these challenges, we introduce Gesplat, a 3DGS-based framework that enables robust novel view synthesis and geometrically consistent reconstruction from unposed sparse images. Unlike prior works that rely on COLMAP for sparse point cloud initialization, we leverage the VGGT foundation model to obtain more reliable initial poses and dense point clouds. Our approach integrates several key innovations: 1) a hybrid Gaussian representation with dual position-shape optimization enhanced by inter-view matching consistency; 2) a graph-guided attribute refinement module to enhance scene details; and 3) flow-based depth regularization that improves depth estimation accuracy for more effective supervision. Comprehensive quantitative and qualitative experiments demonstrate that our approach achieves more robust performance on both forward-facing and large-scale complex datasets compared to other pose-free methods.

---

### CardRewriter: Leveraging Knowledge Cards for Long-Tail Query Rewriting on Short-Video Platforms

Authors: Peiyuan Gong, Feiran Zhu, Yaqi Yin, Chenglei Dai, Chao Zhang, Kai Zheng, Wentian Bao, Jiaxin Mao, Yi Zhang
Published: 2025-10-11
Categories: cs.IR, cs.CL
arXiv: https://arxiv.org/abs/2510.10095v1

Short-video platforms have rapidly become a new generation of information retrieval systems, where users formulate queries to access desired videos. However, user queries, especially long-tail ones, often suffer from spelling errors, incomplete phrasing, and ambiguous intent, resulting in mismatches between user expectations and retrieved results. While large language models (LLMs) have shown success in long-tail query rewriting within e-commerce, they struggle on short-video platforms, where proprietary content such as short videos, live streams, micro dramas, and user social networks falls outside their training distribution. To address this challenge, we introduce \textbf{CardRewriter}, an LLM-based framework that incorporates domain-specific knowledge to enhance long-tail query rewriting. For each query, our method aggregates multi-source knowledge relevant to the query and summarizes it into an informative and query-relevant knowledge card. This card then guides the LLM to better capture user intent and produce more effective query rewrites. We optimize CardRewriter using a two-stage training pipeline: supervised fine-tuning followed by group relative policy optimization, with a tailored reward system balancing query relevance and retrieval effectiveness. Offline experiments show that CardRewriter substantially improves rewriting quality for queries targeting proprietary content. Online A/B testing further confirms significant gains in long-view rate (LVR) and click-through rate (CTR), along with a notable reduction in initiative query reformulation rate (IQRR). Since September 2025, CardRewriter has been deployed on Kuaishou, one of China's largest short-video platforms, serving hundreds of millions of users daily.

---

### What Makes Looped Transformers Perform Better Than Non-Recursive Ones

Authors: Zixuan Gong, Yong Liu, Jiaye Teng
Published: 2025-10-11
Categories: cs.LG, cs.AI, stat.ML
arXiv: https://arxiv.org/abs/2510.10089v3

While looped transformers (termed as Looped-Attn) often outperform standard transformers (termed as Single-Attn) on complex reasoning tasks, the mechanism for this advantage remains underexplored. In this paper, we explain this phenomenon through the lens of loss landscape geometry, inspired by empirical observations of their distinct dynamics at both sample and Hessian levels. To formalize this, we extend the River-Valley landscape model by distinguishing between U-shaped valleys (flat) and V-shaped valleys (steep). Based on empirical observations, we conjecture that the recursive architecture of Looped-Attn induces a landscape-level inductive bias towards River-V-Valley. This inductive bias suggest a better loss convergence along the river due to valley hopping, and further encourage learning about complex patterns compared to the River-U-Valley induced by Single-Attn. Building on this insight, we propose SHIFT (Staged HIerarchical Framework for Progressive Training), a principled training strategy that accelerates the training process of Looped-Attn while achieving comparable performances.

---

### Pharmacist: Safety Alignment Data Curation for Large Language Models against Harmful Fine-tuning

Authors: Guozhi Liu, Qi Mu, Tiansheng Huang, Xinhua Wang, Li Shen, Weiwei Lin, Zhang Li
Published: 2025-10-11
Categories: cs.CR, cs.AI, cs.LG
arXiv: https://arxiv.org/abs/2510.10085v1

Harmful fine-tuning issues present significant safety challenges for fine-tuning-as-a-service in large language models. Existing alignment-stage defenses, e.g., Vaccine, Repnoise, Booster, and T-Vaccine, mitigate harmful fine-tuning issues by enhancing the model's robustness during the alignment phase. While these methods have been proposed to mitigate the issue, they often overlook a critical upstream factor: the role of the original safety-alignment data. We observe that their defense performance and computational efficiency remain constrained by the quality and composition of the alignment dataset. To address this limitation, we propose Pharmacist, a safety alignment data curation solution that enhances defense against harmful fine-tuning by selecting a high-quality and safety-critical core subset from the original alignment data. The core idea of Pharmacist is to train an alignment data selector to rank alignment data. Specifically, up-ranking high-quality and safety-critical alignment data, down-ranking low-quality and non-safety-critical data. Empirical results indicate that models trained on datasets selected by Pharmacist outperform those trained on datasets selected by existing selection methods in both defense and inference performance. In addition, Pharmacist can be effectively integrated with mainstream alignment-stage defense methods. For example, when applied to RepNoise and T-Vaccine, using the dataset selected by Pharmacist instead of the full dataset leads to improvements in defense performance by 2.60\% and 3.30\%, respectively, and enhances inference performance by 3.50\% and 1.10\%. Notably, it reduces training time by 56.83\% and 57.63\%, respectively. Our code is available at https://github.com/Lslland/Pharmacist.

---

### Tracking the Spatiotemporal Evolution of Landslide Scars Using a Vision Foundation Model: A Novel and Universal Framework

Authors: Meijun Zhou, Gang Mei, Zhengjing Ma, Nengxiong Xu, Jianbing Peng
Published: 2025-10-11
Categories: cs.CV
arXiv: https://arxiv.org/abs/2510.10084v1

Tracking the spatiotemporal evolution of large-scale landslide scars is critical for understanding the evolution mechanisms and failure precursors, enabling effective early-warning. However, most existing studies have focused on single-phase or pre- and post-failure dual-phase landslide identification. Although these approaches delineate post-failure landslide boundaries, it is challenging to track the spatiotemporal evolution of landslide scars. To address this problem, this study proposes a novel and universal framework for tracking the spatiotemporal evolution of large-scale landslide scars using a vision foundation model. The key idea behind the proposed framework is to reconstruct discrete optical remote sensing images into a continuous video sequence. This transformation enables a vision foundation model, which is developed for video segmentation, to be used for tracking the evolution of landslide scars. The proposed framework operates within a knowledge-guided, auto-propagation, and interactive refinement paradigm to ensure the continuous and accurate identification of landslide scars. The proposed framework was validated through application to two representative cases: the post-failure Baige landslide and the active Sela landslide (2017-2025). Results indicate that the proposed framework enables continuous tracking of landslide scars, capturing both failure precursors critical for early warning and post-failure evolution essential for assessing secondary hazards and long-term stability.

---

### Enabling High-Quality In-the-Wild Imaging from Severely Aberrated Metalens Bursts

Authors: Debabrata Mandal, Zhihan Peng, Yujie Wang, Praneeth Chakravarthula
Published: 2025-10-11
Categories: physics.optics, cs.CV
arXiv: https://arxiv.org/abs/2510.10083v1

We tackle the challenge of robust, in-the-wild imaging using ultra-thin nanophotonic metalens cameras. Meta-lenses, composed of planar arrays of nanoscale scatterers, promise dramatic reductions in size and weight compared to conventional refractive optics. However, severe chromatic aberration, pronounced light scattering, narrow spectral bandwidth, and low light efficiency continue to limit their practical adoption. In this work, we present an end-to-end solution for in-the-wild imaging that pairs a metalens several times thinner than conventional optics with a bespoke multi-image restoration framework optimized for practical metalens cameras. Our method centers on a lightweight convolutional network paired with a memory-efficient burst fusion algorithm that adaptively corrects noise, saturation clipping, and lens-induced distortions across rapid sequences of extremely degraded metalens captures. Extensive experiments on diverse, real-world handheld captures demonstrate that our approach consistently outperforms existing burst-mode and single-image restoration techniques.These results point toward a practical route for deploying metalens-based cameras in everyday imaging applications.

---

### Diversity Augmentation of Dynamic User Preference Data for Boosting Personalized Text Summarizers

Authors: Parthiv Chatterjee, Shivam Sonawane, Amey Hengle, Aditya Tanna, Sourish Dasgupta, Tanmoy Chakraborty
Published: 2025-10-11
Categories: cs.CL, cs.LG
arXiv: https://arxiv.org/abs/2510.10082v1

Document summarization enables efficient extraction of user-relevant content but is inherently shaped by individual subjectivity, making it challenging to identify subjective salient information in multifaceted documents. This complexity underscores the necessity for personalized summarization. However, training models for personalized summarization has so far been challenging, particularly because diverse training data containing both user preference history (i.e., click-skip trajectory) and expected (gold-reference) summaries are scarce. The MS/CAS PENS dataset is a valuable resource but includes only preference history without target summaries, preventing end-to-end supervised learning, and its limited topic-transition diversity further restricts generalization. To address this, we propose $\mathrm{PerAugy}$, a novel cross-trajectory shuffling and summary-content perturbation based data augmentation technique that significantly boosts the accuracy of four state-of-the-art baseline (SOTA) user-encoders commonly used in personalized summarization frameworks (best result: $\text{0.132}$$\uparrow$ w.r.t AUC). We select two such SOTA summarizer frameworks as baselines and observe that when augmented with their corresponding improved user-encoders, they consistently show an increase in personalization (avg. boost: $\text{61.2\%}\uparrow$ w.r.t. PSE-SU4 metric). As a post-hoc analysis of the role of induced diversity in the augmented dataset by \peraugy, we introduce three dataset diversity metrics -- $\mathrm{TP}$, $\mathrm{RTC}$, and \degreed\ to quantify the induced diversity. We find that $\mathrm{TP}$ and $\mathrm{DegreeD}$ strongly correlate with user-encoder performance on the PerAugy-generated dataset across all accuracy metrics, indicating that increased dataset diversity is a key factor driving performance gains.

---

### How AI Companionship Develops: Evidence from a Longitudinal Study

Authors: Angel Hsing-Chi Hwang, Fiona Li, Jacy Reese Anthis, Hayoun Noh
Published: 2025-10-11
Categories: cs.HC, cs.AI
arXiv: https://arxiv.org/abs/2510.10079v1

The quickly growing popularity of AI companions poses risks to mental health, personal wellbeing, and social relationships. Past work has identified many individual factors that can drive human-companion interaction, but we know little about how these factors interact and evolve over time. In Study 1, we surveyed AI companion users (N = 303) to map the psychological pathway from users' mental models of the agent to parasocial experiences, social interaction, and the psychological impact of AI companions. Participants' responses foregrounded multiple interconnected variables (agency, parasocial interaction, and engagement) that shape AI companionship. In Study 2, we conducted a longitudinal study with a subset of participants (N = 110) using a new generic chatbot. Participants' perceptions of the generic chatbot significantly converged to perceptions of their own companions by Week 3. These results suggest a longitudinal model of AI companionship development and demonstrate an empirical method to study human-AI companionship.

---

### Improving Speech Emotion Recognition with Mutual Information Regularized Generative Model

Authors: Chung-Soo Ahn, Rajib Rana, Sunil Sivadas, Carlos Busso, Jagath C. Rajapakse
Published: 2025-10-11
Categories: cs.SD, cs.LG
arXiv: https://arxiv.org/abs/2510.10078v4

Lack of large, well-annotated emotional speech corpora continues to limit the performance and robustness of speech emotion recognition (SER), particularly as models grow more complex and the demand for multimodal systems increases. While generative data augmentation offers a promising solution, existing approaches often produce emotionally inconsistent samples due to oversimplified conditioning on categorical labels. This paper introduces a novel mutual-information-regularised generative framework that combines cross-modal alignment with feature-level synthesis. Building on an InfoGAN-style architecture, our method first learns a semantically aligned audio-text representation space using pre-trained transformers and contrastive objectives. A feature generator is then trained to produce emotion-aware audio features while employing mutual information as a quantitative regulariser to ensure strong dependency between generated features and their conditioning variables. We extend this approach to multimodal settings, enabling the generation of novel, paired (audio, text) features. Comprehensive evaluation on three benchmark datasets (IEMOCAP, MSP-IMPROV, MSP-Podcast) demonstrates that our framework consistently outperforms existing augmentation methods, achieving state-of-the-art performance with improvements of up to 2.6% in unimodal SER and 3.2% in multimodal emotion recognition. Most importantly, we demonstrate that mutual information functions as both a regulariser and a measurable metric for generative quality, offering a systematic approach to data augmentation in affective computing.

---

### A-IPO: Adaptive Intent-driven Preference Optimization

Authors: Wenqing Wang, Muhammad Asif Ali, Ali Shoker, Ruohan Yang, Junyang Chen, Ying Sha, Huan Wang
Published: 2025-10-11
Categories: cs.CL
arXiv: https://arxiv.org/abs/2510.10077v3

Human preferences are diverse and dynamic, shaped by regional, cultural, and social factors. Existing alignment methods like Direct Preference Optimization (DPO) and its variants often default to majority views, overlooking minority opinions and failing to capture latent user intentions in prompts. To address these limitations, we introduce \underline{\textbf{A}}daptive \textbf{\underline{I}}ntent-driven \textbf{\underline{P}}reference \textbf{\underline{O}}ptimization (\textbf{A-IPO}). Specifically,A-IPO introduces an intention module that infers the latent intent behind each user prompt and explicitly incorporates this inferred intent into the reward function, encouraging stronger alignment between the preferred model's responses and the user's underlying intentions. We demonstrate, both theoretically and empirically, that incorporating an intention--response similarity term increases the preference margin (by a positive shift of $λ\,Δ\mathrm{sim}$ in the log-odds), resulting in clearer separation between preferred and dispreferred responses compared to DPO. For evaluation, we introduce two new benchmarks, Real-pref, Attack-pref along with an extended version of an existing dataset, GlobalOpinionQA-Ext, to assess real-world and adversarial preference alignment. Through explicit modeling of diverse user intents,A-IPO facilitates pluralistic preference optimization while simultaneously enhancing adversarial robustness in preference alignment. Comprehensive empirical evaluation demonstrates that A-IPO consistently surpasses existing baselines, yielding substantial improvements across key metrics: up to +24.8 win-rate and +45.6 Response-Intention Consistency on Real-pref; up to +38.6 Response Similarity and +52.2 Defense Success Rate on Attack-pref; and up to +54.6 Intention Consistency Score on GlobalOpinionQA-Ext.

---

### Gradient-based Model Shortcut Detection for Time Series Classification

Authors: Salomon Ibarra, Frida Cantu, Kaixiong Zhou, Li Zhang
Published: 2025-10-11
Categories: cs.LG, cs.AI
arXiv: https://arxiv.org/abs/2510.10075v2

Deep learning models have attracted lots of research attention in time series classification (TSC) task in the past two decades. Recently, deep neural networks (DNN) have surpassed classical distance-based methods and achieved state-of-the-art performance. Despite their promising performance, deep neural networks (DNNs) have been shown to rely on spurious correlations present in the training data, which can hinder generalization. For instance, a model might incorrectly associate the presence of grass with the label ``cat" if the training set have majority of cats lying in grassy backgrounds. However, the shortcut behavior of DNNs in time series remain under-explored. Most existing shortcut work are relying on external attributes such as gender, patients group, instead of focus on the internal bias behavior in time series models. In this paper, we take the first step to investigate and establish point-based shortcut learning behavior in deep learning time series classification. We further propose a simple detection method based on other class to detect shortcut occurs without relying on test data or clean training classes. We test our proposed method in UCR time series datasets.

---

### StepFly: Agentic Troubleshooting Guide Automation for Incident Diagnosis

Authors: Jiayi Mao, Liqun Li, Yanjie Gao, Zegang Peng, Shilin He, Chaoyun Zhang, Si Qin, Samia Khalid, Qingwei Lin, Saravan Rajmohan, Sitaram Lanka, Dongmei Zhang
Published: 2025-10-11
Categories: cs.AI
arXiv: https://arxiv.org/abs/2510.10074v2

Effective incident management in large-scale IT systems relies on troubleshooting guides (TSGs), but their manual execution is slow and error-prone. While recent advances in LLMs offer promise for automating incident management tasks, existing LLM-based solutions lack specialized support for several key challenges, including managing TSG quality issues, interpreting complex control flow, handling data-intensive queries, and exploiting execution parallelism. We first conducted an empirical study on 92 real-world TSGs, and, guided by our findings, we present StepFly, a novel end-to-end agentic framework for troubleshooting guide automation. Our approach features a three-stage workflow: the first stage provides a comprehensive guide together with a tool, TSG Mentor, to assist site reliability engineers (SREs) in improving TSG quality; the second stage performs offline preprocessing using LLMs to extract structured execution directed acyclic graphs (DAGs) from unstructured TSGs and to create dedicated Query Preparation Plugins (QPPs); and the third stage executes online using a DAG-guided scheduler-executor framework with a memory system to ensure correct workflow and support parallel execution of independent steps. Our empirical evaluation on a collection of real-world TSGs and incidents demonstrates that StepFly achieves a ~94% success rate on GPT-4.1, outperforming baselines with less time and token consumption. Furthermore, it achieves a remarkable execution time reduction of 32.9% to 70.4% for parallelizable TSGs. Our code and sample data are publicly available at https://github.com/microsoft/StepFly.

---

### SecureWebArena: A Holistic Security Evaluation Benchmark for LVLM-based Web Agents

Authors: Zonghao Ying, Yangguang Shao, Jianle Gan, Gan Xu, Wenxin Zhang, Quanchen Zou, Junzheng Shi, Zhenfei Yin, Mingchuan Zhang, Aishan Liu, Xianglong Liu
Published: 2025-10-11
Categories: cs.CR, cs.CV
arXiv: https://arxiv.org/abs/2510.10073v2

Large vision-language model (LVLM)-based web agents are emerging as powerful tools for automating complex online tasks. However, when deployed in real-world environments, they face serious security risks, motivating the design of security evaluation benchmarks. Existing benchmarks provide only partial coverage, typically restricted to narrow scenarios such as user-level prompt manipulation, and thus fail to capture the broad range of agent vulnerabilities. To address this gap, we present \tool{}, the first holistic benchmark for evaluating the security of LVLM-based web agents. \tool{} first introduces a unified evaluation suite comprising six simulated but realistic web environments (\eg, e-commerce platforms, community forums) and includes 2,970 high-quality trajectories spanning diverse tasks and attack settings. The suite defines a structured taxonomy of six attack vectors spanning both user-level and environment-level manipulations. In addition, we introduce a multi-layered evaluation protocol that analyzes agent failures across three critical dimensions: internal reasoning, behavioral trajectory, and task outcome, facilitating a fine-grained risk analysis that goes far beyond simple success metrics. Using this benchmark, we conduct large-scale experiments on 9 representative LVLMs, which fall into three categories: general-purpose, agent-specialized, and GUI-grounded. Our results show that all tested agents are consistently vulnerable to subtle adversarial manipulations and reveal critical trade-offs between model specialization and security. By providing (1) a comprehensive benchmark suite with diverse environments and a multi-layered evaluation pipeline, and (2) empirical insights into the security challenges of modern LVLM-based web agents, \tool{} establishes a foundation for advancing trustworthy web agent deployment.

---

### Unilaw-R1: A Large Language Model for Legal Reasoning with Reinforcement Learning and Iterative Inference

Authors: Hua Cai, Shuang Zhao, Liang Zhang, Xuli Shen, Qing Xu, Weilin Shen, Zihao Wen, Tianke Ban
Published: 2025-10-11
Categories: cs.CL
arXiv: https://arxiv.org/abs/2510.10072v2

Reasoning-focused large language models (LLMs) are rapidly evolving across various domains, yet their capabilities in handling complex legal problems remains underexplored. In this paper, we introduce Unilaw-R1, a large language model tailored for legal reasoning. With a lightweight 7-billion parameter scale, Unilaw-R1 significantly reduces deployment cost while effectively tackling three core challenges in the legal domain: insufficient legal knowledge, unreliable reasoning logic, and weak business generalization. To address these issues, we first construct Unilaw-R1-Data, a high-quality dataset containing 17K distilled and screened chain-of-thought (CoT) samples. Based on this, we adopt a two-stage training strategy combining Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL), which significantly boosts the performance on complex legal reasoning tasks and supports interpretable decision-making in legal AI applications. To assess legal reasoning ability, we also introduce Unilaw-R1-Eval, a dedicated benchmark designed to evaluate models across single- and multi-choice legal tasks. Unilaw-R1 demonstrates strong results on authoritative benchmarks, outperforming all models of similar scale and achieving performance on par with the much larger DeepSeek-R1-Distill-Qwen-32B (54.9%). Following domain-specific training, it also showed significant gains on LawBench and LexEval, exceeding Qwen-2.5-7B-Instruct (46.6%) by an average margin of 6.6%.

---

### ADEPT: Continual Pretraining via Adaptive Expansion and Dynamic Decoupled Tuning

Authors: Jinyang Zhang, Yue Fang, Hongxin Ding, Weibin Liao, Muyang Ye, Xu Chu, Junfeng Zhao, Yasha Wang
Published: 2025-10-11
Categories: cs.LG
arXiv: https://arxiv.org/abs/2510.10071v1

Conventional continual pretraining (CPT) for large language model (LLM) domain adaptation often suffers from catastrophic forgetting and limited domain capacity. Existing strategies adopt layer expansion, introducing additional trainable parameters to accommodate new knowledge. However, the uniform expansion and updates still entangle general and domain learning, undermining its effectiveness. Our pilot studies reveal that LLMs exhibit functional specialization, where layers and units differentially encode general-critical capabilities, suggesting that parameter expansion and optimization should be function-aware. We then propose ADEPT, Adaptive Expansion and Dynamic Decoupled Tuning for continual pretraining, a two-stage framework for domain-adaptive CPT. ADEPT first performs General-Competence Guided Selective Layer Expansion, duplicating layers least critical for the general domain to increase representational capacity while minimizing interference with general knowledge. It then applies Adaptive Unit-Wise Decoupled Tuning, disentangling parameter units within expanded layers according to their general-domain importance and assigning asymmetric learning rates to balance knowledge injection and retention. Experiments on mathematical and medical benchmarks show that ADEPT outperforms full-parameter CPT by up to 5.76% on the general domain and 5.58% on the target domain with only 15% of parameters tuned and less than 50% training time. Ablation studies, theoretical analysis, and extended investigations further demonstrate the necessity of targeted expansion and decoupled optimization, providing new principles for efficient and robust domain-adaptive CPT. Our code is open-sourced at https://github.com/PuppyKnightUniversity/ADEPT

---

### SyncLipMAE: Contrastive Masked Pretraining for Audio-Visual Talking-Face Representation

Authors: Zeyu Ling, Xiaodong Gu, Jiangnan Tang, Changqing Zou
Published: 2025-10-11
Categories: cs.AI, cs.MM
arXiv: https://arxiv.org/abs/2510.10069v2

We introduce SyncLipMAE, a self-supervised pretraining framework for talking-face video that learns synchronization-aware and transferable facial dynamics from unlabeled audio-visual streams. Our approach couples masked visual modeling with cross-modal contrastive alignment and employs three per-frame prompt tokens that explicitly encode the essential factors of a talking-face frame - identity, vocal motion (speech-synchronized facial dynamics), and ambient motion (audio-agnostic movements such as blinks and head pose). The contrastive objective uses time-aligned vocal-motion and audio tokens as positives and misaligned pairs as negatives, driving both modalities into a shared embedding space and yielding token-level audio-visual stream synchronization. After pretraining, the aligned audio tokens together with the visual prompt tokens (identity, vocal motion, ambient motion) form a unified interface for four disparate downstream settings: (i) audio-visual stream synchronization; (ii) facial emotion and head/face action recognition; (iii) visual speech recognition; and (iv) visual dubbing, for which we enable indistinguishable audio- or video-driven control within a single model. Across four task families that require distinct capabilities, SyncLipMAE achieves state-of-the-art results, underscoring the effectiveness of synchronization-aware, factorized self-supervised pretraining.

---

### Probabilistic Hyper-Graphs using Multiple Randomly Masked Autoencoders for Semi-supervised Multi-modal Multi-task Learning

Authors: Pîrvu Mihai-Cristian, Marius Leordeanu
Published: 2025-10-11
Categories: cs.CV
arXiv: https://arxiv.org/abs/2510.10068v2

The computer vision domain has greatly benefited from an abundance of data across many modalities to improve on various visual tasks. Recently, there has been a lot of focus on self-supervised pre-training methods through Masked Autoencoders (MAE) \cite{he2022masked,bachmann2022multimae}, usually used as a first step before optimizing for a downstream task, such as classification or regression. This is very useful as it doesn't require any manually labeled data. In this work, we introduce Probabilistic Hyper-Graphs using Masked Autoencoders (PHG-MAE): a novel model that unifies the classical work on neural graphs \cite{leordeanu2021semi} with the modern approach of masked autoencoders under a common theoretical framework. Through random masking of entire modalities, not just patches, the model samples from the distribution of hyper-edges on each forward pass. Additionally, the model adapts the standard MAE algorithm by combining pre-training and fine-tuning into a single training loop. Moreover, our approach enables the creation of inference-time ensembles which, through aggregation, boost the final prediction performance and consistency. Lastly, we show that we can apply knowledge distillation on top of the ensembles with little loss in performance, even with models that have fewer than 1M parameters. While our work mostly focuses on outdoor UAV scenes that contain multiple world interpretations and modalities, the same steps can be followed in other similar domains, such as autonomous driving or indoor robotics. In order to streamline the process of integrating external pre-trained experts for computer vision multi-modal multi-task learning (MTL) scenarios, we developed a data-pipeline software. Using this tool, we have created and released a fully-automated extension of the Dronescapes dataset. All the technical details, code and reproduction steps are publicly released.

---

### OBsmith: LLM-Powered JavaScript Obfuscator Testing

Authors: Shan Jiang, Chenguang Zhu, Sarfraz Khurshid
Published: 2025-10-11
Categories: cs.SE, cs.AI, cs.PL
arXiv: https://arxiv.org/abs/2510.10066v2

JavaScript obfuscators are widely deployed to protect intellectual property and resist reverse engineering, yet their correctness has been largely overlooked compared to performance and resilience. Existing evaluations typically measure resistance to deobfuscation, leaving the critical question of whether obfuscators preserve program semantics unanswered. Incorrect transformations can silently alter functionality, compromise reliability, and erode security-undermining the very purpose of obfuscation. To address this gap, we present OBsmith, a novel framework to systematically test JavaScript obfuscators using large language models (LLMs). OBsmith leverages LLMs to generate program sketches abstract templates capturing diverse language constructs, idioms, and corner cases-which are instantiated into executable programs and subjected to obfuscation under different configurations. Besides LLM-powered sketching, OBsmith also employs a second source: automatic extraction of sketches from real programs. This extraction path enables more focused testing of project specific features and lets developers inject domain knowledge into the resulting test cases. OBsmith uncovers 11 previously unknown correctness bugs. Under an equal program budget, five general purpose state-of-the-art JavaScript fuzzers (FuzzJIT, Jsfunfuzz, Superion, DIE, Fuzzilli) failed to detect these issues, highlighting OBsmith's complementary focus on obfuscation induced misbehavior. An ablation shows that all components except our generic MRs contribute to at least one bug class; the negative MR result suggests the need for obfuscator-specific metamorphic relations. Our results also seed discussion on how to balance obfuscation presets and performance cost. We envision OBsmith as an important step towards automated testing and quality assurance of obfuscators and other semantic-preserving toolchains.

---

### CLMN: Concept based Language Models via Neural Symbolic Reasoning

Authors: Yibo Yang
Published: 2025-10-11
Categories: cs.CL, cs.AI
arXiv: https://arxiv.org/abs/2510.10063v2

Deep learning has advanced NLP, but interpretability remains limited, especially in healthcare and finance. Concept bottleneck models tie predictions to human concepts in vision, but NLP versions either use binary activations that harm text representations or latent concepts that weaken semantics, and they rarely model dynamic concept interactions such as negation and context. We introduce the Concept Language Model Network (CLMN), a neural-symbolic framework that keeps both performance and interpretability. CLMN represents concepts as continuous, human-readable embeddings and applies fuzzy-logic reasoning to learn adaptive interaction rules that state how concepts affect each other and the final decision. The model augments original text features with concept-aware representations and automatically induces interpretable logic rules. Across multiple datasets and pre-trained language models, CLMN achieves higher accuracy than existing concept-based methods while improving explanation quality. These results show that integrating neural representations with symbolic reasoning in a unified concept space can yield practical, transparent NLP systems.

---

### HUME: Measuring the Human-Model Performance Gap in Text Embedding Tasks

Authors: Adnan El Assadi, Isaac Chung, Roman Solomatin, Niklas Muennighoff, Kenneth Enevoldsen
Published: 2025-10-11
Categories: cs.CL
arXiv: https://arxiv.org/abs/2510.10062v3

Comparing human and model performance offers a valuable perspective for understanding the strengths and limitations of embedding models, highlighting where they succeed and where they fail to capture meaning and nuance. However, such comparisons are rarely made, as human performance on embedding tasks is difficult to measure. To fill this gap, we introduce HUME: Human Evaluation Framework for Text Embeddings. While frameworks like MTEB provide broad model evaluation, they lack reliable estimates of human performance, limiting the interpretability of model scores. We measure human performance across 16 MTEB datasets spanning reranking, classification, clustering, and semantic textual similarity across linguistically diverse high- and low-resource languages. Humans achieve an average performance of 77.6% compared to 80.1% for the best embedding model, though with substantial variation: models reach high performance on some datasets while struggling on notably low-resource languages. Our human annotations also reveal multiple dataset issues. We additionally benchmark nine LLMs as annotators on reranking, classification, and STS tasks, finding that they fall short of human performance (76.1% vs. 81.2%) despite offering scalability advantages. We provide human performance baselines, insights into task difficulty patterns, and an extensible evaluation framework that enables a more meaningful interpretation of results and informs the development of both models and benchmarks. Our code, dataset, and leaderboard are publicly available at https://github.com/embeddings-benchmark/mteb.

---

### Translution: Unifying Self-attention and Convolution for Adaptive and Relative Modeling

Authors: Hehe Fan, Yi Yang, Mohan Kankanhalli, Fei Wu
Published: 2025-10-11
Categories: cs.LG, cs.AI, cs.CL, cs.CV
arXiv: https://arxiv.org/abs/2510.10060v1

When modeling a given type of data, we consider it to involve two key aspects: 1) identifying relevant elements (e.g., image pixels or textual words) to a central element, as in a convolutional receptive field, or to a query element, as in self-attention, and 2) encoding these tokens effectively. Self-attention can adaptively identify these elements but relies on absolute positional embedding for structural representation learning. In contrast, convolution encodes elements in a relative manner, yet their fixed kernel size limits their ability to adaptively select the relevant elements. In this paper, we introduce Translution, an operation that unifies the adaptive identification capability of self-attention and the relative encoding advantage of convolution. However, this integration leads to a substantial increase in the number of parameters, exceeding most currently available computational resources. Therefore, we propose a lightweight variant of Translution, named α-Translution. Experiments on computer vision and natural language processing tasks show that Translution (including α-Translution) achieves superior accuracy compared to self-attention. The code is available at https://github.com/hehefan/Translution.

---

### One4Many-StablePacker: An Efficient Deep Reinforcement Learning Framework for the 3D Bin Packing Problem

Authors: Lei Gao, Shihong Huang, Shengjie Wang, Hong Ma, Feng Zhang, Hengda Bao, Qichang Chen, Weihua Zhou
Published: 2025-10-11
Categories: cs.LG
arXiv: https://arxiv.org/abs/2510.10057v2

The three-dimensional bin packing problem (3D-BPP) is widely applied in logistics and warehousing. Existing learning-based approaches often neglect practical stability-related constraints and exhibit limitations in generalizing across diverse bin dimensions. To address these limitations, we propose a novel deep reinforcement learning framework, One4Many-StablePacker (O4M-SP). The primary advantage of O4M-SP is its ability to handle various bin dimensions in a single training process while incorporating support and weight constraints common in practice. Our training method introduces two innovative mechanisms. First, it employs a weighted reward function that integrates loading rate and a new height difference metric for packing layouts, promoting improved bin utilization through flatter packing configurations. Second, it combines clipped policy gradient optimization with a tailored policy drifting method to mitigate policy entropy collapse, encouraging exploration at critical decision nodes during packing to avoid suboptimal solutions. Extensive experiments demonstrate that O4M-SP generalizes successfully across diverse bin dimensions and significantly outperforms baseline methods. Furthermore, O4M-SP exhibits strong practical applicability by effectively addressing packing scenarios with stability constraints.

---

### Incomplete Multi-Label Image Recognition by Co-learning Semantic-Aware Features and Label Recovery

Authors: Zhi-Fen He, Ren-Dong Xie, Bo Li, Bin Liu, Jin-Yan Hu
Published: 2025-10-11
Categories: cs.CV
arXiv: https://arxiv.org/abs/2510.10055v2

Multi-label image recognition with incomplete labels is a challenging yet vital task in computer vision, which faces two fundamental challenges: learning semantic-aware features and recovering missing labels. In this paper, we propose a Co-learning framework for Semantic-aware features and Label recovery (CSL), designed to address both challenges in a unified learning paradigm. Specifically, we develop a semantic-related feature learning module that captures robust semantic-related representations by discovering semantic information and label correlations. Furthermore, a semantic-guided feature enhancement module is introduced to generate highly discriminative semantic-aware features by effectively aligning visual and semantic spaces. Finally, we present a collaborative learning framework that integrates semantic-aware feature learning with label recovery. This framework not only dynamically enhances the discriminability of semantic-aware features but also adaptively infers and recovers missing labels, thereby forming a mutually reinforcing mechanism between the two processes. Extensive experiments on three widely used public datasets (MS-COCO, VOC2007, and NUS-WIDE) demonstrate that CSL outperforms state-of-the-art methods for incomplete multi-label image recognition.

---

### DREAM: A Benchmark Study for Deepfake photoREalism AssessMent

Authors: Bo Peng, Zichuan Wang, Sheng Yu, Xiaochuan Jin, Wei Wang, Jing Dong
Published: 2025-10-11
Categories: cs.CV
arXiv: https://arxiv.org/abs/2510.10053v2

Deep learning based face-swap videos, widely known as deepfakes, have drawn wide attention due to their threat to information credibility. Recent works mainly focus on the problem of deepfake detection that aims to reliably tell deepfakes apart from real ones, in an objective way. On the other hand, the subjective perception of deepfakes, especially its computational modeling and imitation, is also a significant problem but lacks adequate study. In this paper, we focus on the photorealism assessment of deepfakes, which is defined as the automatic assessment of deepfake photorealism that approximates human perception of deepfakes. It is important for evaluating the quality and deceptiveness of deepfakes which can be used for predicting the influence of deepfakes on Internet, and it also has potentials in improving the deepfake generation process by serving as a critic. This paper promotes this new direction by presenting a comprehensive benchmark called DREAM, which stands for Deepfake photoREalism AssessMent. It is comprised of a deepfake video dataset of diverse quality, a large scale annotation that includes 140,000 photorealism scores and textual descriptions obtained from 3,500 human annotators, and a comprehensive evaluation and analysis of 18 representative photorealism assessment methods, including recent large vision language model based methods and a newly proposed description-aligned CLIP method. The benchmark and insights included in this study can lay the foundation for future research in this direction and other related areas. We make the dataset available to the research community at https://github.com/bomb2peng/DREAM-A-Benchmark-Study-for-Deepfake-photoREalism-AssessMent.

---

### Think Twice to See More: Iterative Visual Reasoning in Medical VLMs

Authors: Kaitao Chen, Shaohao Rui, Yankai Jiang, Jiamin Wu, Qihao Zheng, Chunfeng Song, Xiaosong Wang, Mu Zhou, Mianxin Liu
Published: 2025-10-11
Categories: cs.CV, cs.AI
arXiv: https://arxiv.org/abs/2510.10052v1

Medical vision-language models (VLMs) excel at image-text understanding but typically rely on a single-pass reasoning that neglects localized visual cues. In clinical practice, however, human experts iteratively scan, focus, and refine the regions of interest before reaching a final diagnosis. To narrow this machine-human perception gap, we introduce ViTAR, a novel VLM framework that emulates the iterative reasoning process of human experts through a cognitive chain of "think-act-rethink-answer". ViTAR treats medical images as interactive objects, enabling models to engage multi-step visual reasoning. To support this approach, we curate a high-quality instruction dataset comprising 1K interactive examples that encode expert-like diagnostic behaviors. In addition, a 16K visual question answering training data has been curated towards fine-grained visual diagnosis. We introduce a two-stage training strategy that begins with supervised fine-tuning to guide cognitive trajectories, followed by the reinforcement learning to optimize decision-making. Extensive evaluations demonstrate that ViTAR outperforms strong state-of-the-art models. Visual attention analysis reveals that from the "think" to "rethink" rounds, ViTAR increasingly anchors visual grounding to clinically critical regions and maintains high attention allocation to visual tokens during reasoning, providing mechanistic insight into its improved performance. These findings demonstrate that embedding expert-style iterative thinking chains into VLMs enhances both performance and trustworthiness of medical AI.

---