Zijian Chen

Zicheng Zhang, Ziheng Jia, Haoning Wu et al.

With the rising interest in research on Large Multi-modal Models (LMMs) for video understanding, many studies have emphasized general video comprehension capabilities, neglecting the systematic exploration into video quality understanding. To address this oversight, we introduce Q-Bench-Video in this paper, a new benchmark specifically designed to evaluate LMMs' proficiency in discerning video quality. a) To ensure video source diversity, Q-Bench-Video encompasses videos from natural scenes, AI-generated Content (AIGC), and Computer Graphics (CG). b) Building on the traditional multiple-choice questions format with the Yes-or-No and What-How categories, we include Open-ended questions to better evaluate complex scenarios. Additionally, we incorporate the video pair quality comparison question to enhance comprehensiveness. c) Beyond the traditional Technical, Aesthetic, and Temporal distortions, we have expanded our evaluation aspects to include the dimension of AIGC distortions, which addresses the increasing demand for video generation. Finally, we collect a total of 2,378 question-answer pairs and test them on 12 open-source & 5 proprietary LMMs. Our findings indicate that while LMMs have a foundational understanding of video quality, their performance remains incomplete and imprecise, with a notable discrepancy compared to human performance. Through Q-Bench-Video, we seek to catalyze community interest, stimulate further research, and unlock the untapped potential of LMMs to close the gap in video quality understanding.

Wei Sun, Weixia Zhang, Yuqin Cao et al.

UHD images, typically with resolutions equal to or higher than 4K, pose a significant challenge for efficient image quality assessment (IQA) algorithms, as adopting full-resolution images as inputs leads to overwhelming computational complexity and commonly used pre-processing methods like resizing or cropping may cause substantial loss of detail. To address this problem, we design a multi-branch deep neural network (DNN) to assess the quality of UHD images from three perspectives: global aesthetic characteristics, local technical distortions, and salient content perception. Specifically, aesthetic features are extracted from low-resolution images downsampled from the UHD ones, which lose high-frequency texture information but still preserve the global aesthetics characteristics. Technical distortions are measured using a fragment image composed of mini-patches cropped from UHD images based on the grid mini-patch sampling strategy. The salient content of UHD images is detected and cropped to extract quality-aware features from the salient regions. We adopt the Swin Transformer Tiny as the backbone networks to extract features from these three perspectives. The extracted features are concatenated and regressed into quality scores by a two-layer multi-layer perceptron (MLP) network. We employ the mean square error (MSE) loss to optimize prediction accuracy and the fidelity loss to optimize prediction monotonicity. Experimental results show that the proposed model achieves the best performance on the UHD-IQA dataset while maintaining the lowest computational complexity, demonstrating its effectiveness and efficiency. Moreover, the proposed model won first prize in ECCV AIM 2024 UHD-IQA Challenge. The code is available at https://github.com/sunwei925/UIQA.

Zhichao Zhang, Wei Sun, Xinyue Li et al.

In recent years, artificial intelligence (AI)-driven video generation has gained significant attention. Consequently, there is a growing need for accurate video quality assessment (VQA) metrics to evaluate the perceptual quality of AI-generated content (AIGC) videos and optimize video generation models. However, assessing the quality of AIGC videos remains a significant challenge because these videos often exhibit highly complex distortions, such as unnatural actions and irrational objects. To address this challenge, we systematically investigate the AIGC-VQA problem, considering both subjective and objective quality assessment perspectives. For the subjective perspective, we construct the Large-scale Generated Video Quality assessment (LGVQ) dataset, consisting of 2,808 AIGC videos generated by 6 video generation models using 468 carefully curated text prompts. We evaluate the perceptual quality of AIGC videos from three critical dimensions: spatial quality, temporal quality, and text-video alignment. For the objective perspective, we establish a benchmark for evaluating existing quality assessment metrics on the LGVQ dataset. Our findings show that current metrics perform poorly on this dataset, highlighting a gap in effective evaluation tools. To bridge this gap, we propose the Unify Generated Video Quality assessment (UGVQ) model, designed to accurately evaluate the multi-dimensional quality of AIGC videos. The UGVQ model integrates the visual and motion features of videos with the textual features of their corresponding prompts, forming a unified quality-aware feature representation tailored to AIGC videos. Experimental results demonstrate that UGVQ achieves state-of-the-art performance on the LGVQ dataset across all three quality dimensions. Both the LGVQ dataset and the UGVQ model are publicly available on https://github.com/zczhang-sjtu/UGVQ.git.

Kaiyuan Ji, Yixuan Gao, Lu Sun et al.

Advertising images significantly impact commercial conversion rates and brand equity, yet current evaluation methods rely on subjective judgments, lacking scalability, standardized criteria, and interpretability. To address these challenges, we present A^3 (Advertising Aesthetic Assessment), a comprehensive framework encompassing four components: a paradigm (A^3-Law), a dataset (A^3-Dataset), a multimodal large language model (A^3-Align), and a benchmark (A^3-Bench). Central to A^3 is a theory-driven paradigm, A^3-Law, comprising three hierarchical stages: (1) Perceptual Attention, evaluating perceptual image signals for their ability to attract attention; (2) Formal Interest, assessing formal composition of image color and spatial layout in evoking interest; and (3) Desire Impact, measuring desire evocation from images and their persuasive impact. Building on A^3-Law, we construct A^3-Dataset with 120K instruction-response pairs from 30K advertising images, each richly annotated with multi-dimensional labels and Chain-of-Thought (CoT) rationales. We further develop A^3-Align, trained under A^3-Law with CoT-guided learning on A^3-Dataset. Extensive experiments on A^3-Bench demonstrate that A^3-Align achieves superior alignment with A^3-Law compared to existing models, and this alignment generalizes well to quality advertisement selection and prescriptive advertisement critique, indicating its potential for broader deployment. Dataset, code, and models can be found at: https://github.com/euleryuan/A3-Align.

Zijian Chen, Wei Sun, Jun Jia et al.

Banding, also known as staircase-like contours, frequently occurs in flat areas of images/videos processed by the compression or quantization algorithms. As undesirable artifacts, banding destroys the original image structure, thus degrading users' quality of experience (QoE). In this paper, we systematically investigate the banding image quality assessment (IQA) problem, aiming to detect the image banding artifacts and evaluate their perceptual visual quality. Considering that the existing image banding databases only contain limited content sources and banding generation methods, and lack perceptual quality labels (i.e. mean opinion scores), we first build the largest banding IQA database so far, named Banding Artifact Noticeable Database (BAND-2k), which consists of 2,000 banding images generated by 15 compression and quantization schemes. A total of 23 workers participated in the subjective IQA experiment, yielding over 214,000 patch-level banding class labels and 44,371 reliable image-level quality ratings. Subsequently, we develop an effective no-reference (NR) banding evaluator for banding detection and quality assessment by leveraging frequency characteristics of banding artifacts. A dual convolutional neural network is employed to concurrently learn the feature representation from the high-frequency and low-frequency maps, thereby enhancing the ability to discern banding artifacts. The quality score of a banding image is generated by pooling the banding detection maps masked by the spatial frequency filters. Experiments demonstrate that our banding evaluator achieves a remarkably high accuracy in banding detection and also exhibits high SRCC and PLCC results with the perceptual quality labels. These findings unveil the strong correlations between the intensity of banding artifacts and the perceptual visual quality, thus validating the necessity of banding quality assessment.

Linhan Cao, Wei Sun, Xiongkuo Min et al.

Just noticeable distortion (JND), representing the threshold of distortion in an image that is minimally perceptible to the human visual system (HVS), is crucial for image compression algorithms to achieve a trade-off between transmission bit rate and image quality. However, traditional JND prediction methods only rely on pixel-level or sub-band level features, lacking the ability to capture the impact of image content on JND. To bridge this gap, we propose a Semantic-Guided JND (SG-JND) network to leverage semantic information for JND prediction. In particular, SG-JND consists of three essential modules: the image preprocessing module extracts semantic-level patches from images, the feature extraction module extracts multi-layer features by utilizing the cross-scale attention layers, and the JND prediction module regresses the extracted features into the final JND value. Experimental results show that SG-JND achieves the state-of-the-art performance on two publicly available JND datasets, which demonstrates the effectiveness of SG-JND and highlight the significance of incorporating semantic information in JND assessment.

Zijian Chen, Wei Sun, Zicheng Zhang et al.

Banding artifact, as known as staircase-like contour, is a common quality annoyance that happens in compression, transmission, etc. scenarios, which largely affects the user's quality of experience (QoE). The banding distortion typically appears as relatively small pixel-wise variations in smooth backgrounds, which is difficult to analyze in the spatial domain but easily reflected in the frequency domain. In this paper, we thereby study the banding artifact from the frequency aspect and propose a no-reference banding detection model to capture and evaluate banding artifacts, called the Frequency-Sensitive BANding Detector (FS-BAND). The proposed detector is able to generate a pixel-wise banding map with a perception correlated quality score. Experimental results show that the proposed FS-BAND method outperforms state-of-the-art image quality assessment (IQA) approaches with higher accuracy in banding classification task.

Bohan Liang, Zijian Chen, Qi Jia et al.

Stock prediction, a subject closely related to people's investment activities in fully dynamic and live environments, has been widely studied. Current large language models (LLMs) have shown remarkable potential in various domains, exhibiting expert-level performance through advanced reasoning and contextual understanding. In this paper, we introduce PriceSeer, a live, dynamic, and data-uncontaminated benchmark specifically designed for LLMs performing stock prediction tasks. Specifically, PriceSeer includes 110 U.S. stocks from 11 industrial sectors, with each containing 249 historical data points. Our benchmark implements both internal and external information expansion, where LLMs receive extra financial indicators, news, and fake news to perform stock price prediction. We evaluate six cutting-edge LLMs under different prediction horizons, demonstrating their potential in generating investment strategies after obtaining accurate price predictions for different sectors. Additionally, we provide analyses of LLMs' suboptimal performance in long-term predictions, including the vulnerability to fake news and specific industries. The code and evaluation data will be open-sourced at https://github.com/BobLiang2113/PriceSeer.

Yuting Zheng, Zijian Chen, Qi Jia

Unraveling the hierarchical structure-property relationships is the central challenge of materials science, necessitating the interpretation of data across vast physical scales from micro to macro. Despite the rapid integration of Large Multimodal Models (LMMs) into scientific workflows, existing scientific benchmarks primarily focus on general chart interpretation or isolated common-sense reasoning, failing to capture reasoning ability across intricate physical dimensions. To address this, we introduce CSMBench, a dataset comprising 1,041 high-quality figures curated from premier journals up to September 2025. CSMBench categorizes data into four scientifically distinct regimes: atomic, micro, meso, and macro scales, strictly aligning with the focus and definitions in materials study. Through open-ended figure description and multiple-choice caption matching tasks, we evaluate state-of-the-art open-source and closed-source models. Our analysis identifies that performance varies significantly across physical scales due to the distinct visual characteristics, highlighting the limitations of current generalist models and identifying critical directions for achieving hierarchical and accurate understanding in materials research. The CSMBench is publicly released at: https://huggingface.co/datasets/lututu/CSMBench.

Zijian Chen, Maria Varkanitsa, Prakash Ishwar et al.

We propose a lesion-aware graph neural network (LEGNet) to predict language ability from resting-state fMRI (rs-fMRI) connectivity in patients with post-stroke aphasia. Our model integrates three components: an edge-based learning module that encodes functional connectivity between brain regions, a lesion encoding module, and a subgraph learning module that leverages functional similarities for prediction. We use synthetic data derived from the Human Connectome Project (HCP) for hyperparameter tuning and model pretraining. We then evaluate the performance using repeated 10-fold cross-validation on an in-house neuroimaging dataset of post-stroke aphasia. Our results demonstrate that LEGNet outperforms baseline deep learning methods in predicting language ability. LEGNet also exhibits superior generalization ability when tested on a second in-house dataset that was acquired under a slightly different neuroimaging protocol. Taken together, the results of this study highlight the potential of LEGNet in effectively learning the relationships between rs-fMRI connectivity and language ability in a patient cohort with brain lesions for improved post-stroke aphasia evaluation.

Shaojie Shi, Zhengyu Shi, Lingran Zheng et al.

Causal inference in social science relies on end-to-end, intervention-centered research-design reasoning grounded in real-world policy interventions, but current benchmarks fail to evaluate this capability of large language models (LLMs). We present InterveneBench, a benchmark designed to assess such reasoning in realistic social settings. Each instance in InterveneBench is derived from an empirical social science study and requires models to reason about policy interventions and identification assumptions without access to predefined causal graphs or structural equations. InterveneBench comprises 744 peer-reviewed studies across diverse policy domains. Experimental results show that state-of-the-art LLMs struggle under this setting. To address this limitation, we further propose a multi-agent framework, STRIDES. It achieves significant performance improvements over state-of-the-art reasoning models. Our code and data are available at https://github.com/Sii-yuning/STRIDES.

Xianfeng Wang, Kaiwei Zhang, Qi Jia et al.

While Multimodal Large Language Models (MLLMs) have demonstrated impressive proficiency in high-level reasoning tasks, such as complex diagrammatic interpretation, it remains an open question whether they possess the fundamental visual primitives comparable to human intuition. To investigate this, we introduce KidVis, a novel benchmark grounded in the theory of human visual development. KidVis deconstructs visual intelligence into six atomic capabilities - Concentration, Tracking, Discrimination, Memory, Spatial, and Closure - already possessed by 6-7 year old children, comprising 10 categories of low-semantic-dependent visual tasks. Evaluating 20 state-of-the-art MLLMs against a human physiological baseline reveals a stark performance disparity. Results indicate that while human children achieve a near-perfect average score of 95.32, the state-of-the-art GPT-5 attains only 67.33. Crucially, we observe a "Scaling Law Paradox": simply increasing model parameters fails to yield linear improvements in these foundational visual capabilities. This study confirms that current MLLMs, despite their reasoning prowess, lack the essential physiological perceptual primitives required for generalized visual intelligence.

Yushuo Zheng, Zicheng Zhang, Huiyu Duan et al.

Geoscience intelligence is expected to understand, reason about, and predict earth system changes to support human decision-making in critical domains such as disaster response, climate adaptation and environmental protection. Although current research has shown promising progress on specific geoscience tasks, such as remote sensing interpretation, geographic question-answering, existing benchmarks remain largely task-specific which failing to capture the open-ended real world geoscience problems. As a result, it remains unclear how far current AI systems are from achieving genuine geoscience intelligence. To address this gap, we present \textbf{GeoR-Bench}, a \underline{Bench}mark for evaluating \underline{Geo}science visual \underline{R}easoning through reasoning informed visual editing tasks. GeoR-Bench contains 440 curated samples spanning 6 geoscience categories and 24 task types, covering earth observation imagery and structured scientific representations such as maps and diagrams. We evaluate outputs along three dimensions, including reasoning, consistency, and quality. Benchmark results of 21 closed- and open-source multimodal models reveal that geoscience reasoning remains a critical bottleneck. The highest-performing model achieves 42.7\% overall strict accuracy, while the best open-source models only get 10.3\%. Notably, the visual consistency and image quality of the outputs frequently surpass their scientific accuracy. Ultimately, these findings indicate that current models generate superficially plausible results but fail to capture underlying earth science processes.

Zijian Chen, Xueguang Ma, Shengyao Zhuang et al.

Deep Research agents are rapidly emerging as primary consumers of modern retrieval systems. Unlike human users who issue and refine queries without documenting their intermediate thought processes, Deep Research agents generate explicit natural language reasoning before each search call, revealing rich intent and contextual information that existing retrievers entirely ignore. To exploit this overlooked signal, we introduce: (1) Reasoning-Aware Retrieval, a retrieval paradigm that jointly embeds the agent's reasoning trace alongside its query; and (2) DR-Synth, a data synthesis method that generates Deep Research retriever training data from standard QA datasets. We demonstrate that both components are independently effective, and their combination yields a trained embedding model, AgentIR-4B, with substantial gains. On the challenging BrowseComp-Plus benchmark, AgentIR-4B achieves 68\% accuracy with the open-weight agent Tongyi-DeepResearch, compared to 50\% with conventional embedding models twice its size, and 37\% with BM25. Code and data are available at: https://texttron.github.io/AgentIR/.

Xiaoxiao Wang, Chunxiao Li, Junying Wang et al.

As comprehensive large model evaluation becomes prohibitively expensive, predicting model performance from limited observations has become essential. However, existing statistical methods struggle with pattern shifts, data sparsity, and lack of explanation, while pure LLM methods remain unreliable. We propose STAR, a framework that bridges data-driven STatistical expectations with knowledge-driven Agentic Reasoning. STAR leverages specialized retrievers to gather external knowledge and embeds semantic features into Constrained Probabilistic Matrix Factorization (CPMF) to generate statistical expectations with uncertainty. A reasoning module guided by Expectation Violation Theory (EVT) then refines predictions through intra-family analysis, cross-model comparison, and credibility-aware aggregation, producing adjustments with traceable explanations. Extensive experiments show that STAR consistently outperforms all baselines on both score-based and rank-based metrics, delivering a 14.46% gain in total score over the strongest statistical method under extreme sparsity, with only 1--2 observed scores per test model.

Lin Yao, Jianfei Song, Ruizhuo Xu et al.

Historically lower-level tasks such as automatic speech recognition (ASR) and speaker identification are the main focus in the speech field. Interest has been growing in higher-level spoken language understanding (SLU) tasks recently, like sentiment analysis (SA). However, improving performances on SLU tasks remains a big challenge. Basically, there are two main methods for SLU tasks: (1) Two-stage method, which uses a speech model to transfer speech to text, then uses a language model to get the results of downstream tasks; (2) One-stage method, which just fine-tunes a pre-trained speech model to fit in the downstream tasks. The first method loses emotional cues such as intonation, and causes recognition errors during ASR process, and the second one lacks necessary language knowledge. In this paper, we propose the Wave BERT (WaBERT), a novel end-to-end model combining the speech model and the language model for SLU tasks. WaBERT is based on the pre-trained speech and language model, hence training from scratch is not needed. We also set most parameters of WaBERT frozen during training. By introducing WaBERT, audio-specific information and language knowledge are integrated in the short-time and low-resource training process to improve results on the dev dataset of SLUE SA tasks by 1.15% of recall score and 0.82% of F1 score. Additionally, we modify the serial Continuous Integrate-and-Fire (CIF) mechanism to achieve the monotonic alignment between the speech and text modalities.

Zijian Chen, Xueguang Ma, Shengyao Zhuang et al.

Deep-Research agents, which integrate large language models (LLMs) with search tools, have shown success in improving the effectiveness of handling complex queries that require iterative search planning and reasoning over search results. Evaluations on current benchmarks like BrowseComp relies on black-box live web search APIs, have notable limitations in (1) fairness: dynamic and opaque web APIs hinder fair comparisons and reproducibility of deep research methods; (2) transparency: lack of control over the document corpus makes it difficult to isolate retriever contributions. In other words, the current evaluations may compare a complete deep research system at a given time, but they do not foster well-controlled experiments to provide insights into the capability of underlying deep research LLMs. To address these challenges, we introduce BrowseComp-Plus, a benchmark derived from BrowseComp, employing a fixed, carefully curated corpus. Each query in BrowseComp-Plus includes human-verified supporting documents and mined challenging negatives, enabling controlled experimentation. The benchmark is shown to be effective in distinguishing the performance of deep research systems. For instance, the open-source model Search-R1, when paired with the BM25 retriever, achieves 3.86% accuracy, whereas the GPT-5 achieves 55.9%. Integrating the GPT-5 with the Qwen3-Embedding-8B retriever further enhances its accuracy to 70.1% with fewer search calls. This benchmark allows comprehensive evaluation and disentangled analysis of deep research agents and retrieval methods, fostering insights into retrieval effectiveness, citation accuracy, and context engineering in Deep-Research system.

Zijian Chen, Tingzhu Chen, Wenjun Zhang et al.

We introduce OBI-Bench, a holistic benchmark crafted to systematically evaluate large multi-modal models (LMMs) on whole-process oracle bone inscriptions (OBI) processing tasks demanding expert-level domain knowledge and deliberate cognition. OBI-Bench includes 5,523 meticulously collected diverse-sourced images, covering five key domain problems: recognition, rejoining, classification, retrieval, and deciphering. These images span centuries of archaeological findings and years of research by front-line scholars, comprising multi-stage font appearances from excavation to synthesis, such as original oracle bone, inked rubbings, oracle bone fragments, cropped single characters, and handprinted characters. Unlike existing benchmarks, OBI-Bench focuses on advanced visual perception and reasoning with OBI-specific knowledge, challenging LMMs to perform tasks akin to those faced by experts. The evaluation of 6 proprietary LMMs as well as 17 open-source LMMs highlights the substantial challenges and demands posed by OBI-Bench. Even the latest versions of GPT-4o, Gemini 1.5 Pro, and Qwen-VL-Max are still far from public-level humans in some fine-grained perception tasks. However, they perform at a level comparable to untrained humans in deciphering tasks, indicating remarkable capabilities in offering new interpretative perspectives and generating creative guesses. We hope OBI-Bench can facilitate the community to develop domain-specific multi-modal foundation models towards ancient language research and delve deeper to discover and enhance these untapped potentials of LMMs.

Xiangyang Zhu, Yuan Tian, Qi Jia et al.

The success of large language models (LLMs) in scientific domains has heightened safety concerns, prompting numerous benchmarks to evaluate their scientific safety. Existing benchmarks often suffer from limited risk coverage and a reliance on subjective evaluation. To address these problems, we introduce SafeSci, a comprehensive framework for safety evaluation and enhancement in scientific contexts. SafeSci comprises SafeSciBench, a multi-disciplinary benchmark with 0.25M samples, and SafeSciTrain, a large-scale dataset containing 1.5M samples for safety enhancement. SafeSciBench distinguishes between safety knowledge and risk to cover extensive scopes and employs objective metrics such as deterministically answerable questions to mitigate evaluation bias. We evaluate 24 advanced LLMs, revealing critical vulnerabilities in current models. We also observe that LLMs exhibit varying degrees of excessive refusal behaviors on safety-related issues. For safety enhancement, we demonstrate that fine-tuning on SafeSciTrain significantly enhances the safety alignment of models. Finally, we argue that knowledge is a double-edged sword, and determining the safety of a scientific question should depend on specific context, rather than universally categorizing it as safe or unsafe. Our work provides both a diagnostic tool and a practical resource for building safer scientific AI systems.

Zijian Chen, Yuze Sun, Yuan Tian et al.

Hundreds of benchmarks dedicated to evaluating large models from multiple perspectives have been presented over the past few years. Albeit substantial efforts, most of them remain closed-ended and are prone to overfitting due to the potential data contamination in the ever-growing training corpus of large models, thereby undermining the credibility of the evaluation. Moreover, the increasing scale and scope of current benchmarks with transient metrics, as well as the heavily human-dependent curation procedure, pose significant challenges for timely maintenance and adaptation to gauge the advancing capabilities of large models. In this paper, we introduce MACEval, a \Multi-Agent Continual Evaluation network for dynamic evaluation of large models, and define a new set of metrics to quantify performance longitudinally and sustainably. MACEval adopts an interactive and autonomous evaluation mode that employs role assignment, in-process data generation, and evaluation routing through a cascaded agent network. Extensive experiments on 9 open-ended tasks with 23 participating large models demonstrate that MACEval is (1) human-free and automatic, mitigating laborious result processing with inter-agent judgment guided; (2) efficient and economical, reducing a considerable amount of data and overhead to obtain similar results compared to related benchmarks; and (3) flexible and scalable, migrating or integrating existing benchmarks via customized evaluation topologies. We hope that MACEval can broaden future directions of large model evaluation.

Zijian Chen, Wenjun Zhang, Guangtao Zhai

Contemporary benchmarks are struggling to keep pace with the development of large language models (LLMs). Although they are indispensable to evaluate model performance on various tasks, it is uncertain whether the models trained on Internet data have genuinely learned how to solve problems or merely seen the questions before. This potential data contamination issue presents a fundamental challenge to establishing trustworthy evaluation frameworks. Meanwhile, existing benchmarks predominantly assume benign, resource-rich settings, leaving the behavior of LLMs under pressure unexplored. In this paper, we introduce Squid Game, a dynamic and adversarial evaluation environment with resource-constrained and asymmetric information settings elaborated to evaluate LLMs through interactive gameplay against other LLM opponents. Notably, Squid Game consists of six elimination-style levels, focusing on multi-faceted abilities, such as instruction-following, code, reasoning, planning, and safety alignment. We evaluate over 50 LLMs on Squid Game, presenting the largest behavioral evaluation study of general LLMs on dynamic adversarial scenarios. We observe a clear generational phase transition on performance in the same model lineage and find evidence that some models resort to speculative shortcuts to win the game, indicating the possibility of higher-level evaluation paradigm contamination in static benchmarks. Furthermore, we compare prominent LLM benchmarks and Squid Game with correlation analyses, highlighting that dynamic evaluation can serve as a complementary part for static evaluations. The code and data will be released in the future.

Songtao Jiang, Yuan Wang, Sibo Song et al.

Real-world clinical decision-making requires integrating heterogeneous data, including medical text, 2D images, 3D volumes, and videos, while existing AI systems fail to unify all these signals, limiting their utility. In this paper, we introduce Hulu-Med, a transparent, generalist medical Vision-Language Model (VLM) designed to unify language-only, 2D/3D vision-language, and video understanding within a single architecture. Hulu-Med is trained on a curated corpus of 16.7 million samples, comprising exclusively public or synthetic data, spanning 12 major anatomical systems and 14 medical imaging modalities. Hulu-Med employs a medical-aware token-reduction strategy that prunes redundant visual tokens, achieving up to a 55% reduction for 3D and video inputs, improving cross-modal efficiency, and enabling training at 7B-32B parameter scales in approximately 4,000-40,000 GPU hours. Across 30 public in-domain and out-of-domain medical benchmarks-covering text reasoning, visual question answering, report generation, multilingual dialogue, video understanding, and rare disease diagnosis-Hulu-Med surpasses existing open-source models on 27 of 30 benchmarks and outperforms proprietary systems such as GPT-4o on 16 benchmarks. Despite being a VLM, Hulu-Med outperforms GPT-4o and matches GPT-o1 on the text-only HealthBench. For the first time in the community, we provide a fully transparent, reproducible and cost-effective pipeline for holistic medical vision-language understanding by releasing our end-to-end data curation, training procedures, and model parameters. Code and models are available at https://github.com/ZJUI-AI4H/Hulu-Med.

Kaiwei Zhang, Qi Jia, Zijian Chen et al.

Large language models (LLMs), while increasingly used in domains requiring factual rigor, often display a troubling behavior: sycophancy, the tendency to align with user beliefs regardless of correctness. This tendency is reinforced by preference-based alignment techniques that optimize for user satisfaction but can undermine truthfulness. While relatively benign in casual dialogue, sycophancy poses serious risks in high-stakes settings such as scientific question answering (QA), where model outputs may shape collaborative reasoning, decision-making, and knowledge formation. Despite its importance, this phenomenon remains underexamined in factual QA contexts. We address this gap by introducing a unified evaluation framework to quantify the impact of sycophantic context on model behavior in scientific QA, measuring how much user-imposed social pressure distorts model outputs. The framework incorporates adversarial prompting setups and targeted metrics, such as misleading resistance and sycophancy resistance, that capture a model's ability to maintain factual consistency under misleading cues. Systematic evaluations across open-source and proprietary models reveal pervasive sycophantic tendencies, driven more by alignment strategy than by model size. To mitigate this issue, we propose Pressure-Tune, a lightweight post-training method that fine-tunes models on synthetic adversarial dialogues paired with chain-of-thought rationales. These rationales reject user misinformation while reinforcing factual commitments. Experiments on challenging scientific QA benchmarks show that Pressure-Tune significantly enhances sycophancy resistance without compromising accuracy or responsiveness to valid feedback, offering a practical pathway toward more truthful and principled model behavior.

Zijian Chen, Lirong Deng, Zhengyu Chen et al.

Evaluating the abilities of large models and manifesting their gaps are challenging. Current benchmarks adopt either ground-truth-based score-form evaluation on static datasets or indistinct textual chatbot-style human preferences collection, which may not provide users with immediate, intuitive, and perceptible feedback on performance differences. In this paper, we introduce BioMotion Arena, a novel framework for evaluating large language models (LLMs) and multimodal large language models (MLLMs) via visual animation. Our methodology draws inspiration from the inherent visual perception of motion patterns characteristic of living organisms that utilizes point-light source imaging to amplify the performance discrepancies between models. Specifically, we employ a pairwise comparison evaluation and collect more than 45k votes for 53 mainstream LLMs and MLLMs on 90 biological motion variants. Data analyses show that the crowd-sourced human votes are in good agreement with those of expert raters, demonstrating the superiority of our BioMotion Arena in offering discriminative feedback. We also find that over 90\% of evaluated models, including the cutting-edge open-source InternVL3 and proprietary Claude-4 series, fail to produce fundamental humanoid point-light groups, much less smooth and biologically plausible motions. This enables BioMotion Arena to serve as a challenging benchmark for performance visualization and a flexible evaluation framework without restrictions on ground-truth.

Junying Wang, Zicheng Zhang, Yijin Guo et al.

As foundation models grow rapidly in capability and deployment, evaluating their scientific understanding becomes increasingly critical. Existing science benchmarks have made progress towards broad Range, wide Reach, and high Rigor, yet they often face two major challenges: data leakage risks that compromise benchmarking validity, and evaluation inefficiency due to large-scale testing. To address these issues, we introduce the Ever-Evolving Science Exam (EESE), a dynamic benchmark designed to reliably assess scientific capabilities in foundation models. Our approach consists of two components: 1) a non-public EESE-Pool with over 100K expertly constructed science instances (question-answer pairs) across 5 disciplines and 500+ subfields, built through a multi-stage pipeline ensuring Range, Reach, and Rigor, 2) a periodically updated 500-instance subset EESE, sampled and validated to enable leakage-resilient, low-overhead evaluations. Experiments on 32 open- and closed-source models demonstrate that EESE effectively differentiates the strengths and weaknesses of models in scientific fields and cognitive dimensions. Overall, EESE provides a robust, scalable, and forward-compatible solution for science benchmark design, offering a realistic measure of how well foundation models handle science questions. The project page is at: https://github.com/aiben-ch/EESE.

Zijian Chen, Yuan Tian, Yuze Sun et al.

Just noticeable difference (JND), the minimum change that the human visual system (HVS) can perceive, has been studied for decades. Although recent work has extended this line of research into machine vision, there has been a scarcity of studies systematically exploring its perceptual boundaries across multiple tasks and stimulus types, particularly in the current era of rapidly advancing large multimodal models (LMMs), where studying the multifaceted capabilities of models has become a mainstream focus. Moreover, the perceptual defects of LMMs are not investigated thoroughly, resulting in potential security issues and suboptimal response efficiency. In this paper, we take an initial attempt and demonstrate that there exist significant visual blind spots in current LMMs. To systemically quantify this characteristic, we propose a new concept, {\bf LMM-JND}, together with its determination pipeline. Targeting uncovering the behavior commonalities in HVS-aligned visual perception tasks, we delve into several LMM families and construct a large-scale dataset, named VPA-JND, which contains 21.5k reference images with over 489k stimuli across 12 distortion types, to facilitate LMM-JND studies. VPA-JND exposes areas where state-of-the-art LMMs, including GPT-4o and the InternVL2.5 series, struggle with basic comparison queries and fall significantly short of human-level visual performance. We further explore the effects of vision and language backbones and find a notable correlation between their design philosophy that may instruct the future refinement of LMMs for their visual acuity. Together, our research underscores the significance of LMM-JND as a unique perspective for studying LMMs, and predictable LMM-JND is crucial for security concerns. This work will be available at https://github.com/zijianchen98/LMM-JND.

Wei Sun, Linhan Cao, Yuqin Cao et al.

The rapid proliferation of user-generated content (UGC) on short-form video platforms has made video engagement prediction increasingly important for optimizing recommendation systems and guiding content creation. However, this task remains challenging due to the complex interplay of factors such as semantic content, visual quality, audio characteristics, and user background. Prior studies have leveraged various types of features from different modalities, such as visual quality, semantic content, background sound, etc., but often struggle to effectively model their cross-feature and cross-modality interactions. In this work, we empirically investigate the potential of large multimodal models (LMMs) for video engagement prediction. We adopt two representative LMMs: VideoLLaMA2, which integrates audio, visual, and language modalities, and Qwen2.5-VL, which models only visual and language modalities. Specifically, VideoLLaMA2 jointly processes key video frames, text-based metadata, and background sound, while Qwen2.5-VL utilizes only key video frames and text-based metadata. Trained on the SnapUGC dataset, both models demonstrate competitive performance against state-of-the-art baselines, showcasing the effectiveness of LMMs in engagement prediction. Notably, VideoLLaMA2 consistently outperforms Qwen2.5-VL, highlighting the importance of audio features in engagement prediction. By ensembling two types of models, our method achieves first place in the ICCV VQualA 2025 EVQA-SnapUGC Challenge on short-form video engagement prediction. The code is available at https://github.com/sunwei925/LMM-EVQA.git.

Jiarui Wang, Huiyu Duan, Ziheng Jia et al.

Recent advancements in large multimodal models (LMMs) have driven substantial progress in both text-to-video (T2V) generation and video-to-text (V2T) interpretation tasks. However, current AI-generated videos (AIGVs) still exhibit limitations in terms of perceptual quality and text-video alignment. Therefore, a reliable and scalable automatic model for AIGV evaluation is desirable, which heavily relies on the scale and quality of human annotations. To this end, we present AIGVE-60K, a comprehensive dataset and benchmark for AI-Generated Video Evaluation, which features (i) comprehensive tasks, encompassing 3,050 extensive prompts across 20 fine-grained task dimensions, (ii) the largest human annotations, including 120K mean-opinion scores (MOSs) and 60K question-answering (QA) pairs annotated on 58,500 videos generated from 30 T2V models, and (iii) bidirectional benchmarking and evaluating for both T2V generation and V2T interpretation capabilities. Based on AIGVE-60K, we propose LOVE, a LMM-based metric for AIGV Evaluation from multiple dimensions including perceptual preference, text-video correspondence, and task-specific accuracy in terms of both instance level and model level. Comprehensive experiments demonstrate that LOVE not only achieves state-of-the-art performance on the AIGVE-60K dataset, but also generalizes effectively to a wide range of other AIGV evaluation benchmarks. These findings highlight the significance of the AIGVE-60K dataset. Database and codes are anonymously available at https://github.com/IntMeGroup/LOVE.

Jinhao Li, Zijian Chen, Tingzhu Chen et al.

Oracle bone inscriptions (OBIs) are the earliest known form of Chinese characters and serve as a valuable resource for research in anthropology and archaeology. However, most excavated fragments are severely degraded due to thousands of years of natural weathering, corrosion, and man-made destruction, making automatic OBI recognition extremely challenging. Previous methods either focus on pixel-level information or utilize vanilla transformers for glyph-based OBI denoising, which leads to tremendous computational overhead. Therefore, this paper proposes a fast attentive denoising framework for oracle bone inscriptions, i.e., OBIFormer. It leverages channel-wise self-attention, glyph extraction, and selective kernel feature fusion to reconstruct denoised images precisely while being computationally efficient. Our OBIFormer achieves state-of-the-art denoising performance for PSNR and SSIM metrics on synthetic and original OBI datasets. Furthermore, comprehensive experiments on a real oracle dataset demonstrate the great potential of our OBIFormer in assisting automatic OBI recognition. The code will be made available at https://github.com/LJHolyGround/OBIFormer.

Zijian Chen, Wenjie Hua, Jinhao Li et al.

Deciphering oracle bone characters (OBCs), the oldest attested form of written Chinese, has remained the ultimate, unwavering goal of scholars, offering an irreplaceable key to understanding humanity's early modes of production. Current decipherment methodologies of OBC are primarily constrained by the sporadic nature of archaeological excavations and the limited corpus of inscriptions. With the powerful visual perception capability of large multimodal models (LMMs), the potential of using LMMs for visually deciphering OBCs has increased. In this paper, we introduce PictOBI-20k, a dataset designed to evaluate LMMs on the visual decipherment tasks of pictographic OBCs. It includes 20k meticulously collected OBC and real object images, forming over 15k multi-choice questions. We also conduct subjective annotations to investigate the consistency of the reference point between humans and LMMs in visual reasoning. Experiments indicate that general LMMs possess preliminary visual decipherment skills, and LMMs are not effectively using visual information, while most of the time they are limited by language priors. We hope that our dataset can facilitate the evaluation and optimization of visual attention in future OBC-oriented LMMs. The code and dataset will be available at https://github.com/OBI-Future/PictOBI-20k.

Zijian Chen, John-Michael Gamble, Micaela Jantzi et al.

Manual assignment of Anatomical Therapeutic Chemical (ATC) codes to prescription records is a significant bottleneck in healthcare research and operations at Ontario Health and InterRAI Canada, requiring extensive expert time and effort. To automate this process while maintaining data privacy, we develop a practical approach using locally deployable large language models (LLMs). Inspired by recent advances in automatic International Classification of Diseases (ICD) coding, our method frames ATC coding as a hierarchical information extraction task, guiding LLMs through the ATC ontology level by level. We evaluate our approach using GPT-4o as an accuracy ceiling and focus development on open-source Llama models suitable for privacy-sensitive deployment. Testing across Health Canada drug product data, the RABBITS benchmark, and real clinical notes from Ontario Health, our method achieves 78% exact match accuracy with GPT-4o and 60% with Llama 3.1 70B. We investigate knowledge grounding through drug definitions, finding modest improvements in accuracy. Further, we show that fine-tuned Llama 3.1 8B matches zero-shot Llama 3.1 70B accuracy, suggesting that effective ATC coding is feasible with smaller models. Our results demonstrate the feasibility of automatic ATC coding in privacy-sensitive healthcare environments, providing a foundation for future deployments.

Nandan Thakur, Zijian Chen, Xueguang Ma et al.

Search agents, which integrate language models (LMs) with web search, are becoming crucial for answering complex user queries. Constructing training datasets for deep research tasks, involving multi-step retrieval and reasoning, remains challenging due to expensive human annotation, or cumbersome prerequisites. In this work, we introduce ORBIT, a training dataset with 20K reasoning-intensive queries with short verifiable answers, generated using a frugal framework without relying on paid API services. The modular framework relies on four stages: seed creation, question--answer pair generation, and two stages of verification: self and external. ORBIT spans 15 domains and each training pair requires 4--5 reasoning steps, with external search verification required from the complete web. We train Qwen3-4B as the base model on ORBIT using GRPO and evaluate it on Wikipedia question answering tasks. Extensive experiment results demonstrate that ORBIT-4B achieves strong performance among sub-4B LLMs as search agents, proving the utility of synthetic datasets. Our framework, code and datasets are open-sourced and available publicly.

Jinnao Li, Zijian Chen, Tingzhu Chen et al.

Geo-localization aims to infer the geographic location where an image was captured using observable visual evidence. Traditional methods achieve impressive results through large-scale training on massive image corpora. With the emergence of multi-modal large language models (MLLMs), recent studies have explored their applications in geo-localization, benefiting from improved accuracy and interpretability. However, existing benchmarks largely ignore the temporal information inherent in images, which can further constrain the location. To bridge this gap, we introduce GTPred, a novel benchmark for geo-temporal prediction. GTPred comprises 370 globally distributed images spanning over 120 years. We evaluate MLLM predictions by jointly considering year and hierarchical location sequence matching, and further assess intermediate reasoning chains using meticulously annotated ground-truth reasoning processes. Experiments on 8 proprietary and 7 open-source MLLMs show that, despite strong visual perception, current models remain limited in world knowledge and geo-temporal reasoning. Results also demonstrate that incorporating temporal information significantly enhances location inference performance.

Zicheng Zhang, Haoning Wu, Chunyi Li et al.

How to accurately and efficiently assess AI-generated images (AIGIs) remains a critical challenge for generative models. Given the high costs and extensive time commitments required for user studies, many researchers have turned towards employing large multi-modal models (LMMs) as AIGI evaluators, the precision and validity of which are still questionable. Furthermore, traditional benchmarks often utilize mostly natural-captured content rather than AIGIs to test the abilities of LMMs, leading to a noticeable gap for AIGIs. Therefore, we introduce A-Bench in this paper, a benchmark designed to diagnose whether LMMs are masters at evaluating AIGIs. Specifically, A-Bench is organized under two key principles: 1) Emphasizing both high-level semantic understanding and low-level visual quality perception to address the intricate demands of AIGIs. 2) Various generative models are utilized for AIGI creation, and various LMMs are employed for evaluation, which ensures a comprehensive validation scope. Ultimately, 2,864 AIGIs from 16 text-to-image models are sampled, each paired with question-answers annotated by human experts, and tested across 18 leading LMMs. We hope that A-Bench will significantly enhance the evaluation process and promote the generation quality for AIGIs. The benchmark is available at https://github.com/Q-Future/A-Bench.

Qi Jia, Haodong Zhao, Dun Pei et al.

Benchmarking AI systems in multi-turn interactive scenarios is essential for understanding their practical capabilities in real-world applications. However, existing evaluation protocols are highly heterogeneous, differing significantly in dataset formats, model interfaces, and evaluation pipelines, which severely impedes systematic comparison. In this work, we present UniDial-EvalKit (UDE), a unified evaluation toolkit for assessing interactive AI systems. The core contribution of UDE lies in its holistic unification: it standardizes heterogeneous data formats into a universal schema, streamlines complex evaluation pipelines through a modular architecture, and aligns metric calculations under a consistent scoring interface. It also supports efficient large-scale evaluation through parallel generation and scoring, as well as checkpoint-based caching to eliminate redundant computation. Validated across diverse multi-turn benchmarks, UDE not only guarantees high reproducibility through standardized workflows and transparent logging, but also significantly improves evaluation efficiency and extensibility. We make the complete toolkit and evaluation scripts publicly available to foster a standardized benchmarking ecosystem and accelerate future breakthroughs in interactive AI.

Zijian Chen, Wei Sun, Haoning Wu et al.

The proliferation of Artificial Intelligence-Generated Images (AGIs) has greatly expanded the Image Naturalness Assessment (INA) problem. Different from early definitions that mainly focus on tone-mapped images with limited distortions (e.g., exposure, contrast, and color reproduction), INA on AI-generated images is especially challenging as it has more diverse contents and could be affected by factors from multiple perspectives, including low-level technical distortions and high-level rationality distortions. In this paper, we take the first step to benchmark and assess the visual naturalness of AI-generated images. First, we construct the AI-Generated Image Naturalness (AGIN) database by conducting a large-scale subjective study to collect human opinions on the overall naturalness as well as perceptions from technical and rationality perspectives. AGIN verifies that naturalness is universally and disparately affected by technical and rationality distortions. Second, we propose the Joint Objective Image Naturalness evaluaTor (JOINT), to automatically predict the naturalness of AGIs that aligns human ratings. Specifically, JOINT imitates human reasoning in naturalness evaluation by jointly learning both technical and rationality features. We demonstrate that JOINT significantly outperforms baselines for providing more subjectively consistent results on naturalness assessment.

Zijian Chen, Jueqi Wang, Archana Venkataraman

We propose an image-conditioned diffusion model to estimate high angular resolution diffusion weighted imaging (DWI) from a low angular resolution acquisition. Our model, which we call QID$^2$, takes as input a set of low angular resolution DWI data and uses this information to estimate the DWI data associated with a target gradient direction. We leverage a U-Net architecture with cross-attention to preserve the positional information of the reference images, further guiding the target image generation. We train and evaluate QID$^2$ on single-shell DWI samples curated from the Human Connectome Project (HCP) dataset. Specifically, we sub-sample the HCP gradient directions to produce low angular resolution DWI data and train QID$^2$ to reconstruct the missing high angular resolution samples. We compare QID$^2$ with two state-of-the-art GAN models. Our results demonstrate that QID$^2$ not only achieves higher-quality generated images, but it consistently outperforms the GAN models in downstream tensor estimation across multiple metrics. Taken together, this study highlights the potential of diffusion models, and QID$^2$ in particular, for q-space up-sampling, thus offering a promising toolkit for clinical and research applications.

Marcos V. Conde, Saman Zadtootaghaj, Nabajeet Barman et al.

This paper reviews the AIS 2024 Video Quality Assessment (VQA) Challenge, focused on User-Generated Content (UGC). The aim of this challenge is to gather deep learning-based methods capable of estimating the perceptual quality of UGC videos. The user-generated videos from the YouTube UGC Dataset include diverse content (sports, games, lyrics, anime, etc.), quality and resolutions. The proposed methods must process 30 FHD frames under 1 second. In the challenge, a total of 102 participants registered, and 15 submitted code and models. The performance of the top-5 submissions is reviewed and provided here as a survey of diverse deep models for efficient video quality assessment of user-generated content.

Dasong Li, Sizhuo Ma, Hang Hua et al.

This paper presents an overview of the VQualA 2025 Challenge on Engagement Prediction for Short Videos, held in conjunction with ICCV 2025. The challenge focuses on understanding and modeling the popularity of user-generated content (UGC) short videos on social media platforms. To support this goal, the challenge uses a new short-form UGC dataset featuring engagement metrics derived from real-world user interactions. This objective of the Challenge is to promote robust modeling strategies that capture the complex factors influencing user engagement. Participants explored a variety of multi-modal features, including visual content, audio, and metadata provided by creators. The challenge attracted 97 participants and received 15 valid test submissions, contributing significantly to progress in short-form UGC video engagement prediction.

Zeyu Zhang, Zijian Chen, Zicheng Zhang et al.

Large Multimodal Models (LMMs) have demonstrated impressive capabilities across a wide range of multimodal tasks, achieving ever-increasing performance on various evaluation benchmarks. However, existing benchmarks are typically static and often overlap with pre-training datasets, leading to fixed complexity constraints and substantial data contamination issues. Meanwhile, manually annotated datasets are labor-intensive, time-consuming, and subject to human bias and inconsistency, leading to reliability and reproducibility issues. To address these problems, we propose a fully dynamic multimodal evaluation framework, named Open-ended Visual Puzzle Generation (OVPG), which aims to generate fresh, diverse, and verifiable evaluation data automatically in puzzle-solving tasks. Specifically, the OVPG pipeline consists of a raw material sampling module, a visual content generation module, and a puzzle rule design module, which ensures that each evaluation instance is primitive, highly randomized, and uniquely solvable, enabling continual adaptation to the evolving capabilities of LMMs. Built upon OVPG, we construct PuzzleBench, a dynamic and scalable benchmark comprising 11,840 VQA samples. It features six carefully designed puzzle tasks targeting three core LMM competencies, visual recognition, logical reasoning, and context understanding. PuzzleBench differs from static benchmarks that quickly become outdated. It enables ongoing dataset refreshing through OVPG and a rich set of open-ended puzzle designs, allowing seamless adaptation to the evolving capabilities of LMMs.

Zijian Chen, Ronak Pradeep, Jimmy Lin

Recent advances have demonstrated that large language models (LLMs) excel as listwise rerankers, but their high computational demands remain a barrier to widespread adoption. Further, the traditional language modeling (LM) objective is not ideally suited for reranking tasks. FIRST is a novel approach that addresses these challenges by integrating a learning-to-rank objective and leveraging the logits of only the first generated token, thereby significantly reducing inference latency compared to traditional LLM rerankers. In this study, we extend the evaluation of FIRST to the TREC Deep Learning datasets (DL19-22), validating its robustness across diverse domains. We investigate the influence of different first-stage retrievers on FIRST rerankers, observing diminishing returns and patterns consistent with traditional LLM rerankers. Through applying the FIRST objective to a broader range of backbone models, we achieve effectiveness surpassing the original implementation. Our experiments confirm that fast reranking with single-token logits does not compromise out-of-domain reranking quality. To better quantify the computational savings in the original study, we measure and compare latency to find a 21%-42% gain across various models and benchmarks. Moreover, while LM training implicitly improves zero-shot single-token reranking, our experiments also raise questions about whether LM pre-training may hinder subsequent fine-tuning with the FIRST objective. These findings pave the way for more efficient and effective listwise reranking in future applications.

Jinhao Li, Zijian Chen, Runze Jiang et al.

The oracle bone inscription (OBI) recognition plays a significant role in understanding the history and culture of ancient China. However, the existing OBI datasets suffer from a long-tail distribution problem, leading to biased performance of OBI recognition models across majority and minority classes. With recent advancements in generative models, OBI synthesis-based data augmentation has become a promising avenue to expand the sample size of minority classes. Unfortunately, current OBI datasets lack large-scale structure-aligned image pairs for generative model training. To address these problems, we first present the Oracle-P15K, a structure-aligned OBI dataset for OBI generation and denoising, consisting of 14,542 images infused with domain knowledge from OBI experts. Second, we propose a diffusion model-based pseudo OBI generator, called OBIDiff, to achieve realistic and controllable OBI generation. Given a clean glyph image and a target rubbing-style image, it can effectively transfer the noise style of the original rubbing to the glyph image. Extensive experiments on OBI downstream tasks and user preference studies show the effectiveness of the proposed Oracle-P15K dataset and demonstrate that OBIDiff can accurately preserve inherent glyph structures while transferring authentic rubbing styles effectively.

Zijian Chen, Mei Wang, Weihong Deng et al.

2D face recognition encounters challenges in unconstrained environments due to varying illumination, occlusion, and pose. Recent studies focus on RGB-D face recognition to improve robustness by incorporating depth information. However, collecting sufficient paired RGB-D training data is expensive and time-consuming, hindering wide deployment. In this work, we first construct a diverse depth dataset generated by 3D Morphable Models for depth model pre-training. Then, we propose a domain-independent pre-training framework that utilizes readily available pre-trained RGB and depth models to separately perform face recognition without needing additional paired data for retraining. To seamlessly integrate the two distinct networks and harness the complementary benefits of RGB and depth information for improved accuracy, we propose an innovative Adaptive Confidence Weighting (ACW). This mechanism is designed to learn confidence estimates for each modality to achieve modality fusion at the score level. Our method is simple and lightweight, only requiring ACW training beyond the backbone models. Experiments on multiple public RGB-D face recognition benchmarks demonstrate state-of-the-art performance surpassing previous methods based on depth estimation and feature fusion, validating the efficacy of our approach.

Qi Jia, Xiujie Song, Zicheng Zhang et al.

Existing benchmarks for large language models (LLMs) predominantely focus on assessing their capabilities through verifiable tasks. Such objective and static benchmarks offer limited utility for practical LLM selection, making it difficult for users to find suitable models for their individual needs. To bridge this gap, we present the first User-Centric Subjective Leaderboard (USL), which provides a preference-driven, dynamic ranking of LLMs across diverse real-world scenarios. Our work is built upon a thorough investigation of real human preference data, involving more than 10K subjective queries. Our investigation reveals significant diversity and contradictions in human preferences, which limit the effectiveness of state-of-the-art reward models. To address this, we introduce Customizable Reward Models (CRMs). With only 4B parameters, our CRM surpasses the performance of leading models such as GPT-4.1 and Gemini-2.5-pro, showing exceptional generalization capabilities across new topics and criteria. The USL, powered by CRMs, exhibits strong negative correlations to contradictory preferences.

Yuan Tian, Shuo Wang, Rongzhao Zhang et al.

Medical imaging has significantly advanced computer-aided diagnosis, yet its re-identification (ReID) risks raise critical privacy concerns, calling for de-identification (DeID) techniques. Unfortunately, existing DeID methods neither particularly preserve medical semantics, nor are flexibly adjustable towards different privacy levels. To address these issues, we propose a divide-and-conquer framework comprising two steps: (1) Identity-Blocking, which blocks varying proportions of identity-related regions, to achieve different privacy levels; and (2) Medical-Semantics-Compensation, which leverages pre-trained Medical Foundation Models (MFMs) to extract medical semantic features to compensate the blocked regions. Moreover, recognizing that features from MFMs may still contain residual identity information, we introduce a Minimum Description Length principle-based feature decoupling strategy, to effectively decouple and discard such identity components. Extensive evaluations against existing approaches across seven datasets and three downstream tasks, demonstrates our state-of-the-art performance.

Ziheng Jia, Linhan Cao, Jinliang Han et al.

Developing a robust visual quality assessment (VQualA) large multi-modal model (LMM) requires achieving versatility, powerfulness, and transferability. However, existing VQualA LMMs typically focus on a single task and rely on full-parameter fine-tuning, which makes them prone to overfitting on specific modalities or task types, thereby limiting their generalization capacity and transferability. To address this, we propose a vision-encoder-centered generative pre-training pipeline and develop the VITAL-Series LMMs. (1) We adopt a machine-executed annotation-scrutiny paradigm, constructing over 4.5M vision-language (VL) pairs-the largest VQualA training dataset to date. (2) We employ a multi-task training workflow that simultaneously enhances the model's quantitative scoring precision and strengthens its capability for quality interpretation across both image and video modalities. (3) Building upon the vision encoder, we realize an efficient model zoo extension: the model zoo exhibits strong zero-shot performance, and each paired decoder requires only a swift warm-up using less than 1/1000 of the pre-training data to achieve performance comparable to the fully trained counterpart. Overall, our work lays a cornerstone for advancing toward the foundation LMM for VQualA.

Jinhao Li, Zijian Chen, Lirong Deng et al.

Person re-identification (ReID) aims to retrieve the images of an interested person in the gallery images, with wide applications in medical rehabilitation, abnormal behavior detection, and public security. However, traditional person ReID models suffer from uni-modal capability, leading to poor generalization ability in multi-modal data, such as RGB, thermal, infrared, sketch images, textual descriptions, etc. Recently, the emergence of multi-modal large language models (MLLMs) shows a promising avenue for addressing this problem. Despite this potential, existing methods merely regard MLLMs as feature extractors or caption generators, which do not fully unleash their reasoning, instruction-following, and cross-modal understanding capabilities. To bridge this gap, we introduce MMReID-Bench, the first multi-task multi-modal benchmark specifically designed for person ReID. The MMReID-Bench includes 20,710 multi-modal queries and gallery images covering 10 different person ReID tasks. Comprehensive experiments demonstrate the remarkable capabilities of MLLMs in delivering effective and versatile person ReID. Nevertheless, they also have limitations in handling a few modalities, particularly thermal and infrared data. We hope MMReID-Bench can facilitate the community to develop more robust and generalizable multimodal foundation models for person ReID.

Ziheng Jia, Jiaying Qian, Zicheng Zhang et al.

Reinforcement fine-tuning (RFT) is a proliferating paradigm for LMM training. Analogous to high-level reasoning tasks, RFT is similarly applicable to low-level vision domains, including image quality assessment (IQA). Existing RFT-based IQA methods typically use rule-based output rewards to verify the model's rollouts but provide no reward supervision for the "think" process, leaving its correctness and efficacy uncontrolled. Furthermore, these methods typically fine-tune directly on downstream IQA tasks without explicitly enhancing the model's native low-level visual quality perception, which may constrain its performance upper bound. In response to these gaps, we propose the multi-stage RFT IQA framework (Refine-IQA). In Stage-1, we build the Refine-Perception-20K dataset (with 12 main distortions, 20,907 locally-distorted images, and over 55K RFT samples) and design multi-task reward functions to strengthen the model's visual quality perception. In Stage-2, targeting the quality scoring task, we introduce a probability difference reward involved strategy for "think" process supervision. The resulting Refine-IQA Series Models achieve outstanding performance on both perception and scoring tasks-and, notably, our paradigm activates a robust "think" (quality interpreting) capability that also attains exceptional results on the corresponding quality interpreting benchmark.

Renjie Li, Wenjie Wei, Qi Xin et al.

Large language models (LLMs) are rapidly pushing the limits of contemporary computing hardware. For example, training GPT-3 has been estimated to consume around 1300 MWh of electricity, and projections suggest future models may require city-scale (gigawatt) power budgets. These demands motivate exploration of computing paradigms beyond conventional von Neumann architectures. This review surveys emerging photonic hardware optimized for next-generation generative AI computing. We discuss integrated photonic neural network architectures (e.g., Mach-Zehnder interferometer meshes, lasers, wavelength-multiplexed microring resonators) that perform ultrafast matrix operations. We also examine promising alternative neuromorphic devices, including spiking neural network circuits and hybrid spintronic-photonic synapses, which combine memory and processing. The integration of two-dimensional materials (graphene, TMDCs) into silicon photonic platforms is reviewed for tunable modulators and on-chip synaptic elements. Transformer-based LLM architectures (self-attention and feed-forward layers) are analyzed in this context, identifying strategies and challenges for mapping dynamic matrix multiplications onto these novel hardware substrates. We then dissect the mechanisms of mainstream LLMs, such as ChatGPT, DeepSeek, and LLaMA, highlighting their architectural similarities and differences. We synthesize state-of-the-art components, algorithms, and integration methods, highlighting key advances and open issues in scaling such systems to mega-sized LLM models. We find that photonic computing systems could potentially surpass electronic processors by orders of magnitude in throughput and energy efficiency, but require breakthroughs in memory, especially for long-context windows and long token sequences, and in storage of ultra-large datasets.

Zijian Chen, Wei Sun, Yuan Tian et al.

Assessing action quality is both imperative and challenging due to its significant impact on the quality of AI-generated videos, further complicated by the inherently ambiguous nature of actions within AI-generated video (AIGV). Current action quality assessment (AQA) algorithms predominantly focus on actions from real specific scenarios and are pre-trained with normative action features, thus rendering them inapplicable in AIGVs. To address these problems, we construct GAIA, a Generic AI-generated Action dataset, by conducting a large-scale subjective evaluation from a novel causal reasoning-based perspective, resulting in 971,244 ratings among 9,180 video-action pairs. Based on GAIA, we evaluate a suite of popular text-to-video (T2V) models on their ability to generate visually rational actions, revealing their pros and cons on different categories of actions. We also extend GAIA as a testbed to benchmark the AQA capacity of existing automatic evaluation methods. Results show that traditional AQA methods, action-related metrics in recent T2V benchmarks, and mainstream video quality methods perform poorly with an average SRCC of 0.454, 0.191, and 0.519, respectively, indicating a sizable gap between current models and human action perception patterns in AIGVs. Our findings underscore the significance of action quality as a unique perspective for studying AIGVs and can catalyze progress towards methods with enhanced capacities for AQA in AIGVs.