arXiv subjects

Guangtao Zhai

Publications and source records attributed to Guangtao Zhai.

At least 109 records · Page 6Linked to original sources

arXiv2026

Evaluating from Benign to Dynamic Adversarial: A Squid Game for Large Language Models

The potential data contamination issue in contemporary large language models (LLMs) benchmarks presents a fundamental challenge to establishing trustworthy evaluation frameworks. Meanwhile, they predominantly assume benign, resource-rich settings, leaving the behavior of LLMs under pressure unexplored. In this paper, we introduce \textsc{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. Squid Game consists of six elimination-style levels, focusing on multi-faceted abilities, including 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 in 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. We also compare prominent LLM benchmarks and \textsc{Squid Game}, highlighting that dynamic evaluation can serve as a complementary part for static evaluations. Project page: https://github.com/zijianchen98/LLM_Squid_Game.

cs.CL↗

arXiv2026

SFQA: A Comprehensive Perceptual Quality Assessment Dataset for Singing Face Generation

The Talking Face Generation task has enormous potential for various applications in digital humans and agents, etc. Singing, as a common facial movement second only to talking, can be regarded as a universal language across ethnicities and cultures. However, it is often underestimated in the field due to lack of singing face datasets and the domain gap between singing and talking in rhythm and amplitude. More significantly, the quality of Singing Face Generation (SFG) often falls short and is uneven or limited by different applicable scenarios, which prompts us to propose timely and effective quality assessment methods to ensure user experience. To address existing gaps in this domain, this paper introduces a new SFG content quality assessment dataset SFQA, built using 12 representative generation methods. During the construction of the dataset, 100 photographs or portraits, as well as 36 music clips from 7 different styles, are utilized to generate 5,184 singing face videos that constitute the SFQA dataset. To further explore the quality of SFG methods, subjective quality assessment is conducted by evaluators, whose ratings reveal a significant variation in quality among different generation methods. Based on our proposed SFQA dataset, we comprehensively benchmark the current objective quality assessment algorithms.

cs.MM↗

arXiv2026

Decoupling Perception and Calibration: Label-Efficient Image Quality Assessment Framework

Recent multimodal large language models (MLLMs) have demonstrated strong capabilities in image quality assessment (IQA) tasks. However, adapting such large-scale models is computationally expensive and still relies on substantial Mean Opinion Score (MOS) annotations. We argue that for MLLM-based IQA, the core bottleneck lies not in the quality perception capacity of MLLMs, but in MOS scale calibration. Therefore, we propose LEAF, a Label-Efficient Image Quality Assessment Framework that distills perceptual quality priors from an MLLM teacher into a lightweight student regressor, enabling MOS calibration with minimal human supervision. Specifically, the teacher conducts dense supervision through point-wise judgments and pair-wise preferences, with an estimate of decision reliability. Guided by these signals, the student learns the teacher's quality perception patterns through joint distillation and is calibrated on a small MOS subset to align with human annotations. Experiments on both user-generated and AI-generated IQA benchmarks demonstrate that our method significantly reduces the need for human annotations while maintaining strong MOS-aligned correlations, making lightweight IQA practical under limited annotation budgets.

cs.CV↗

arXiv2026

QualiRAG: Retrieval-Augmented Generation for Visual Quality Understanding

Visual quality assessment (VQA) is increasingly shifting from scalar score prediction toward interpretable quality understanding -- a paradigm that demands \textit{fine-grained spatiotemporal perception} and \textit{auxiliary contextual information}. Current approaches rely on supervised fine-tuning or reinforcement learning on curated instruction datasets, which involve labor-intensive annotation and are prone to dataset-specific biases. To address these challenges, we propose \textbf{QualiRAG}, a \textit{training-free} \textbf{R}etrieval-\textbf{A}ugmented \textbf{G}eneration \textbf{(RAG)} framework that systematically leverages the latent perceptual knowledge of large multimodal models (LMMs) for visual quality perception. Unlike conventional RAG that retrieves from static corpora, QualiRAG dynamically generates auxiliary knowledge by decomposing questions into structured requests and constructing four complementary knowledge sources: \textit{visual metadata}, \textit{subject localization}, \textit{global quality summaries}, and \textit{local quality descriptions}, followed by relevance-aware retrieval for evidence-grounded reasoning. Extensive experiments show that QualiRAG achieves substantial improvements over open-source general-purpose LMMs and VQA-finetuned LMMs on visual quality understanding tasks, and delivers competitive performance on visual quality comparison tasks, demonstrating robust quality assessment capabilities without any task-specific training. The code will be publicly available at https://github.com/clh124/QualiRAG.

cs.CV↗

arXiv2026

Q-Bench-Portrait: Benchmarking Multimodal Large Language Models on Portrait Image Quality Perception

Recent advances in multimodal large language models (MLLMs) have demonstrated impressive performance on existing low-level vision benchmarks, which primarily focus on generic images. However, their capabilities to perceive and assess portrait images, a domain characterized by distinct structural and perceptual properties, remain largely underexplored. To this end, we introduce Q-Bench-Portrait, the first holistic benchmark specifically designed for portrait image quality perception, comprising 2,765 image-question-answer triplets and featuring (1) diverse portrait image sources, including natural, synthetic distortion, AI-generated, artistic, and computer graphics images; (2) comprehensive quality dimensions, covering technical distortions, AIGC-specific distortions, and aesthetics; and (3) a range of question formats, including single-choice, multiple-choice, true/false, and open-ended questions, at both global and local levels. Based on Q-Bench-Portrait, we evaluate 20 open-source and 5 closed-source MLLMs, revealing that although current models demonstrate some competence in portrait image perception, their performance remains limited and imprecise, with a clear gap relative to human judgments. We hope that the proposed benchmark will foster further research into enhancing the portrait image perception capabilities of both general-purpose and domain-specific MLLMs.

cs.CV↗

arXiv2026

REF-VLM: Triplet-Based Referring Paradigm for Unified Visual Decoding

Multimodal Large Language Models (MLLMs) demonstrate robust zero-shot capabilities across diverse vision-language tasks after training on mega-scale datasets. However, dense prediction tasks, such as semantic segmentation and keypoint detection, pose significant challenges for MLLMs when represented solely as text outputs. Simultaneously, current MLLMs utilizing latent embeddings for visual task decoding generally demonstrate limited adaptability to both multi-task learning and multi-granularity scenarios. In this work, we present \textbf{REF-VLM}, an end-to-end framework for unified training of various visual decoding tasks. To address complex visual decoding scenarios, we introduce the \textbf{Triplet-Based Referring Paradigm (TRP)}, which explicitly decouples three critical dimensions in visual decoding tasks through a triplet structure: concepts, decoding types, and targets. TRP employs symbolic delimiters to enforce structured representation learning, enhancing the parsability and interpretability of model outputs. Additionally, we construct \textbf{Visual-Task Instruction Following Dataset (VT-Instruct)}, a large-scale multi-task dataset containing over 100 million multimodal dialogue samples across 25 task types. Beyond text inputs and outputs, VT-Instruct incorporates various visual prompts such as point, box, scribble, and mask, and generates outputs composed of text and visual units like box, keypoint, depth and mask. The combination of different visual prompts and visual units generates a wide variety of task types, expanding the applicability of REF-VLM significantly. Both qualitative and quantitative experiments demonstrate that our REF-VLM outperforms other MLLMs across a variety of standard benchmarks. The code, dataset, and demo will be publicly available.

cs.CV↗

arXiv2026

KidVis: Do Multimodal Large Language Models Possess the Visual Perceptual Capabilities of a 6-Year-Old?

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.

cs.CV↗

arXiv2026

Enhancing Image Quality Assessment Ability of LMMs via Retrieval-Augmented Generation

Large Multimodal Models (LMMs) have recently shown remarkable promise in low-level visual perception tasks, particularly in Image Quality Assessment (IQA), demonstrating strong zero-shot capability. However, achieving state-of-the-art performance often requires computationally expensive fine-tuning methods, which aim to align the distribution of quality-related token in output with image quality levels. Inspired by recent training-free works for LMM, we introduce IQARAG, a novel, training-free framework that enhances LMMs' IQA ability. IQARAG leverages Retrieval-Augmented Generation (RAG) to retrieve some semantically similar but quality-variant reference images with corresponding Mean Opinion Scores (MOSs) for input image. These retrieved images and input image are integrated into a specific prompt. Retrieved images provide the LMM with a visual perception anchor for IQA task. IQARAG contains three key phases: Retrieval Feature Extraction, Image Retrieval, and Integration & Quality Score Generation. Extensive experiments across multiple diverse IQA datasets, including KADID, KonIQ, LIVE Challenge, and SPAQ, demonstrate that the proposed IQARAG effectively boosts the IQA performance of LMMs, offering a resource-efficient alternative to fine-tuning for quality assessment.

cs.CV↗

arXiv2026

MoGen: A Unified Collaborative Framework for Controllable Multi-Object Image Generation

Existing multi-object image generation methods face difficulties in achieving precise alignment between localized image generation regions and their corresponding semantics based on language descriptions, frequently resulting in inconsistent object quantities and attribute aliasing. To mitigate this limitation, mainstream approaches typically rely on external control signals to explicitly constrain the spatial layout, local semantic and visual attributes of images. However, this strong dependency makes the input format rigid, rendering it incompatible with the heterogeneous resource conditions of users and diverse constraint requirements. To address these challenges, we propose MoGen, a user-friendly multi-object image generation method. First, we design a Regional Semantic Anchor (RSA) module that precisely anchors phrase units in language descriptions to their corresponding image regions during the generation process, enabling text-to-image generation that follows quantity specifications for multiple objects. Building upon this foundation, we further introduce an Adaptive Multi-modal Guidance (AMG) module, which adaptively parses and integrates various combinations of multi-source control signals to formulate corresponding structured intent. This intent subsequently guides selective constraints on scene layouts and object attributes, achieving dynamic fine-grained control. Experimental results demonstrate that MoGen significantly outperforms existing methods in generation quality, quantity consistency, and fine-grained control, while exhibiting superior accessibility and control flexibility. Code is available at: https://github.com/Tear-kitty/MoGen/tree/master.

cs.CV↗

arXiv2026

One Battle After Another: Probing LLMs' Limits on Multi-Turn Instruction Following with a Benchmark Evolving Framework

Evaluating LLMs' instruction-following ability in multi-topic dialogues is essential yet challenging. Existing benchmarks are limited to a fixed number of turns, susceptible to saturation and failing to account for users' interactive experience. In this work, we propose a novel framework featuring a three-layer tracking mechanism and a query synthesis agent to mimic sequential user behaviors. Grounded in Flow Theory, we introduce process-centric metrics and terminate a conversational evaluation only upon exhausting user patience. Leveraging this framework, we present EvolIF, an evolving benchmark covering 12 constraint groups. Our analysis reveals deficiencies in failure recovery and fine-grained instruction following, with performance stratification becoming evident as conversational depth increases. GPT-5 demonstrates the most sustained resilience, maintaining a 66.40% robustness score, outperforming Gemini-3-Pro by 5.59%, while other models lag behind. Data and code will be released at https://github.com/JiaQiSJTU/EvolIF.

cs.CL↗

arXiv2026

EvolMem: A Cognitive-Driven Benchmark for Multi-Session Dialogue Memory

Despite recent advances in understanding and leveraging long-range conversational memory, existing benchmarks still lack systematic evaluation of large language models(LLMs) across diverse memory dimensions, particularly in multi-session settings. In this work, we propose EvolMem, a new benchmark for assessing multi-session memory capabilities of LLMs and agent systems. EvolMem is grounded in cognitive psychology and encompasses both declarative and non-declarative memory, further decomposed into multiple fine-grained abilities. To construct the benchmark, we introduce a hybrid data synthesis framework that consists of topic-initiated generation and narrative-inspired transformations. This framework enables scalable generation of multi-session conversations with controllable complexity, accompanied by sample-specific evaluation guidelines. Extensive evaluation reveals that no LLM consistently outperforms others across all memory dimensions. Moreover, agent memory mechanisms do not necessarily enhance LLMs' capabilities and often exhibit notable efficiency limitations. Data and code will be released at https://github.com/shenye7436/EvolMem.

cs.CL↗

arXiv2026

Enhancing Blind Video Quality Assessment with Rich Quality-aware Features

Blind video quality assessment (BVQA) is a highly challenging task due to the intrinsic complexity of video content and visual distortions, especially given the high popularity of social media videos, which originate from a wide range of sources, and are often processed by various compression and enhancement algorithms. While recent BVQA and blind image quality assessment (BIQA) studies have made remarkable progress, their models typically perform well on the datasets they were trained on but generalize poorly to unseen videos, making them less effective for accurately evaluating the perceptual quality of diverse social media videos. In this paper, we propose Rich Quality-aware features enabled Video Quality Assessment (RQ-VQA), a simple yet effective method to enhance BVQA by leveraging rich quality-aware features extracted from off-the-shelf BIQA and BVQA models. Our approach exploits the expertise of existing quality assessment models within their trained domains to improve generalization. Specifically, we design a multi-source feature framework that integrates:(1) Learnable spatial features} from a base model fine-tuned on the target VQA dataset to capture domain-specific quality cues; (2) Temporal motion features from the fast pathway of SlowFast pre-trained on action recognition datasets to model motion-related distortions; (3) Spatial quality-aware features from BIQA models trained on diverse IQA datasets to enhance frame-level distortion representation; and (4) Spatiotemporal quality-aware features from a BVQA model trained on large-scale VQA datasets to jointly encode spatial structure and temporal dynamics. These features are concatenated and fed into a multi-layer perceptron (MLP) to regress them into quality scores. Experimental results demonstrate that our model achieves state-of-the-art performance on three public social media VQA datasets.

eess.IV↗

arXiv2025

PriceSeer: Evaluating Large Language Models in Real-Time Stock Prediction

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.

q-fin.ST↗

arXiv2025

LiveProteinBench: A Contamination-Free Benchmark for Assessing Models' Specialized Capabilities in Protein Science

In contrast to their remarkable performance on general knowledge QA, the true abilities of Large Language Models (LLMs) in tasks demanding deep, specialized reasoning, such as in protein biology, have yet to be thoroughly investigated. Current benchmarks suffer from critical deficiencies, such as data contamination due to outdated test sets, insufficient focus on essential protein-specific tasks, and a neglect of multimodal assessments. To resolve these issues, we introduce LiveProteinBench, a contamination-free, multimodal benchmark of 12 tasks for evaluating LLM performance on protein property and function prediction. Its central innovation lies in a test set composed exclusively of proteins validated after the start of 2025, guaranteeing that the data is novel to all tested models. We benchmarked a suite of prominent general-purpose LLMs and specialized biological LLMs using both unimodal and multimodal input schemes. Our results show that: 1) General-purpose proprietary large models demonstrate superior zero-shot performance when encountering new protein data, outperforming their open-source and domain-specific counterparts by over 20\% accuracy. 2) The effective use of multi-view structural information remains a significant challenge, as the inclusion of structural images often fails to provide a consistent benefit and can even degrade performance. This highlights the limitations of current models in effectively fusing information across different modalities. 3) Models' performance scales more directly with the computational cost during inference than with its parameter count, underscoring the critical role of Chain-of-Thought reasoning capabilities for protein-specific tasks. LiveProteinBench delineates the current performance frontiers for LLMs in bioinformatics and presents new challenges for the development of future multimodal foundation models for biology

q-bio.QM↗

arXiv2025

Generative Human-Object Interaction Detection via Differentiable Cognitive Steering of Multi-modal LLMs

Human-object interaction (HOI) detection aims to localize human-object pairs and the interactions between them. Existing methods operate under a closed-world assumption, treating the task as a classification problem over a small, predefined verb set, which struggles to generalize to the long-tail of unseen or ambiguous interactions in the wild. While recent multi-modal large language models (MLLMs) possess the rich world knowledge required for open-vocabulary understanding, they remain decoupled from existing HOI detectors since fine-tuning them is computationally prohibitive. To address these constraints, we propose \GRASP-HO}, a novel Generative Reasoning And Steerable Perception framework that reformulates HOI detection from the closed-set classification task to the open-vocabulary generation problem. To bridge the vision and cognitive, we first extract hybrid interaction representations, then design a lightweight learnable cognitive steering conduit (CSC) module to inject the fine-grained visual evidence into a frozen MLLM for effective reasoning. To address the supervision mismatch between classification-based HOI datasets and open-vocabulary generative models, we introduce a hybrid guidance strategy that coupling the language modeling loss and auxiliary classification loss, enabling discriminative grounding without sacrificing generative flexibility. Experiments demonstrate state-of-the-art closed-set performance and strong zero-shot generalization, achieving a unified paradigm that seamlessly bridges discriminative perception and generative reasoning for open-world HOI detection.

cs.CV↗

arXiv2025

Using GUI Agent for Electronic Design Automation

Graphical User Interface (GUI) agents adopt an end-to-end paradigm that maps a screenshot to an action sequence, thereby automating repetitive tasks in virtual environments. However, existing GUI agents are evaluated almost exclusively on commodity software such as Microsoft Word and Excel. Professional Computer-Aided Design (CAD) suites promise an order-of-magnitude higher economic return, yet remain the weakest performance domain for existing agents and are still far from replacing expert Electronic-Design-Automation (EDA) engineers. We therefore present the first systematic study that deploys GUI agents for EDA workflows. Our contributions are: (1) a large-scale dataset named GUI-EDA, including 5 CAD tools and 5 physical domains, comprising 2,000+ high-quality screenshot-answer-action pairs recorded by EDA scientists and engineers during real-world component design; (2) a comprehensive benchmark that evaluates 30+ mainstream GUI agents, demonstrating that EDA tasks constitute a major, unsolved challenge; and (3) an EDA-specialized metric named EDAgent, equipped with a reflection mechanism that achieves reliable performance on industrial CAD software and, for the first time, outperforms Ph.D. students majored in Electrical Engineering. This work extends GUI agents from generic office automation to specialized, high-value engineering domains and offers a new avenue for advancing EDA productivity. The dataset will be released at: https://github.com/aiben-ch/GUI-EDA.

cs.CV↗

arXiv2025

Embodied Image Compression

Image Compression for Machines (ICM) has emerged as a pivotal research direction in the field of visual data compression. However, with the rapid evolution of machine intelligence, the target of compression has shifted from task-specific virtual models to Embodied agents operating in real-world environments. To address the communication constraints of Embodied AI in multi-agent systems and ensure real-time task execution, this paper introduces, for the first time, the scientific problem of Embodied Image Compression. We establish a standardized benchmark, EmbodiedComp, to facilitate systematic evaluation under ultra-low bitrate conditions in a closed-loop setting. Through extensive empirical studies in both simulated and real-world settings, we demonstrate that existing Vision-Language-Action models (VLAs) fail to reliably perform even simple manipulation tasks when compressed below the Embodied bitrate threshold. We anticipate that EmbodiedComp will catalyze the development of domain-specific compression tailored for Embodied agents , thereby accelerating the Embodied AI deployment in the Real-world.

cs.CV↗

arXiv2025

Parameter-Efficient Fine-Tuning for Pre-Trained Vision Models: A Survey and Benchmark

Pre-trained vision models (PVMs) have demonstrated remarkable adaptability across a wide range of downstream vision tasks, showcasing exceptional performance. However, as these models scale to billions or even trillions of parameters, conventional full fine-tuning has become increasingly impractical due to its high computational and storage demands. To address these challenges, parameter-efficient fine-tuning (PEFT) has emerged as a promising alternative, aiming to achieve performance comparable to full fine-tuning while making minimal adjustments to the model parameters. This paper presents a comprehensive survey of the latest advancements in the visual PEFT field, systematically reviewing current methodologies and categorizing them into four primary categories: addition-based, partial-based, unified-based, and multi-task tuning. In addition, this paper offers an in-depth analysis of widely used visual datasets and real-world applications where PEFT methods have been successfully applied. Furthermore, this paper introduces the V-PEFT Bench, a unified benchmark designed to standardize the evaluation of PEFT methods across a diverse set of vision tasks, ensuring consistency and fairness in comparison. Finally, the paper outlines potential directions for future research to propel advances in the PEFT field. A comprehensive collection of resources is available at https://github.com/synbol/Awesome-Parameter-Efficient-Transfer-Learning.

cs.CV↗