Yichen Guo

Chengzhi Liu, Yuzhe Yang, Sophia Xiao Pu et al.

Multimodal large language models are increasingly deployed as long-horizon agents, where memory must do more than recall: it must track an evolving world, revise what has gone stale, and surface the right evidence at decision time. Existing benchmarks measure recall over static dialogue, collapse memory into a single end-of-task accuracy, and reduce visual observations to captions, leaving us unable to localize failures to writing, maintenance, retrieval, or use. The rise of agent harnesses that author their own memory sharpens this gap, since we have no principled way to compare hand-designed pipelines with self-managing alternatives. To close these gaps, we formulate multimodal agent memory as an Action-World Interaction Loop with an observable four-stage lifecycle, and instantiate it in WorldMemArena: 400 multi-session multimodal tasks spanning Lifelong Evolution (evolving personal and task states) and Agentic Execution (memory from real observations, actions, and feedback), annotated with gold memory points, updates, distractors, and evidence chains for stage-level diagnosis. This enables the first head-to-head comparison of long-context, manually designed (RAG and external memory systems), and harness-based memory agents. Results show that: (1) better memory writing and storage do not guarantee better performance; (2) multimodal memory still struggles to fully use visual evidence; (3) systems are unstable across domains and degrade on realistic agentic trajectories; and (4) harness memory is more flexible but remains costly and less reliable.

Hanze Li, Jinhao You, Yichen Guo et al.

Large Language Models (LLMs) have achieved strong performance across diverse natural language tasks, yet their outputs often suffer from hallucinations -- content that is misaligned with factual information. In this work, we conduct a comprehensive layer-wise analysis of the decoding process and reveal that hallucinations tend to originate from deeper decoder layers. To address this issue, we introduce \textbf{DeLask} (\textbf{De}coder \textbf{La}yer \textbf{Sk}ipping), a novel decoding framework that dynamically skips layers prone to producing hallucinations. DeLask leverages the theoretical insight that the forward computation of an $L$-layer Transformer is conditionally equivalent to $L$ steps of gradient descent. We define a \emph{driftance value} by computing the cosine similarity between gradients derived from consecutive decoder steps, identifying problematic layers when the descent direction reverses. Rather than discarding such layers entirely, DeLask partially aggregates their hidden states with preceding layers, thereby preserving consistency while suppressing erroneous signals. Extensive experiments across diverse LLMs and benchmarks demonstrate that DeLask consistently mitigates hallucinations and enhances overall reliability, providing a lightweight and generalizable decoding framework for improving the robustness of large-scale language models.

Junpeng Jing, Jiankun Li, Pengfei Xiong et al.

Correlation based stereo matching has achieved outstanding performance, which pursues cost volume between two feature maps. Unfortunately, current methods with a fixed model do not work uniformly well across various datasets, greatly limiting their real-world applicability. To tackle this issue, this paper proposes a new perspective to dynamically calculate correlation for robust stereo matching. A novel Uncertainty Guided Adaptive Correlation (UGAC) module is introduced to robustly adapt the same model for different scenarios. Specifically, a variance-based uncertainty estimation is employed to adaptively adjust the sampling area during warping operation. Additionally, we improve the traditional non-parametric warping with learnable parameters, such that the position-specific weights can be learned. We show that by empowering the recurrent network with the UGAC module, stereo matching can be exploited more robustly and effectively. Extensive experiments demonstrate that our method achieves state-of-the-art performance over the ETH3D, KITTI, and Middlebury datasets when employing the same fixed model over these datasets without any retraining procedure. To target real-time applications, we further design a lightweight model based on UGAC, which also outperforms other methods over KITTI benchmarks with only 0.6 M parameters.

Qunliang Xing, Mai Xu, Xin Deng et al.

Image defocus is inherent in the physics of image formation caused by the optical aberration of lenses, providing plentiful information on image quality. Unfortunately, existing quality enhancement approaches for compressed images neglect the inherent characteristic of defocus, resulting in inferior performance. This paper finds that in compressed images, significantly defocused regions have better compression quality, and two regions with different defocus values possess diverse texture patterns. These observations motivate our defocus-aware quality enhancement (DAQE) approach. Specifically, we propose a novel dynamic region-based deep learning architecture of the DAQE approach, which considers the regionwise defocus difference of compressed images in two aspects. (1) The DAQE approach employs fewer computational resources to enhance the quality of significantly defocused regions and more resources to enhance the quality of other regions; (2) The DAQE approach learns to separately enhance diverse texture patterns for regions with different defocus values, such that texture-specific enhancement can be achieved. Extensive experiments validate the superiority of our DAQE approach over state-of-the-art approaches in terms of quality enhancement and resource savings.

Xiaoye Liang, Lai Jiang, Minglang Qiao et al.

In recent years, the development of burst imaging technology has improved the capture and processing capabilities of visual data, enabling a wide range of applications. However, the redundancy in burst images leads to the increased storage and transmission demands, as well as reduced efficiency of downstream tasks. To address this, we propose a new task of Burst Image Quality Assessment (BuIQA), to evaluate the task-driven quality of each frame within a burst sequence, providing reasonable cues for burst image selection. Specifically, we establish the first benchmark dataset for BuIQA, consisting of $7,346$ burst sequences with $45,827$ images and $191,572$ annotated quality scores for multiple downstream scenarios. Inspired by the data analysis, a unified BuIQA framework is proposed to achieve an efficient adaption for BuIQA under diverse downstream scenarios. Specifically, a task-driven prompt generation network is developed with heterogeneous knowledge distillation, to learn the priors of the downstream task. Then, the task-aware quality assessment network is introduced to assess the burst image quality based on the task prompt. Extensive experiments across 10 downstream scenarios demonstrate the impressive BuIQA performance of the proposed approach, outperforming the state-of-the-art. Furthermore, it can achieve $0.33$ dB PSNR improvement in the downstream tasks of denoising and super-resolution, by applying our approach to select the high-quality burst frames.

Chengzhi Liu, Yichen Guo, Yepeng Liu et al.

LLM agents increasingly run inside execution harnesses that dispatch tools, allocate resources, and route messages between specialized components. However, a harness can return a correct, benign answer over a trajectory that accesses unauthorized resources or leaks context to the wrong agent. Output-level evaluation cannot see these failures, yet most safety benchmarks score only final outputs or terminal states, even though many violations occur mid-trajectory rather than at termination. The central question is whether the harness respects user intent, permission boundaries, and information-flow constraints throughout execution. To address this gap, we propose HarnessAudit, a framework that audits full execution trajectories across boundary compliance, execution fidelity, and system stability, with a focus on multi-agent harnesses where these risks are most pronounced. We further introduce HarnessAudit-Bench, a benchmark of 210 tasks across eight real-world domains, instantiated in both single-agent and multi-agent configurations with embedded safety constraints. Evaluating ten harness configurations across frontier models and three multi-agent frameworks, we find that: (i) task completion is misaligned with safe execution, and violations accumulate with trajectory length; (ii) safety risks vary across domains, task types, and agent roles; (iii) most violations concentrate in resource access and inter-agent information transfer; and (iv) multi-agent collaboration expands the safety risk surface, while harness design sets the upper bound of safe deployment.

Qianqi Yan, Yichen Guo, Ching-Chen Kuo et al.

Modern multimodal large language models (MLLMs) generate fluent responses from interleaved text, image, audio, and video inputs. However, identifying which input sources support each generated statement remains an open challenge. Existing attribution methods are primarily designed for classification settings, fixed prediction targets, or single-modality architectures, and do not naturally extend to autoregressive, decoder-only models performing open-ended multimodal generation. We introduce OmniTrace, a lightweight and model-agnostic framework that formalizes attribution as a generation-time tracing problem over the causal decoding process. OmniTrace provides a unified protocol that converts arbitrary token-level signals such as attention weights or gradient-based scores into coherent span-level, cross-modal explanations during decoding. It traces each generated token to multimodal inputs, aggregates signals into semantically meaningful spans, and selects concise supporting sources through confidence-weighted and temporally coherent aggregation, without retraining or supervision. Evaluations on Qwen2.5-Omni and MiniCPM-o-4.5 across visual, audio, and video tasks demonstrate that generation-aware span-level attribution produces more stable and interpretable explanations than naive self-attribution and embedding-based baselines, while remaining robust across multiple underlying attribution signals. Our results suggest that treating attribution as a structured generation-time tracing problem provides a scalable foundation for transparency in omni-modal language models.

Li Yang, Mai Xu, Shengxi Li et al.

Blind visual quality assessment (BVQA) on 360{\textdegree} video plays a key role in optimizing immersive multimedia systems. When assessing the quality of 360{\textdegree} video, human tends to perceive its quality degradation from the viewport-based spatial distortion of each spherical frame to motion artifact across adjacent frames, ending with the video-level quality score, i.e., a progressive quality assessment paradigm. However, the existing BVQA approaches for 360{\textdegree} video neglect this paradigm. In this paper, we take into account the progressive paradigm of human perception towards spherical video quality, and thus propose a novel BVQA approach (namely ProVQA) for 360{\textdegree} video via progressively learning from pixels, frames and video. Corresponding to the progressive learning of pixels, frames and video, three sub-nets are designed in our ProVQA approach, i.e., the spherical perception aware quality prediction (SPAQ), motion perception aware quality prediction (MPAQ) and multi-frame temporal non-local (MFTN) sub-nets. The SPAQ sub-net first models the spatial quality degradation based on spherical perception mechanism of human. Then, by exploiting motion cues across adjacent frames, the MPAQ sub-net properly incorporates motion contextual information for quality assessment on 360{\textdegree} video. Finally, the MFTN sub-net aggregates multi-frame quality degradation to yield the final quality score, via exploring long-term quality correlation from multiple frames. The experiments validate that our approach significantly advances the state-of-the-art BVQA performance on 360{\textdegree} video over two datasets, the code of which has been public in \url{https://github.com/yanglixiaoshen/ProVQA.}

Kai Tang, Jinhao You, Xiuqi Ge et al.

Despite the impressive capabilities of Large Vision-Language Models (LVLMs), they remain susceptible to hallucinations-generating content that is inconsistent with the input image. Existing training-free hallucination mitigation methods often suffer from unstable performance and high sensitivity to hyperparameter settings, limiting their practicality and broader adoption. In this paper, we propose a novel decoding mechanism, Decoding with Inter-layer Consistency via Layer Aggregation (DCLA), which requires no retraining, fine-tuning, or access to external knowledge bases. Specifically, our approach constructs a dynamic semantic reference by aggregating representations from previous layers, and corrects semantically deviated layers to enforce inter-layer consistency. The method allows DCLA to robustly mitigate hallucinations across multiple LVLMs. Experiments on hallucination benchmarks such as MME and POPE demonstrate that DCLA effectively reduces hallucinations while enhancing the reliability and performance of LVLMs.

Chenxi Li, Yichen Guo, Benfang Qian et al.

Large Vision-Language Models (LVLMs) have achieved impressive performance in multimodal tasks, but they still suffer from hallucinations, i.e., generating content that is grammatically accurate but inconsistent with visual inputs. In this work, we introduce a novel map-level perspective to mitigate hallucinations in LVLMs, interpreting the hidden states of the model as a 2D semantic map. We observe that factual information is widely distributed across this map, extending beyond the localized inter- or intra-layer regions targeted by most existing methods (e.g., contrastive decoding and layer-wise consistency). Building on this insight, we propose Map-Level Attention Processing (MAP), a training-free decoding method that effectively leverages factual information through attention-based map-level operations to improve factual consistency. Specifically, we employ Layer-Wise Criss-Cross Attention to progressively refine token representations at each decoding layer by aggregating tokens from both inter- and intra-layer dimensions. Additionally, a Global-Local Logit Fusion mechanism combines logits obtained before and after global attention to further refine predictions and improve accuracy. Our method consistently improves the truthfulness and performance of LVLMs across benchmarks, such as POPE, MME, and MMHal-Bench, demonstrating the potential of the map-level decoding strategy.

Yichen Guo, Hanze Li, Zonghao Zhang et al.

Although large vision-language models (LVLMs) leverage rich visual token representations to achieve strong performance on multimodal tasks, these tokens also introduce significant computational overhead during inference. Existing training-free token pruning methods typically adopt a single-stage strategy, focusing either on visual self-attention or visual-textual cross-attention. However, such localized perspectives often overlook the broader information flow across the model, leading to substantial performance degradation, especially under high pruning ratios. In this work, we propose STAR (Stage-wise Attention-guided token Reduction), a training-free, plug-and-play framework that approaches token pruning from a global perspective. Instead of pruning at a single point, STAR performs attention-guided reduction in two complementary stages: an early-stage pruning based on visual self-attention to remove redundant low-level features, and a later-stage pruning guided by cross-modal attention to discard task-irrelevant tokens. This holistic approach allows STAR to significantly reduce computational cost while better preserving task-critical information. Extensive experiments across multiple LVLM architectures and benchmarks show that STAR achieves strong acceleration while maintaining comparable, and in some cases even improved performance.