arXiv subjects

Yiwu Zhong

Publications and source records attributed to Yiwu Zhong.

At least 19 recordsLinked to original sources

arXiv2026

OmniReasoning: Pushing the Limits of Audio-Visual Joint Reasoning

Recent advances have enabled unified omni-modal models in understanding audio, vision, and language. However, existing benchmarks, training data, and learning methods largely treat the modalities independently, leaving the capability of audio-visual joint reasoning poorly evaluated and insufficiently elicited. We address this gap with a benchmark, data engine, and learning method. First, we introduce OmniReasoningBench, a benchmark where both audio and visual evidence are indispensable. It comprises 1,150 multiple-choice and open-ended questions across two tasks, reasoning over video and reasoning beyond video. Second, we develop a data engine OmniQA. It automatically constructs evidence-grounded QA pairs that explicitly necessitate audio-visual joint reasoning, together with time-stamped clue chains that guide the annotation of thinking process. Besides our benchmark, this engine produces training data OmniReasoning-SFT-112K and OmniReasoning-RL-19K. Finally, we propose an on-policy self-distillation method Modality-Factored Self-Distillation (MFSD). It evaluates each sampled response under modality-specific clue contexts, disentangling the contributions of individual clues and their cross-modal interactions for token-level credit assignment. With our training data and learning method, our model OmniReasoning-30B-A3B achieves 50.0% on OmniVideoBench and 42.5% on OmniReasoningBench, improving the base model Qwen3-Omni-30B-A3B-Thinking by 12.8 and 9.3 percentage points, respectively. Moreover, it delivers substantial gains on general and long-video benchmarks, including Video-MME-v2. We hope our work offers a solid step for facilitating future research in omni-modal joint reasoning.

cs.CV↗

arXiv2026

OmniEcho: Audio-Visual Spatial Understanding for Omni-Modal Embodied Agents

Humans can effortlessly localize the direction of a sound source and integrate it with visual cues for reasoning, yet this remains challenging for embodied agents. In particular, it is still unclear how to effectively evaluate and model spatial audio understanding in embodied settings. To address this gap, we introduce \textbf{OmniEchoBench}, a unified benchmark for spatial audio-visual perception and audio-vision-language navigation. OmniEchoBench comprises six tasks over 197 real-world spatial audio-visual scenes, 2,972 question-answer pairs, and 900 navigation samples with first-order ambisonics (FOA) audio collected from 30 real-world environments. To enable scalable training supervision, we develop a controllable rendering pipeline for spatial audio. It preserves geometric consistency among sound sources, visual observations, and agent trajectories. Building on this, we propose \textbf{OmniEcho}, a spatially aware omni-modal model. It introduces an FOA spatial encoder alongside a pretrained semantic audio pathway. Extensive experiments show that OmniEcho achieves state-of-the-art performance on spatial audio-visual perception. For our sound-guided navigation, OmniEcho reaches a performance level close to that of traditional vision-language navigation. These results demonstrate that spatial audio can serve as a valuable signal for embodied scene reasoning and navigation, while also highlighting fine-grained spatial localization and distance estimation as important open challenges. Our code and data will be available in https://github.com/PKU-VaLuE-Lab/OmniEcho/tree/main

cs.SD↗

arXiv2026

SinkPruner: Sink-Free Visual Token Pruning for Multimodal Large Language Models

Despite their strong multimodal understanding ability, multimodal large language models (MLLMs) incur substantial computational overhead when processing long visual token sequences. To reduce inference costs, recent studies have explored visual token pruning through vision-centric or text-guided strategies. However, these methods often overlook high-norm outlier tokens, i.e., tokens with abnormally large feature norms, leading to suboptimal pruning decisions. In this work, we show that such high-norm outlier tokens are highly redundant in both feature and spatial dimensions, yet are often mistakenly preserved as informative cues by existing methods. Motivated by this observation, we propose SinkPruner, a training-free visual token pruning framework for efficient MLLM inference. SinkPruner follows a coarse-to-fine design with two key modules: a visual sanitizer that filters high-norm redundancies and alleviates attention sink and attention dispersion, and a text-guided pruner that further retains tokens semantically aligned with the text query. Extensive experiments on twelve image-language and four video-language benchmarks demonstrate the effectiveness, efficiency, and generalizability of our framework. Notably, SinkPruner preserves 96.5% (91.8%) of the original performance of LLaVA-1.5 (Qwen2.5-VL) under an 89% token reduction. Experiments further indicate that our visual sanitizer exhibits promising transferability in enhancing the performance of existing pruning methods. Our code is available at https://github.com/LaVi-Lab/SinkPruner.

cs.CV↗

arXiv2026

Data Pyramid for Embodied Manipulation: A Survey

Multimodal foundation models learned to see and to speak by consuming the whole internet. Embodied agents admit no such shortcut, since they require data that couple observations with physical states and actions. These signals can be provided, to varying degrees, by multiple data sources. In this work, we organize the embodied data ecosystem as a "pyramid" spanning five complementary sources: real-robot data, UMI-style data, egocentric and exocentric data, simulation data, and general vision-language data. We organize the pyramid around the tension between scalability and robot alignment, and further characterize each source in terms of data quality, diversity, reusability, and physical fidelity. We then analyze recent embodied foundation models through the lens of their data recipes, examining how different sources are selected, aligned, and mixed during pretraining. For embodied brain models, vision-language-action models, and world-action models alike, we relate data composition to capabilities in perception, reasoning, planning, action generation, and world prediction. We close by discussing six open challenges: building large-scale tactile datasets, collecting failure and recovery data, developing scalable data-collection pipelines, aligning actions across embodiments, leveraging egocentric data for dexterous manipulation, and designing principled data recipes for robot learning. We hope this work paves the foundation for the design of next-generation embodied systems.

cs.RO↗

arXiv2026

M$^3$Eval: Multi-Modal Memory Evaluation through Cognitively-Grounded Video Tasks

As multi-modal models advance towards long-form video understanding, memory emerges as a critical capability. Despite substantial efforts in developing video datasets and benchmarks, existing works primarily focus on perception and reasoning, without systematically evaluating memory: what models retain, how faithfully information is preserved, and how robust memory remains under interference. To address this gap, we introduce M$^3$Eval, the first comprehensive evaluation framework and benchmark for probing different memory dimensions in multi-modal models. Grounded in cognitive psychology, our design features carefully constructed tasks that isolate key aspects of memory. Leveraging M$^3$Eval, we conduct extensive experiments across representative multi-modal models, revealing consistent weaknesses and distinctive behaviors. We find that models struggle to maintain disentangled representations when processing parallel video streams, exhibit interference patterns differing substantially from those observed in human memory, ground memory sources more reliably in the spatial domain than the temporal domain, and demonstrate limited symbolic memory. Collectively, our benchmark provides a valuable resource for future research, while our findings highlight memory as a fundamental yet underexplored capability and offer insights for designing more effective memory mechanisms in multi-modal models. Our code and dataset are available at https://pku-value-lab.github.io/m3eval-homepage.

cs.CV↗

arXiv2026

Preserve Support, Not Correspondence: Dynamic Routing for Offline Reinforcement Learning

One-step offline RL actors are attractive because they avoid backpropagating through long iterative samplers and keep inference cheap, but they still have to improve under a critic without drifting away from actions that the dataset can support. In recent one-step extraction pipelines, a strong iterative teacher provides one target action for each latent draw, and the same student output is asked to do both jobs: move toward higher Q and stay near that paired endpoint. If those two directions disagree, the loss resolves them as a compromise on that same sample, even when a nearby better action remains locally supported by the data. We propose DROL, a latent-conditioned one-step actor trained with top-1 dynamic routing. For each state, the actor samples $K$ candidate actions from a bounded latent prior, assigns each dataset action to its nearest candidate, and updates only that winner with Behavior Cloning and critic guidance. Because the routing is recomputed from the current candidate geometry, ownership of a supported region can shift across candidates over the course of learning. This gives a one-step actor room to make local improvements that pointwise extraction struggles to capture, while retaining single-pass inference at test time. On OGBench and D4RL, DROL is competitive with the one-step FQL baseline, improving many OGBench task groups while remaining strong on both AntMaze and Adroit. Project page: https://muzhancun.github.io/preprints/DROL.

cs.LG↗

arXiv2026

DOSE: Data Selection for Multi-Modal LLMs via Off-the-Shelf Models

High-quality and diverse multimodal data are essential for improving vision-language models (VLMs), yet existing datasets often contain noisy, redundant, and poorly aligned samples. To address these problems, data filtering is commonly used to enhance the efficiency and performance of multimodal learning, but it introduces extra computational cost because filtering models are usually trained on the same data they are meant to screen. To reduce this cost, we study DOSE, which explores whether off-the-shelf pretrained models that have never seen the target data can be used to select training samples for larger and stronger multimodal models without any task-specific training. Even without fine-tuning, these models can effectively assess text quality and image-text alignment to guide data selection. Based on this, we build a joint quality-alignment distribution and apply adaptive weighted sampling to select informative samples while maintaining long-tail diversity. This approach enhances data diversity, enabling models trained on DOSE-filtered data to match or surpass those trained on the full dataset on standard VQA and math benchmarks. Extensive experiments demonstrate its effectiveness, efficiency, and scalability.

cs.CV↗

arXiv2026

TextShield-R1: Reinforced Reasoning for Tampered Text Detection

The growing prevalence of tampered images poses serious security threats, highlighting the urgent need for reliable detection methods. Multimodal large language models (MLLMs) demonstrate strong potential in analyzing tampered images and generating interpretations. However, they still struggle with identifying micro-level artifacts, exhibit low accuracy in localizing tampered text regions, and heavily rely on expensive annotations for forgery interpretation. To this end, we introduce TextShield-R1, the first reinforcement learning based MLLM solution for tampered text detection and reasoning. Specifically, our approach introduces Forensic Continual Pre-training, an easy-to-hard curriculum that well prepares the MLLM for tampered text detection by harnessing the large-scale cheap data from natural image forensic and OCR tasks. During fine-tuning, we perform Group Relative Policy Optimization with novel reward functions to reduce annotation dependency and improve reasoning capabilities. At inference time, we enhance localization accuracy via OCR Rectification, a method that leverages the MLLM's strong text recognition abilities to refine its predictions. Furthermore, to support rigorous evaluation, we introduce the Text Forensics Reasoning (TFR) benchmark, comprising over 45k real and tampered images across 16 languages, 10 tampering techniques, and diverse domains. Rich reasoning-style annotations are included, allowing for comprehensive assessment. Our TFR benchmark simultaneously addresses seven major limitations of existing benchmarks and enables robust evaluation under cross-style, cross-method, and cross-language conditions. Extensive experiments demonstrate that TextShield-R1 significantly advances the state of the art in interpretable tampered text detection.

cs.CV↗

arXiv2025

Webly-Supervised Image Manipulation Localization via Category-Aware Auto-Annotation

Images manipulated by image editing tools can mislead viewers and pose significant risks to social security. However, accurately localizing manipulated image regions remains challenging due to the severe scarcity of high-quality annotated data, which is laborious to create. To address this, we propose a novel approach that mitigates data scarcity by leveraging readily available web data. We utilize a large collection of manually forged images from the web, as well as automatically generated annotations derived from a simpler auxiliary task, constrained image manipulation localization.Specifically, we introduce CAAAv2, a novel auto-annotation framework that operates on a category-aware, prior-feature-denoising paradigm that notably reduces task complexity. To further ensure annotation reliability, we propose QES, a novel metric that filters out low-quality annotations. Combining CAAAv2 and QES, we construct MIMLv2, a large-scale, diverse, and high-quality dataset containing 246,212 manually forged images with pixel-level mask annotations. This is over 120 times larger than existing handcrafted datasets like IMD20. Additionally, we introduce Object Jitter, a technique that further enhances model training by generating high-quality manipulation artifacts. Building on these advances, we develop Web-IML, a new model designed to effectively leverage web-scale supervision for the task of image manipulation localization. Extensive experiments demonstrate that our approach substantially alleviates the data scarcity problem and significantly improves the performance of various models on multiple real-world forgery benchmarks. With the proposed web supervision, our Web-IML achieves a striking performance gain of 31% and surpasses the previous state-of-the-art SparseViT by 21.6 average IoU points. The dataset and code will be released at https://github.com/qcf-568/MIML.

cs.CV↗

arXiv2025

Rethinking Chain-of-Thought Reasoning for Videos

Chain-of-thought (CoT) reasoning has been highly successful in solving complex tasks in natural language processing, and recent multimodal large language models (MLLMs) have extended this paradigm to video reasoning. However, these models typically build on lengthy reasoning chains and large numbers of input visual tokens. Motivated by empirical observations from our benchmark study, we hypothesize that concise reasoning combined with a reduced set of visual tokens can be sufficient for effective video reasoning. To evaluate this hypothesis, we design and validate an efficient post-training and inference framework that enhances a video MLLM's reasoning capability. Our framework enables models to operate on compressed visual tokens and generate brief reasoning traces prior to answering. The resulting models achieve substantially improved inference efficiency, deliver competitive performance across diverse benchmarks, and avoid reliance on manual CoT annotations or supervised fine-tuning. Collectively, our results suggest that long, human-like CoT reasoning may not be necessary for general video reasoning, and that concise reasoning can be both effective and efficient. Our code will be released at https://github.com/LaVi-Lab/Rethink_CoT_Video.

cs.CV↗

arXiv2025

Fine-grained Spatiotemporal Grounding on Egocentric Videos

Spatiotemporal video grounding aims to localize target entities in videos based on textual queries. While existing research has made significant progress in exocentric videos, the egocentric setting remains relatively underexplored, despite its growing importance in applications such as augmented reality and robotics. In this work, we conduct a systematic analysis of the discrepancies between egocentric and exocentric videos, revealing key challenges such as shorter object durations, sparser trajectories, smaller object sizes, and larger positional shifts. To address these challenges, we introduce EgoMask, the first pixel-level benchmark for fine-grained spatiotemporal grounding in egocentric videos. It is constructed by our proposed automatic annotation pipeline, which annotates referring expressions and object masks across short-, medium-, and long-term videos. Additionally, we create EgoMask-Train, a large-scale training dataset to facilitate model development. Experiments demonstrate that the state-of-the-art spatiotemporal grounding models perform poorly on our benchmark EgoMask, but fine-tuning on EgoMask-Train yields significant improvements, while preserving performance on exocentric datasets. Our work thus provides essential resources and insights for advancing egocentric video understanding. Our code is available at https://github.com/LaVi-Lab/EgoMask .

cs.CV↗

arXiv2025

AIM: Adaptive Inference of Multi-Modal LLMs via Token Merging and Pruning

Large language models (LLMs) have enabled the creation of multi-modal LLMs that exhibit strong comprehension of visual data such as images and videos. However, these models usually rely on extensive visual tokens from visual encoders, leading to high computational demands, which limits their applicability in resource-constrained environments and for long-context tasks. In this work, we propose a training-free adaptive inference method for multi-modal LLMs that can accommodate a broad range of efficiency requirements with a minimum performance drop. Our method consists of a) iterative token merging based on embedding similarity before LLMs, and b) progressive token pruning within LLM layers based on multi-modal importance. With a minimalist design, our method can be applied to both video and image LLMs. Extensive experiments on diverse video and image benchmarks demonstrate that our method substantially reduces computation load (e.g., a $\textbf{7-fold}$ reduction in FLOPs) while preserving the performance of video and image LLMs. Further, at a similar computational cost, our method outperforms the state-of-the-art methods in long video understanding (e.g., $\textbf{+4.6}$ on MLVU). Additionally, our in-depth analysis provides insights into token redundancy and LLM layer behaviors, offering guidance for future research in designing efficient multi-modal LLMs. Our code is available at https://github.com/LaVi-Lab/AIM.

cs.CV↗

arXiv2025

Omni-IML: Towards Unified Image Manipulation Localization

Existing Image Manipulation Localization (IML) methods mostly rely heavily on task-specific designs, making them perform well only on the target IML task, while joint training on multiple IML tasks causes significant performance degradation, hindering real applications. To this end, we propose Omni-IML, the first generalist model designed to unify IML across diverse tasks. Specifically, Omni-IML achieves generalization through three key components: (1) a Modal Gate Encoder, which adaptively selects the optimal encoding modality per sample, (2) a Dynamic Weight Decoder, which dynamically adjusts decoder filters to the task at hand, and (3) an Anomaly Enhancement module that leverages box supervision to highlight the tampered regions and facilitate the learning of task-agnostic features. Beyond localization, to support interpretation of the tampered images, we construct Omni-273k, a large high-quality dataset that includes natural language descriptions of tampered artifact. It is annotated through our automatic, chain-of-thoughts annotation technique. We also design a simple-yet-effective interpretation module to better utilize these descriptive annotations. Our extensive experiments show that our single Omni-IML model achieves state-of-the-art performance across all four major IML tasks, providing a valuable solution for practical deployment and a promising direction of generalist models in image forensics. Our code and dataset will be publicly available.

cs.CV↗

arXiv2025

PAVE: Patching and Adapting Video Large Language Models

Pre-trained video large language models (Video LLMs) exhibit remarkable reasoning capabilities, yet adapting these models to new tasks involving additional modalities or data types (e.g., audio or 3D information) remains challenging. In this paper, we present PAVE, a flexible framework for adapting pre-trained Video LLMs to downstream tasks with side-channel signals, such as audio, 3D cues, or multi-view videos. PAVE introduces lightweight adapters, referred to as "patches," which add a small number of parameters and operations to a base model without modifying its architecture or pre-trained weights. In doing so, PAVE can effectively adapt the pre-trained base model to support diverse downstream tasks, including audio-visual question answering, 3D reasoning, multi-view video recognition, and high frame rate video understanding. Across these tasks, PAVE significantly enhances the performance of the base model, surpassing state-of-the-art task-specific models while incurring a minor cost of ~0.1% additional FLOPs and parameters. Further, PAVE supports multi-task learning and generalizes well across different Video LLMs. Our code is available at https://github.com/dragonlzm/PAVE.

cs.CV↗

arXiv2025

Revisiting Tampered Scene Text Detection in the Era of Generative AI

The rapid advancements of generative AI have fueled the potential of generative text image editing, meanwhile escalating the threat of misinformation spreading. However, existing forensics methods struggle to detect unseen forgery types that they have not been trained on, underscoring the need for a model capable of generalized detection of tampered scene text. To tackle this, we propose a novel task: open-set tampered scene text detection, which evaluates forensics models on their ability to identify both seen and previously unseen forgery types. We have curated a comprehensive, high-quality dataset, featuring the texts tampered by eight text editing models, to thoroughly assess the open-set generalization capabilities. Further, we introduce a novel and effective training paradigm that subtly alters the texture of selected texts within an image and trains the model to identify these regions. This approach not only mitigates the scarcity of high-quality training data but also enhances models' fine-grained perception and open-set generalization abilities. Additionally, we present DAF, a novel framework that improves open-set generalization by distinguishing between the features of authentic and tampered text, rather than focusing solely on the tampered text's features. Our extensive experiments validate the remarkable efficacy of our methods. For example, our zero-shot performance can even beat the previous state-of-the-art full-shot model by a large margin. Our dataset and code are available at https://github.com/qcf-568/OSTF.

cs.CV↗

arXiv2024

TemporalBench: Benchmarking Fine-grained Temporal Understanding for Multimodal Video Models

Understanding fine-grained temporal dynamics is crucial for multimodal video comprehension and generation. Due to the lack of fine-grained temporal annotations, existing video benchmarks mostly resemble static image benchmarks and are incompetent at evaluating models for temporal understanding. In this paper, we introduce TemporalBench, a new benchmark dedicated to evaluating fine-grained temporal understanding in videos. TemporalBench consists of ~10K video question-answer pairs, derived from ~2K high-quality human annotations detailing the temporal dynamics in video clips. As a result, our benchmark provides a unique testbed for evaluating various temporal understanding and reasoning abilities such as action frequency, motion magnitude, event order, etc. Moreover, it enables evaluations on various tasks like both video question answering and captioning, both short and long video understanding, as well as different models such as multimodal video embedding models and text generation models. Results show that state-of-the-art models like GPT-4o achieve only 38.5% question answering accuracy on TemporalBench, demonstrating a significant gap (~30%) between humans and AI in temporal understanding. Furthermore, we notice a critical pitfall for multi-choice QA where LLMs can detect the subtle changes in negative captions and find a centralized description as a cue for its prediction, where we propose Multiple Binary Accuracy (MBA) to correct such bias. We hope that TemporalBench can foster research on improving models' temporal reasoning capabilities. Both dataset and evaluation code will be made available.

cs.CV↗

arXiv2024

Enhancing Temporal Modeling of Video LLMs via Time Gating

Video Large Language Models (Video LLMs) have achieved impressive performance on video-and-language tasks, such as video question answering. However, most existing Video LLMs neglect temporal information in video data, leading to struggles with temporal-aware video understanding. To address this gap, we propose a Time Gating Video LLM (TG-Vid) designed to enhance temporal modeling through a novel Time Gating module (TG). The TG module employs a time gating mechanism on its sub-modules, comprising gating spatial attention, gating temporal attention, and gating MLP. This architecture enables our model to achieve a robust understanding of temporal information within videos. Extensive evaluation of temporal-sensitive video benchmarks (i.e., MVBench, TempCompass, and NExT-QA) demonstrates that our TG-Vid model significantly outperforms the existing Video LLMs. Further, comprehensive ablation studies validate that the performance gains are attributed to the designs of our TG module. Our code is available at https://github.com/LaVi-Lab/TG-Vid.

cs.CV↗

arXiv2024

Beyond Embeddings: The Promise of Visual Table in Visual Reasoning

Visual representation learning has been a cornerstone in computer vision, involving typical forms such as visual embeddings, structural symbols, and text-based representations. Despite the success of CLIP-type visual embeddings, they often lack access to world knowledge critical for visual reasoning. In this work, we propose Visual Table, a novel form of visual representation tailored for visual reasoning. Visual tables are constructed as hierarchical descriptions of visual scenes, featuring a scene description and multiple object-centric descriptions covering categories, attributes, and knowledge. Thanks to the structural and textual formats, visual tables offer unique advantages over mere visual embeddings, such as interpretability and controllable editing. Furthermore, they deliver instance-level world knowledge and detailed attributes that are essential for visual reasoning. To create visual tables, we develop a generator trained on the dataset with collected, small-scale annotations. Extensive results on 11 visual reasoning benchmarks demonstrate that the generated visual tables significantly outperform previous structural and text-based representations. Moreover, they consistently enhance state-of-the-art multimodal large language models across diverse benchmarks, showcasing their potential for advancing visual reasoning tasks. Our code is available at https://github.com/LaVi-Lab/Visual-Table.

cs.CV↗