Fanqing Meng

Peng Xu, Wenqi Shao, Kaipeng Zhang et al. · pku

Large Vision-Language Models (LVLMs) have recently played a dominant role in multimodal vision-language learning. Despite the great success, it lacks a holistic evaluation of their efficacy. This paper presents a comprehensive evaluation of publicly available large multimodal models by building a LVLM evaluation Hub (LVLM-eHub). Our LVLM-eHub consists of $8$ representative LVLMs such as InstructBLIP and MiniGPT-4, which are thoroughly evaluated by a quantitative capability evaluation and an online arena platform. The former evaluates $6$ categories of multimodal capabilities of LVLMs such as visual question answering and embodied artificial intelligence on $47$ standard text-related visual benchmarks, while the latter provides the user-level evaluation of LVLMs in an open-world question-answering scenario. The study reveals several innovative findings. First, instruction-tuned LVLM with massive in-domain data such as InstructBLIP heavily overfits many existing tasks, generalizing poorly in the open-world scenario. Second, instruction-tuned LVLM with moderate instruction-following data may result in object hallucination issues (i.e., generate objects that are inconsistent with target images in the descriptions). It either makes the current evaluation metric such as CIDEr for image captioning ineffective or generates wrong answers. Third, employing a multi-turn reasoning evaluation framework can mitigate the issue of object hallucination, shedding light on developing an effective pipeline for LVLM evaluation. The findings provide a foundational framework for the conception and assessment of innovative strategies aimed at enhancing zero-shot multimodal techniques. Our LVLM-eHub will be available at https://github.com/OpenGVLab/Multi-Modality-Arena

Fanqing Meng, Jin Wang, Chuanhao Li et al.

The capability to process multiple images is crucial for Large Vision-Language Models (LVLMs) to develop a more thorough and nuanced understanding of a scene. Recent multi-image LVLMs have begun to address this need. However, their evaluation has not kept pace with their development. To fill this gap, we introduce the Multimodal Multi-image Understanding (MMIU) benchmark, a comprehensive evaluation suite designed to assess LVLMs across a wide range of multi-image tasks. MMIU encompasses 7 types of multi-image relationships, 52 tasks, 77K images, and 11K meticulously curated multiple-choice questions, making it the most extensive benchmark of its kind. Our evaluation of 24 popular LVLMs, including both open-source and proprietary models, reveals significant challenges in multi-image comprehension, particularly in tasks involving spatial understanding. Even the most advanced models, such as GPT-4o, achieve only 55.7% accuracy on MMIU. Through multi-faceted analytical experiments, we identify key performance gaps and limitations, providing valuable insights for future model and data improvements. We aim for MMIU to advance the frontier of LVLM research and development, moving us toward achieving sophisticated multimodal multi-image user interactions.

Fanqing Meng, Wenqi Shao, Zhanglin Peng et al.

This paper investigates an under-explored but important problem: given a collection of pre-trained neural networks, predicting their performance on each multi-modal task without fine-tuning them, such as image recognition, referring, captioning, visual question answering, and text question answering. A brute-force approach is to finetune all models on all target datasets, bringing high computational costs. Although recent-advanced approaches employed lightweight metrics to measure models' transferability,they often depend heavily on the prior knowledge of a single task, making them inapplicable in a multi-modal multi-task scenario. To tackle this issue, we propose an efficient multi-task model selector (EMMS), which employs large-scale foundation models to transform diverse label formats such as categories, texts, and bounding boxes of different downstream tasks into a unified noisy label embedding. EMMS can estimate a model's transferability through a simple weighted linear regression, which can be efficiently solved by an alternating minimization algorithm with a convergence guarantee. Extensive experiments on 5 downstream tasks with 24 datasets show that EMMS is fast, effective, and generic enough to assess the transferability of pre-trained models, making it the first model selection method in the multi-task scenario. For instance, compared with the state-of-the-art method LogME enhanced by our label embeddings, EMMS achieves 9.0\%, 26.3\%, 20.1\%, 54.8\%, 12.2\% performance gain on image recognition, referring, captioning, visual question answering, and text question answering, while bringing 5.13x, 6.29x, 3.59x, 6.19x, and 5.66x speedup in wall-clock time, respectively. The code is available at https://github.com/OpenGVLab/Multitask-Model-Selector.

Wenqi Shao, Meng Lei, Yutao Hu et al. · pku

Recent advancements in Large Vision-Language Models (LVLMs) have demonstrated significant progress in tackling complex multimodal tasks. Among these cutting-edge developments, Google's Bard stands out for its remarkable multimodal capabilities, promoting comprehensive comprehension and reasoning across various domains. This work presents an early and holistic evaluation of LVLMs' multimodal abilities, with a particular focus on Bard, by proposing a lightweight variant of LVLM-eHub, named Tiny LVLM-eHub. In comparison to the vanilla version, Tiny LVLM-eHub possesses several appealing properties. Firstly, it provides a systematic assessment of six categories of multimodal capabilities, including visual perception, visual knowledge acquisition, visual reasoning, visual commonsense, object hallucination, and embodied intelligence, through quantitative evaluation of $42$ standard text-related visual benchmarks. Secondly, it conducts an in-depth analysis of LVLMs' predictions using the ChatGPT Ensemble Evaluation (CEE), which leads to a robust and accurate evaluation and exhibits improved alignment with human evaluation compared to the word matching approach. Thirdly, it comprises a mere $2.1$K image-text pairs, facilitating ease of use for practitioners to evaluate their own offline LVLMs. Through extensive experimental analysis, this study demonstrates that Bard outperforms previous LVLMs in most multimodal capabilities except object hallucination, to which Bard is still susceptible. Tiny LVLM-eHub serves as a baseline evaluation for various LVLMs and encourages innovative strategies aimed at advancing multimodal techniques. Our project is publicly available at \url{https://github.com/OpenGVLab/Multi-Modality-Arena}.

Kimi Team, Tongtong Bai, Yifan Bai et al.

We introduce Kimi K2.5, an open-source multimodal agentic model designed to advance general agentic intelligence. K2.5 emphasizes the joint optimization of text and vision so that two modalities enhance each other. This includes a series of techniques such as joint text-vision pre-training, zero-vision SFT, and joint text-vision reinforcement learning. Building on this multimodal foundation, K2.5 introduces Agent Swarm, a self-directed parallel agent orchestration framework that dynamically decomposes complex tasks into heterogeneous sub-problems and executes them concurrently. Extensive evaluations show that Kimi K2.5 achieves state-of-the-art results across various domains including coding, vision, reasoning, and agentic tasks. Agent Swarm also reduces latency by up to $4.5\times$ over single-agent baselines. We release the post-trained Kimi K2.5 model checkpoint to facilitate future research and real-world applications of agentic intelligence.

Kimi Team, Yu Zhang, Zongyu Lin et al.

We introduce Kimi Linear, a hybrid linear attention architecture that, for the first time, outperforms full attention under fair comparisons across various scenarios -- including short-context, long-context, and reinforcement learning (RL) scaling regimes. At its core lies Kimi Delta Attention (KDA), an expressive linear attention module that extends Gated DeltaNet with a finer-grained gating mechanism, enabling more effective use of limited finite-state RNN memory. Our bespoke chunkwise algorithm achieves high hardware efficiency through a specialized variant of the Diagonal-Plus-Low-Rank (DPLR) transition matrices, which substantially reduces computation compared to the general DPLR formulation while remaining more consistent with the classical delta rule. We pretrain a Kimi Linear model with 3B activated parameters and 48B total parameters, based on a layerwise hybrid of KDA and Multi-Head Latent Attention (MLA). Our experiments show that with an identical training recipe, Kimi Linear outperforms full MLA with a sizeable margin across all evaluated tasks, while reducing KV cache usage by up to 75% and achieving up to 6 times decoding throughput for a 1M context. These results demonstrate that Kimi Linear can be a drop-in replacement for full attention architectures with superior performance and efficiency, including tasks with longer input and output lengths. To support further research, we open-source the KDA kernel and vLLM implementations, and release the pre-trained and instruction-tuned model checkpoints.

Kimi Team, Guangyu Chen, Yu Zhang et al.

Residual connections with PreNorm are standard in modern LLMs, yet they accumulate all layer outputs with fixed unit weights. This uniform aggregation causes uncontrolled hidden-state growth with depth, progressively diluting each layer's contribution. We propose Attention Residuals (AttnRes), which replaces this fixed accumulation with softmax attention over preceding layer outputs, allowing each layer to selectively aggregate earlier representations with learned, input-dependent weights. To address the memory and communication overhead of attending over all preceding layer outputs for large-scale model training, we introduce Block AttnRes, which partitions layers into blocks and attends over block-level representations, reducing the memory footprint while preserving most of the gains of full AttnRes. Combined with cache-based pipeline communication and a two-phase computation strategy, Block AttnRes becomes a practical drop-in replacement for standard residual connections with minimal overhead. Scaling law experiments confirm that the improvement is consistent across model sizes, and ablations validate the benefit of content-dependent depth-wise selection. We further integrate AttnRes into the Kimi Linear architecture (48B total / 3B activated parameters) and pre-train on 1.4T tokens, where AttnRes mitigates PreNorm dilution, yielding more uniform output magnitudes and gradient distribution across depth, and improves downstream performance across all evaluated tasks.

Kaining Ying, Fanqing Meng, Jin Wang et al.

Large Vision-Language Models (LVLMs) show significant strides in general-purpose multimodal applications such as visual dialogue and embodied navigation. However, existing multimodal evaluation benchmarks cover a limited number of multimodal tasks testing rudimentary capabilities, falling short in tracking LVLM development. In this study, we present MMT-Bench, a comprehensive benchmark designed to assess LVLMs across massive multimodal tasks requiring expert knowledge and deliberate visual recognition, localization, reasoning, and planning. MMT-Bench comprises $31,325$ meticulously curated multi-choice visual questions from various multimodal scenarios such as vehicle driving and embodied navigation, covering $32$ core meta-tasks and $162$ subtasks in multimodal understanding. Due to its extensive task coverage, MMT-Bench enables the evaluation of LVLMs using a task map, facilitating the discovery of in- and out-of-domain tasks. Evaluation results involving $30$ LVLMs such as the proprietary GPT-4V, GeminiProVision, and open-sourced InternVL-Chat, underscore the significant challenges posed by MMT-Bench. We anticipate that MMT-Bench will inspire the community to develop next-generation multimodal foundation models aimed at achieving general-purpose multimodal intelligence.

Fanqing Meng, Lingxiao Du, Zongkai Liu et al.

DeepSeek R1, and o1 have demonstrated powerful reasoning capabilities in the text domain through stable large-scale reinforcement learning. To enable broader applications, some works have attempted to transfer these capabilities to multimodal reasoning. However, these efforts have been limited by the limited difficulty of selected tasks and relatively small training scales, making it challenging to demonstrate strong multimodal reasoning abilities. To address this gap, we introduce the MMK12 dataset and MM-EUREKA with 7B and 32B parameters. The former is a high-quality multimodal mathematics reasoning dataset featuring diverse knowledge domains with human-verified answers and solution processes. The latter is a multimodal model employing rule-based reinforcement learning on MMK12, utilizing online filtering and two-stage training strategy to enhance training stability. MM-EUREKA demonstrates remarkable performance gains in multimodal mathematical reasoning, outperforming previous powerful models like InternVL2.5-78B or InternVL2.5-38B-MPO. In particular, MM-EUREKA achieves competitive or superior performance compared to both open-source and closed-source models, and trails slightly behind o1 in multidisciplinary reasoning tasks. We open-source our complete pipeline to foster further research in this area. We release all our codes, models, data, etc. at https://github.com/ModalMinds/MM-EUREKA

Fanqing Meng, Wenqi Shao, Quanfeng Lu et al.

Charts play a vital role in data visualization, understanding data patterns, and informed decision-making. However, their unique combination of graphical elements (e.g., bars, lines) and textual components (e.g., labels, legends) poses challenges for general-purpose multimodal models. While vision-language models trained on chart data excel in comprehension, they struggle with generalization. To address these challenges, we propose ChartAssistant, a chart-based vision-language model for universal chart comprehension and reasoning. ChartAssistant leverages ChartSFT, a comprehensive dataset covering diverse chart-related tasks with basic (e.g. bars and pies) and specialized (e.g. radars, and bubbles) chart types. It undergoes a two-stage training process, starting with pre-training on chart-to-table parsing to align chart and text, followed by multitask instruction-following fine-tuning. This approach enables ChartAssistant to achieve competitive performance across various chart tasks. Experimental results demonstrate significant performance gains over the state-of-the-art UniChart and Chartllama method, especially outperforming them on real-world chart data with zero-shot setting. The code and data are available at https://github.com/OpenGVLab/ChartAst.

Lingxiao Du, Fanqing Meng, Zongkai Liu et al.

While Multimodal Large Language Models (MLLMs) have achieved impressive progress in vision-language understanding, they still struggle with complex multi-step reasoning, often producing logically inconsistent or partially correct solutions. A key limitation lies in the lack of fine-grained supervision over intermediate reasoning steps. To address this, we propose MM-PRM, a process reward model trained within a fully automated, scalable framework. We first build MM-Policy, a strong multimodal model trained on diverse mathematical reasoning data. Then, we construct MM-K12, a curated dataset of 10,000 multimodal math problems with verifiable answers, which serves as seed data. Leveraging a Monte Carlo Tree Search (MCTS)-based pipeline, we generate over 700k step-level annotations without human labeling. The resulting PRM is used to score candidate reasoning paths in the Best-of-N inference setup and achieves significant improvements across both in-domain (MM-K12 test set) and out-of-domain (OlympiadBench, MathVista, etc.) benchmarks. Further analysis confirms the effectiveness of soft labels, smaller learning rates, and path diversity in optimizing PRM performance. MM-PRM demonstrates that process supervision is a powerful tool for enhancing the logical robustness of multimodal reasoning systems. We release all our codes and data at https://github.com/ModalMinds/MM-PRM.

Zongkai Liu, Fanqing Meng, Lingxiao Du et al.

Recent advances in rule-based reinforcement learning (RL) have significantly improved the reasoning capability of language models (LMs) with rule-based rewards. However, existing RL methods -- such as GRPO, REINFORCE++, and RLOO -- often suffer from training instability, where large policy updates and improper clipping can lead to training collapse. To address this issue, we propose Clipped Policy Gradient Optimization with Policy Drift (CPGD), a novel algorithm designed to stabilize policy learning in LMs. CPGD introduces a policy drift constraint based on KL divergence to dynamically regularize policy updates, and leverages a clip mechanism on the logarithm of the ratio to prevent excessive policy updates. We provide theoretical justification for CPGD and demonstrate through empirical analysis that it mitigates the instability observed in prior approaches. Furthermore, we show that CPGD significantly improves performance while maintaining training stability. Our implementation balances theoretical rigor with practical usability, offering a robust alternative for RL in the post-training of LMs. We release our code at https://github.com/ModalMinds/MM-EUREKA.

Charles Chen, Qiming Yu, Yuhang Gu et al.

As agent systems scale, skills accumulate into large reusable libraries, yet their scaling laws remain poorly understood. Across 15 frontier LLMs, 1,141 real-world skills, and over 3M routing or execution decisions, we identify two coupled laws. Routing law: single-step routing accuracy decays logarithmically with library size ($R^2{>}0.97$ for all models), with errors progressing from local skill competition to cross-family drift and capture by overly general "black-hole skills". Execution law: before state realization, joint routing is approximately multiplicative, whereas correct execution can improve difficult downstream decisions by about $4{\times}$. A single parameter, the routing logarithmic decay slope $b$, couples the two laws: routing-side fits predict execution-side rescue across models, showing that the same library property controls both pre-execution collapse and downstream recoverability. The laws are actionable: law-guided optimization raises held-out routing accuracy from 71.3% to 91.7%, reduces hijack from 22.4% to 4.1%, and transfers directionally to downstream ClawBench and ClawMark execution settings, improving mean pass rate from 49.3% to 61.6% on ClawBench and from 28.4% to 34.5% on ClawMark. These results show that agent performance depends not only on model capability, but also on the structure, granularity, and exposure policy of the skill library.

Zheng Yan, Jingxiang Weng, Charles Chen et al.

As coding agents gain access to shells, repositories, and user files, least-privilege authorization becomes a prerequisite for safe deployment: an agent should receive enough authority to complete the task, without unnecessary authority that exposes sensitive surfaces.To study whether current models can infer this boundary themselves, we first introduce permission-boundary inference, where a model maps a task instruction and terminal environment to a file-level read/write/execute policy, and AuthBench, a benchmark of 120 realistic terminal tasks with human-reviewed permission labels and executable validators for utility and attack outcomes.AuthBench shows that authorization is not a simple conservative-versus-permissive calibration problem: frontier models often omit permissions required by the execution chain while also granting unused or sensitive accesses.Increasing inference-time reasoning does not resolve this mismatch. Instead, each model moves toward a model-specific authorization attractor: more reasoning makes it more consistent in its own failure mode, whether broad-but-exposed or tight-but-brittle.This suggests that direct policy generation is the bottleneck, because a single generation must both discover all necessary accesses and reject all unnecessary ones.We therefore propose Sufficiency-Tightness Decomposition, which first generates a coverage-oriented policy by forward-simulating the task and then audits each granted entry for grounding and sensitivity.Across tested models, this decomposition improves sensitive-task success by up to 15.8% on tightness-biased models while reducing attack success across all evaluated models.

Fanqing Meng, Lingxiao Du, Zijian Wu et al.

Language-model agents are increasingly used as persistent coworkers that assist users across multiple working days. During such workflows, the surrounding environment may change independently of the agent: new emails arrive, calendar entries shift, knowledge-base records are updated, and evidence appears across images, scanned PDFs, audio, video, and spreadsheets. Existing benchmarks do not adequately evaluate this setting because they typically run within a single static episode and remain largely text-centric. We introduce \bench{}, a benchmark for coworker agents built around multi-turn multi-day tasks, a stateful sandboxed service environment whose state evolves between turns, and rule-based verification. The current release contains 100 tasks across 13 professional scenarios, executed against five stateful sandboxed services (filesystem, email, calendar, knowledge base, spreadsheet) and scored by 1537 deterministic Python checkers over post-execution service state; no LLM-as-judge is invoked during scoring. We benchmark seven frontier agent systems. The strongest model reaches 75.8 weighted score, but the best strict Task Success is only 20.0\%, indicating that partial progress is common while complete end-to-end workflow completion remains rare. Turn-level analysis shows that performance drops after the first exogenous environment update, highlighting adaptation to changing state as a key open challenge. We release the benchmark, evaluation harness, and construction pipeline to support reproducible coworker-agent evaluation.

Xiangdong Zhang, Jiaqi Liao, Shaofeng Zhang et al.

Recent advancements in text-to-video (T2V) diffusion models have enabled high-fidelity and realistic video synthesis. However, current T2V models often struggle to generate physically plausible content due to their limited inherent ability to accurately understand physics. We found that while the representations within T2V models possess some capacity for physics understanding, they lag significantly behind those from recent video self-supervised learning methods. To this end, we propose a novel framework called VideoREPA, which distills physics understanding capability from video understanding foundation models into T2V models by aligning token-level relations. This closes the physics understanding gap and enable more physics-plausible generation. Specifically, we introduce the Token Relation Distillation (TRD) loss, leveraging spatio-temporal alignment to provide soft guidance suitable for finetuning powerful pre-trained T2V models, a critical departure from prior representation alignment (REPA) methods. To our knowledge, VideoREPA is the first REPA method designed for finetuning T2V models and specifically for injecting physical knowledge. Empirical evaluations show that VideoREPA substantially enhances the physics commonsense of baseline method, CogVideoX, achieving significant improvement on relevant benchmarks and demonstrating a strong capacity for generating videos consistent with intuitive physics. More video results are available at https://videorepa.github.io/.

Jiaqi Liao, Yuwei Niu, Fanqing Meng et al.

Recent years have witnessed remarkable advances in Large Vision-Language Models (LVLMs), which have achieved human-level performance across various complex vision-language tasks. Following LLaVA's paradigm, mainstream LVLMs typically employ a shallow MLP for visual-language alignment through a two-stage training process: pretraining for cross-modal alignment followed by instruction tuning. While this approach has proven effective, the underlying mechanisms of how MLPs bridge the modality gap remain poorly understood. Although some research has explored how LLMs process transformed visual tokens, few studies have investigated the fundamental alignment mechanism. Furthermore, the MLP adapter requires retraining whenever switching LLM backbones. To address these limitations, we first investigate the working principles of MLP adapters and discover that they learn to project visual embeddings into subspaces spanned by corresponding text embeddings progressively. Based on this insight, we propose LangBridge, a novel adapter that explicitly maps visual tokens to linear combinations of LLM vocabulary embeddings. This innovative design enables pretraining-free adapter transfer across different LLMs while maintaining performance. Our experimental results demonstrate that a LangBridge adapter pre-trained on Qwen2-0.5B can be directly applied to larger models such as LLaMA3-8B or Qwen2.5-14B while maintaining competitive performance. Overall, LangBridge enables interpretable vision-language alignment by grounding visual representations in LLM vocab embedding, while its plug-and-play design ensures efficient reuse across multiple LLMs with nearly no performance degradation. See our project page at https://curryx-001.github.io/LangBridge.github.io/

Zijian Wu, Xiangyan Liu, Xinyuan Zhang et al.

MCP standardizes how LLMs interact with external systems, forming the foundation for general agents. However, existing MCP benchmarks remain narrow in scope: they focus on read-heavy tasks or tasks with limited interaction depth, and fail to capture the complexity and realism of real-world workflows. To address this gap, we propose MCPMark, a benchmark designed to evaluate MCP use in a more realistic and comprehensive manner. It consists of $127$ high-quality tasks collaboratively created by domain experts and AI agents. Each task begins with a curated initial state and includes a programmatic script for automatic verification. These tasks demand richer and more diverse interactions with the environment, involving a broad range of create, read, update, and delete (CRUD) operations. We conduct a comprehensive evaluation of cutting-edge LLMs using a minimal agent framework that operates in a tool-calling loop. Empirical results show that the best-performing model, gpt-5-medium, reaches only $52.56$\% pass@1 and $33.86$\% pass^4, while other widely regarded strong models, including claude-sonnet-4 and o3, fall below $30$\% pass@1 and $15$\% pass^4. On average, LLMs require $16.2$ execution turns and $17.4$ tool calls per task, significantly surpassing those in previous MCP benchmarks and highlighting the stress-testing nature of MCPMark.

Lingnan Xu, Chong Feng, Kaiyuan Zhang et al.

While large language models (LLMs) demonstrate impressive capabilities, their reliance on parametric knowledge often leads to factual inaccuracies. Retrieval-Augmented Generation (RAG) mitigates this by leveraging external documents, yet existing approaches treat retrieved passages as isolated chunks, ignoring valuable structure that is crucial for document organization. Motivated by this gap, we propose Retrieve-DocumentRoute-Read (RDR2), a novel framework that explicitly incorporates structural information throughout the RAG process. RDR2 employs an LLM-based router to dynamically navigate document structure trees, jointly evaluating content relevance and hierarchical relationships to assemble optimal evidence. Our key innovation lies in formulating document routing as a trainable task, with automatic action curation and structure-aware passage selection inspired by human reading strategies. Through comprehensive evaluation on five challenging datasets, RDR2 achieves state-of-the-art performance, demonstrating that explicit structural awareness significantly enhances RAG systems' ability to acquire and utilize knowledge, particularly in complex scenarios requiring multi-document synthesis.

Fanqing Meng, Wenqi Shao, Lixin Luo et al.

Text-to-image (T2I) models have made substantial progress in generating images from textual prompts. However, they frequently fail to produce images consistent with physical commonsense, a vital capability for applications in world simulation and everyday tasks. Current T2I evaluation benchmarks focus on metrics such as accuracy, bias, and safety, neglecting the evaluation of models' internal knowledge, particularly physical commonsense. To address this issue, we introduce PhyBench, a comprehensive T2I evaluation dataset comprising 700 prompts across 4 primary categories: mechanics, optics, thermodynamics, and material properties, encompassing 31 distinct physical scenarios. We assess 6 prominent T2I models, including proprietary models DALLE3 and Gemini, and demonstrate that incorporating physical principles into prompts enhances the models' ability to generate physically accurate images. Our findings reveal that: (1) even advanced models frequently err in various physical scenarios, except for optics; (2) GPT-4o, with item-specific scoring instructions, effectively evaluates the models' understanding of physical commonsense, closely aligning with human assessments; and (3) current T2I models are primarily focused on text-to-image translation, lacking profound reasoning regarding physical commonsense. We advocate for increased attention to the inherent knowledge within T2I models, beyond their utility as mere image generation tools. The data will be available soon.

Quanfeng Lu, Wenqi Shao, Zitao Liu et al.

Autonomous Graphical User Interface (GUI) navigation agents can enhance user experience in communication, entertainment, and productivity by streamlining workflows and reducing manual intervention. However, prior GUI agents often trained with datasets comprising tasks that can be completed within a single app, leading to poor performance in cross-app navigation. To address this problem, we present GUIOdyssey, a comprehensive dataset for cross-app mobile GUI navigation. GUIOdyssey comprises 8,334 episodes with an average of 15.3 steps per episode, covering 6 mobile devices, 212 distinct apps, and 1,357 app combinations. Each step is enriched with detailed semantic reasoning annotations, which aid the model in building cognitive processes and enhancing its reasoning abilities for complex cross-app tasks. Building on GUIOdyssey, we develop OdysseyAgent, an exploratory multimodal agent for long-step cross-app navigation equipped with a history resampler module that efficiently attends to historical screenshot tokens, balancing performance and inference speed. Extensive experiments conducted in both in-domain and out-of-domain scenarios validate the effectiveness of our approach. Moreover, we demonstrate that historial information involving actions, screenshots and context in our dataset can significantly enhances OdysseyAgent's performance on complex cross-app tasks.