Shiyin Lu

Haodong Duan, Xinyu Fang, Junming Yang et al. · pku

We present VLMEvalKit: an open-source toolkit for evaluating large multi-modality models based on PyTorch. The toolkit aims to provide a user-friendly and comprehensive framework for researchers and developers to evaluate existing multi-modality models and publish reproducible evaluation results. In VLMEvalKit, we implement over 200+ different large multi-modality models, including both proprietary APIs and open-source models, as well as more than 80 different multi-modal benchmarks. By implementing a single interface, new models can be easily added to the toolkit, while the toolkit automatically handles the remaining workloads, including data preparation, distributed inference, prediction post-processing, and metric calculation. Although the toolkit is currently mainly used for evaluating large vision-language models, its design is compatible with future updates that incorporate additional modalities, such as audio and video. Based on the evaluation results obtained with the toolkit, we host OpenVLM Leaderboard, a comprehensive leaderboard to track the progress of multi-modality learning research. The toolkit is released on https://github.com/open-compass/VLMEvalKit and is actively maintained.

Wenhao Yang, Yu Xia, Jinlong Huang et al.

Recent advances in large Vision-Language Models (VLMs) have exhibited strong reasoning capabilities on complex visual tasks by thinking with images in their Chain-of-Thought (CoT), which is achieved by actively invoking tools to analyze visual inputs rather than merely perceiving them. However, existing models often struggle to reflect on and correct themselves when attempting incorrect reasoning trajectories. To address this limitation, we propose DRIM, a model that enables deep but reliable multi-turn reasoning when thinking with images in its multimodal CoT. Our pipeline comprises three stages: data construction, cold-start SFT and RL. Based on a high-resolution image dataset, we construct high-difficulty and verifiable visual question-answer pairs, where solving each task requires multi-turn tool calls to reach the correct answer. In the SFT stage, we collect tool trajectories as cold-start data, guiding a multi-turn reasoning pattern. In the RL stage, we introduce redundancy-penalized policy optimization, which incentivizes the model to develop a self-reflective reasoning pattern. The basic idea is to impose judgment on reasoning trajectories and penalize those that produce incorrect answers without sufficient multi-scale exploration. Extensive experiments demonstrate that DRIM achieves superior performance on visual understanding benchmarks.

Shiyin Lu, Yang Li, Yu Xia et al.

We present Ovis2.5, a successor to Ovis2 designed for native-resolution visual perception and strong multimodal reasoning. Ovis2.5 integrates a native-resolution vision transformer that processes images at their native, variable resolutions, avoiding the degradation from fixed-resolution tiling and preserving both fine detail and global layout -- crucial for visually dense content like complex charts. To strengthen reasoning, we train the model to move beyond linear chain-of-thought and perform reflection -- including self-checking and revision. This advanced capability is exposed as an optional "thinking mode" at inference time, allowing users to trade latency for enhanced accuracy on difficult inputs. The model is trained via a comprehensive five-phase curriculum that progressively builds its skills. The process begins with foundational visual and multimodal pretraining, advances through large-scale instruction tuning, and culminates in alignment and reasoning enhancement using DPO and GRPO. To scale these upgrades efficiently, we employ multimodal data packing and hybrid parallelism, yielding a significant end-to-end speedup. We release two open-source models: Ovis2.5-9B and Ovis2.5-2B. The latter continues the "small model, big performance" philosophy of Ovis2, making it ideal for resource-constrained, on-device scenarios. On the OpenCompass multimodal leaderboard, Ovis2.5-9B averages 78.3, marking a substantial improvement over its predecessor, Ovis2-8B, and achieving state-of-the-art results among open-source MLLMs in the sub-40B parameter range; Ovis2.5-2B scores 73.9, establishing SOTA for its size. Beyond aggregate scores, Ovis2.5 achieves leading results on STEM benchmarks, exhibits strong capabilities on grounding and video tasks, and achieves open-source SOTA at its scale for complex chart analysis.

Wenhao Yang, Yu Xia, Jinlong Huang et al.

Recent advancements in Multimodal Large Language Models (MLLMs) have incentivized models to ``think with images'' by actively invoking visual tools during multi-turn reasoning. The common Reinforcement Learning (RL) practice of relying on outcome-based rewards ignores the fact that textual plausibility often masks executive failure, meaning that models may exhibit intuitive textual reasoning while executing imprecise or irrelevant visual actions within their agentic reasoning trajectories. This reasoning-action discrepancy introduces noise that accumulates throughout the multi-turn reasoning process, severely degrading the model's multimodal reasoning capabilities and potentially leading to training collapse. In this paper, we introduce Multimodal Agentic Policy Optimization (MAPO), bridging the gap between textual reasoning and visual actions generated by models within their Multimodal Chain-of-Thought (MCoT). Specifically, MAPO mandates the model to generate explicit textual descriptions for the visual content obtained via tool usage. We then employ a novel advantage estimation that couples the semantic alignment between these descriptions and the actual observations with the task reward. Theoretical findings are provided to justify the rationale behind MAPO, which inherently reduces the variance of gradients, and extensive experiments demonstrate that our method achieves superior performance across multiple visual reasoning benchmarks.

Jiaqi Tang, Yu Xia, Yi-Feng Wu et al.

The advent of autonomous agents is transforming interactions with Graphical User Interfaces (GUIs) by employing natural language as a powerful intermediary. Despite the predominance of Supervised Fine-Tuning (SFT) methods in current GUI agents for achieving spatial localization, these methods face substantial challenges due to their limited capacity to accurately perceive positional data. Existing strategies, such as reinforcement learning, often fail to assess positional accuracy effectively, thereby restricting their utility. In response, we introduce Location Preference Optimization (LPO), a novel approach that leverages locational data to optimize interaction preferences. LPO uses information entropy to predict interaction positions by focusing on zones rich in information. Besides, it further introduces a dynamic location reward function based on physical distance, reflecting the varying importance of interaction positions. Supported by Group Relative Preference Optimization (GRPO), LPO facilitates an extensive exploration of GUI environments and significantly enhances interaction precision. Comprehensive experiments demonstrate LPO's superior performance, achieving SOTA results across both offline benchmarks and real-world online evaluations. Our code will be made publicly available soon, at https://github.com/AIDC-AI/LPO.

Hai-Long Sun, Da-Wei Zhou, Yang Li et al.

The rapid development of Multimodal Large Language Models (MLLMs), such as GPT-4o, marks a significant step toward artificial general intelligence. Existing methods typically align vision encoders with LLMs via supervised fine-tuning (SFT), but this often deteriorates their ability to handle multiple languages as training progresses. We empirically observe that imbalanced SFT datasets, largely English-centric, degrade performance on non-English languages due to the failure in multilingual token alignment. To address this, we propose PARROT, a novel approach that leverages textual guidance for visual token alignment at the language level. PARROT conditions visual tokens on diverse language inputs and uses Mixture-of-Experts (MoE) to align multilingual tokens. By computing cross-attention between initial visual features and textual embeddings, we select the most relevant experts, converting visual tokens into language-specific representations. Additionally, we introduce the Massive Multilingual Multimodal Benchmark (MMMB), a new benchmark comprising 6 languages, 15 categories, and 12,000 questions, to assess multilingual capabilities. PARROT achieves state-of-the-art performance on both the multilingual benchmarks and a wide range of multimodal tasks. Code and dataset are available at: https://github.com/AIDC-AI/Parrot

Guo-Hua Wang, Shanshan Zhao, Xinjie Zhang et al.

In this report, we introduce Ovis-U1, a 3-billion-parameter unified model that integrates multimodal understanding, text-to-image generation, and image editing capabilities. Building on the foundation of the Ovis series, Ovis-U1 incorporates a diffusion-based visual decoder paired with a bidirectional token refiner, enabling image generation tasks comparable to leading models like GPT-4o. Unlike some previous models that use a frozen MLLM for generation tasks, Ovis-U1 utilizes a new unified training approach starting from a language model. Compared to training solely on understanding or generation tasks, unified training yields better performance, demonstrating the enhancement achieved by integrating these two tasks. Ovis-U1 achieves a score of 69.6 on the OpenCompass Multi-modal Academic Benchmark, surpassing recent state-of-the-art models such as Ristretto-3B and SAIL-VL-1.5-2B. In text-to-image generation, it excels with scores of 83.72 and 0.89 on the DPG-Bench and GenEval benchmarks, respectively. For image editing, it achieves 4.00 and 6.42 on the ImgEdit-Bench and GEdit-Bench-EN, respectively. As the initial version of the Ovis unified model series, Ovis-U1 pushes the boundaries of multimodal understanding, generation, and editing.

Hui Sun, Shiyin Lu, Huanyu Wang et al.

Video large language models (Video-LLMs) have made significant progress in understanding videos. However, processing multiple frames leads to lengthy visual token sequences, presenting challenges such as the limited context length cannot accommodate the entire video, and the inclusion of irrelevant frames hinders visual perception. Hence, effective frame selection is crucial. This paper emphasizes that frame selection should follow three key principles: query relevance, list-wise diversity, and sequentiality. Existing methods, such as uniform frame sampling and query-frame matching, do not capture all of these principles. Thus, we propose Markov decision determinantal point process with dynamic programming (MDP3) for frame selection, a training-free and model-agnostic method that can be seamlessly integrated into existing Video-LLMs. Our method first estimates frame similarities conditioned on the query using a conditional Gaussian kernel within the reproducing kernel Hilbert space~(RKHS). We then apply the determinantal point process~(DPP) to the similarity matrix to capture both query relevance and list-wise diversity. To incorporate sequentiality, we segment the video and apply DPP within each segment, conditioned on the preceding segment selection, modeled as a Markov decision process~(MDP) for allocating selection sizes across segments. Theoretically, MDP3 provides a \((1 - 1/e)\)-approximate solution to the NP-hard list-wise frame selection problem with pseudo-polynomial time complexity, demonstrating its efficiency. Empirically, MDP3 significantly outperforms existing methods, verifying its effectiveness and robustness.

Jun-Peng Jiang, Yu Xia, Hai-Long Sun et al.

Tabular reasoning involves multi-step information extraction and logical inference over tabular data. While recent advances have leveraged large language models (LLMs) for reasoning over structured tables, such high-quality textual representations are often unavailable in real-world settings, where tables typically appear as images. In this paper, we tackle the task of tabular reasoning from table images, leveraging privileged structured information available during training to enhance multimodal large language models (MLLMs). The key challenges lie in the complexity of accurately aligning structured information with visual representations, and in effectively transferring structured reasoning skills to MLLMs despite the input modality gap. To address these, we introduce TabUlar Reasoning with Bridged infOrmation ({\sc Turbo}), a new framework for multimodal tabular reasoning with privileged structured tables. {\sc Turbo} benefits from a structure-aware reasoning trace generator based on DeepSeek-R1, contributing to high-quality modality-bridged data. On this basis, {\sc Turbo} repeatedly generates and selects the advantageous reasoning paths, further enhancing the model's tabular reasoning ability. Experimental results demonstrate that, with limited ($9$k) data, {\sc Turbo} achieves state-of-the-art performance ($+7.2\%$ vs. previous SOTA) across multiple datasets.

Yi-Kai Zhang, Xu-Xiang Zhong, Shiyin Lu et al.

The rapid advancements in Large Language Models (LLMs) have significantly expanded their applications, ranging from multilingual support to domain-specific tasks and multimodal integration. In this paper, we present OmniEvalKit, a novel benchmarking toolbox designed to evaluate LLMs and their omni-extensions across multilingual, multidomain, and multimodal capabilities. Unlike existing benchmarks that often focus on a single aspect, OmniEvalKit provides a modular, lightweight, and automated evaluation system. It is structured with a modular architecture comprising a Static Builder and Dynamic Data Flow, promoting the seamless integration of new models and datasets. OmniEvalKit supports over 100 LLMs and 50 evaluation datasets, covering comprehensive evaluations across thousands of model-dataset combinations. OmniEvalKit is dedicated to creating an ultra-lightweight and fast-deployable evaluation framework, making downstream applications more convenient and versatile for the AI community.

Yi-Kai Zhang, Shiyin Lu, Yang Li et al.

Multimodal large language models (MLLMs), initiated with a trained LLM, first align images with text and then fine-tune on multimodal mixed inputs. However, the MLLM catastrophically forgets the text-only instructions, which do not include images and can be addressed within the initial LLM. In this paper, we present Wings, a novel MLLM that excels in both text-only dialogues and multimodal comprehension. Analyzing MLLM attention in multimodal instructions reveals that text-only forgetting is related to the attention shifts from pre-image to post-image text. From that, we construct extra modules that act as the boosted learner to compensate for the attention shift. The complementary visual and textual learners, like "wings" on either side, are connected in parallel within each layer's attention block. Initially, image and text inputs are aligned with visual learners operating alongside the main attention, balancing focus on visual elements. Textual learners are later collaboratively integrated with attention-based routing to blend the outputs of the visual and textual learners. We design the Low-Rank Residual Attention (LoRRA) to guarantee high efficiency for learners. Our experimental results demonstrate that Wings outperforms equally-scaled MLLMs in both text-only and visual question-answering tasks. On a newly constructed Interleaved Image-Text (IIT) benchmark, Wings exhibits superior performance from text-only-rich to multimodal-rich question-answering tasks.

Yibo Wang, Wenhao Yang, Wei Jiang et al.

Projection-free online learning has drawn increasing interest due to its efficiency in solving high-dimensional problems with complicated constraints. However, most existing projection-free online methods focus on minimizing the static regret, which unfortunately fails to capture the challenge of changing environments. In this paper, we investigate non-stationary projection-free online learning, and choose dynamic regret and adaptive regret to measure the performance. Specifically, we first provide a novel dynamic regret analysis for an existing projection-free method named $\text{BOGD}_\text{IP}$, and establish an $\mathcal{O}(T^{3/4}(1+P_T))$ dynamic regret bound, where $P_T$ denotes the path-length of the comparator sequence. Then, we improve the upper bound to $\mathcal{O}(T^{3/4}(1+P_T)^{1/4})$ by running multiple $\text{BOGD}_\text{IP}$ algorithms with different step sizes in parallel, and tracking the best one on the fly. Our results are the first general-case dynamic regret bounds for projection-free online learning, and can recover the existing $\mathcal{O}(T^{3/4})$ static regret by setting $P_T = 0$. Furthermore, we propose a projection-free method to attain an $\tilde{\mathcal{O}}(τ^{3/4})$ adaptive regret bound for any interval with length $τ$, which nearly matches the static regret over that interval. The essential idea is to maintain a set of $\text{BOGD}_\text{IP}$ algorithms dynamically, and combine them by a meta algorithm. Moreover, we demonstrate that it is also equipped with an $\mathcal{O}(T^{3/4}(1+P_T)^{1/4})$ dynamic regret bound. Finally, empirical studies verify our theoretical findings.

Lijun Zhang, Wei Jiang, Shiyin Lu et al.

In this paper, we revisit the problem of smoothed online learning, in which the online learner suffers both a hitting cost and a switching cost, and target two performance metrics: competitive ratio and dynamic regret with switching cost. To bound the competitive ratio, we assume the hitting cost is known to the learner in each round, and investigate the simple idea of balancing the two costs by an optimization problem. Surprisingly, we find that minimizing the hitting cost alone is $\max(1, \frac{2}α)$-competitive for $α$-polyhedral functions and $1 + \frac{4}λ$-competitive for $λ$-quadratic growth functions, both of which improve state-of-the-art results significantly. Moreover, when the hitting cost is both convex and $λ$-quadratic growth, we reduce the competitive ratio to $1 + \frac{2}{\sqrtλ}$ by minimizing the weighted sum of the hitting cost and the switching cost. To bound the dynamic regret with switching cost, we follow the standard setting of online convex optimization, in which the hitting cost is convex but hidden from the learner before making predictions. We modify Ader, an existing algorithm designed for dynamic regret, slightly to take into account the switching cost when measuring the performance. The proposed algorithm, named as Smoothed Ader, attains an optimal $O(\sqrt{T(1+P_T)})$ bound for dynamic regret with switching cost, where $P_T$ is the path-length of the comparator sequence. Furthermore, if the hitting cost is accessible in the beginning of each round, we obtain a similar guarantee without the bounded gradient condition, and establish an $Ω(\sqrt{T(1+P_T)})$ lower bound to confirm the optimality.

Lijun Zhang, Shiyin Lu, Tianbao Yang

Regret minimization is treated as the golden rule in the traditional study of online learning. However, regret minimization algorithms tend to converge to the static optimum, thus being suboptimal for changing environments. To address this limitation, new performance measures, including dynamic regret and adaptive regret have been proposed to guide the design of online algorithms. The former one aims to minimize the global regret with respect to a sequence of changing comparators, and the latter one attempts to minimize every local regret with respect to a fixed comparator. Existing algorithms for dynamic regret and adaptive regret are developed independently, and only target one performance measure. In this paper, we bridge this gap by proposing novel online algorithms that are able to minimize the dynamic regret and adaptive regret simultaneously. In fact, our theoretical guarantee is even stronger in the sense that one algorithm is able to minimize the dynamic regret over any interval.

Shiyin Lu, Lijun Zhang

In this paper, we consider the problem of prediction with expert advice in dynamic environments. We choose tracking regret as the performance metric and develop two adaptive and efficient algorithms with data-dependent tracking regret bounds. The first algorithm achieves a second-order tracking regret bound, which improves existing first-order bounds. The second algorithm enjoys a path-length bound, which is generally not comparable to the second-order bound but offers advantages in slowly moving environments. Both algorithms are developed under the online mirror descent framework and draw inspiration from existing algorithms that attain data-dependent bounds of static regret. The key idea is to use a clipped simplex in the updating step of online mirror descent. Finally, we extend our algorithms and analysis to online matrix prediction and provide the first data-dependent tracking regret bound for this problem.

Shiyin Lu, Guanghui Wang, Yao Hu et al.

In this paper, we study the multi-objective bandits (MOB) problem, where a learner repeatedly selects one arm to play and then receives a reward vector consisting of multiple objectives. MOB has found many real-world applications as varied as online recommendation and network routing. On the other hand, these applications typically contain contextual information that can guide the learning process which, however, is ignored by most of existing work. To utilize this information, we associate each arm with a context vector and assume the reward follows the generalized linear model (GLM). We adopt the notion of Pareto regret to evaluate the learner's performance and develop a novel algorithm for minimizing it. The essential idea is to apply a variant of the online Newton step to estimate model parameters, based on which we utilize the upper confidence bound (UCB) policy to construct an approximation of the Pareto front, and then uniformly at random choose one arm from the approximate Pareto front. Theoretical analysis shows that the proposed algorithm achieves an $\tilde O(d\sqrt{T})$ Pareto regret, where $T$ is the time horizon and $d$ is the dimension of contexts, which matches the optimal result for single objective contextual bandits problem. Numerical experiments demonstrate the effectiveness of our method.

Guanghui Wang, Shiyin Lu, Lijun Zhang

In this paper, we study adaptive online convex optimization, and aim to design a universal algorithm that achieves optimal regret bounds for multiple common types of loss functions. Existing universal methods are limited in the sense that they are optimal for only a subclass of loss functions. To address this limitation, we propose a novel online method, namely Maler, which enjoys the optimal $O(\sqrt{T})$, $O(d\log T)$ and $O(\log T)$ regret bounds for general convex, exponentially concave, and strongly convex functions respectively. The essential idea is to run multiple types of learning algorithms with different learning rates in parallel, and utilize a meta algorithm to track the best one on the fly. Empirical results demonstrate the effectiveness of our method.

Guanghui Wang, Shiyin Lu, Weiwei Tu et al.

The Adam algorithm has become extremely popular for large-scale machine learning. Under convexity condition, it has been proved to enjoy a data-dependant $O(\sqrt{T})$ regret bound where $T$ is the time horizon. However, whether strong convexity can be utilized to further improve the performance remains an open problem. In this paper, we give an affirmative answer by developing a variant of Adam (referred to as SAdam) which achieves a data-dependant $O(\log T)$ regret bound for strongly convex functions. The essential idea is to maintain a faster decaying yet under controlled step size for exploiting strong convexity. In addition, under a special configuration of hyperparameters, our SAdam reduces to SC-RMSprop, a recently proposed variant of RMSprop for strongly convex functions, for which we provide the first data-dependent logarithmic regret bound. Empirical results on optimizing strongly convex functions and training deep networks demonstrate the effectiveness of our method.

Lijun Zhang, Shiyin Lu, Zhi-Hua Zhou

In this paper, we study online convex optimization in dynamic environments, and aim to bound the dynamic regret with respect to any sequence of comparators. Existing work have shown that online gradient descent enjoys an $O(\sqrt{T}(1+P_T))$ dynamic regret, where $T$ is the number of iterations and $P_T$ is the path-length of the comparator sequence. However, this result is unsatisfactory, as there exists a large gap from the $Ω(\sqrt{T(1+P_T)})$ lower bound established in our paper. To address this limitation, we develop a novel online method, namely adaptive learning for dynamic environment (Ader), which achieves an optimal $O(\sqrt{T(1+P_T)})$ dynamic regret. The basic idea is to maintain a set of experts, each attaining an optimal dynamic regret for a specific path-length, and combines them with an expert-tracking algorithm. Furthermore, we propose an improved Ader based on the surrogate loss, and in this way the number of gradient evaluations per round is reduced from $O(\log T)$ to $1$. Finally, we extend Ader to the setting that a sequence of dynamical models is available to characterize the comparators.