Bowen Qu

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.

Weixin Liu, Bowen Qu, Juming Xiong et al.

Large language model (LLM) summarization systems may pass compact vector representations of private inputs to downstream retrieval, monitoring, audit, or analytic workflows. Even when source documents remain access-restricted, derived vectors may be handled under different access controls and still support sensitive-information inference, creating a residual information-disclosure risk. We study this issue in clinical discharge-summary generation as a high-stakes case study, using electronic health record (EHR)-recorded race as a controlled sensitive-label audit. We audit two artifacts that a system might retain or expose to downstream components: the final prompt-token hidden state and the mean-pooled prompt representation. Our results show that reducing recoverability of the case-study sensitive label from one exported artifact does not necessarily reduce recoverability from another. As a mitigation case study, we introduce SurfaceLoRA, an exported-vector-targeted parameter-efficient fine-tuning method that uses a gradient-reversal discriminator attached to a designated exported vector. Under a balanced five-way probing protocol, SurfaceLoRA reduces EHR-recorded race recoverability from the targeted final-token artifact toward chance while preserving summarization utility, yet recoverability remains substantially higher from untargeted pooled artifacts. These findings show that privacy auditing and mitigation should be performed on the exact vector artifact retained or exposed to downstream components.

Zhengzhuo Xu, Bowen Qu, Yiyan Qi et al.

Automatic chart understanding is crucial for content comprehension and document parsing. Multimodal Large Language Models (MLLMs) have demonstrated remarkable capabilities in chart understanding through domain-specific alignment and fine-tuning. However, current MLLMs still struggle to provide faithful data and reliable analysis only based on charts. To address it, we propose ChartMoE, which employs the Mixture of Expert (MoE) architecture to replace the traditional linear projector to bridge the modality gap. Specifically, we train several linear connectors through distinct alignment tasks, which are utilized as the foundational initialization parameters for different experts. Additionally, we introduce ChartMoE-Align, a dataset with nearly 1 million chart-table-JSON-code quadruples to conduct three alignment tasks (chart-table/JSON/code). Combined with the vanilla connector, we initialize different experts diversely and adopt high-quality knowledge learning to further refine the MoE connector and LLM parameters. Extensive experiments demonstrate the effectiveness of the MoE connector and our initialization strategy, e.g., ChartMoE improves the accuracy of the previous state-of-the-art from 80.48\% to 84.64\% on the ChartQA benchmark.

Shuqing Luo, Bowen Qu, Wei Gao

In this paper, we explore a critical yet under-investigated issue: how to learn robust and well-generalized 3D representation from pre-trained vision language models such as CLIP. Previous works have demonstrated that cross-modal distillation can provide rich and useful knowledge for 3D data. However, like most deep learning models, the resultant 3D learning network is still vulnerable to adversarial attacks especially the iterative attack. In this work, we propose Dual Denoising, a novel framework for learning robust and well-generalized 3D representations from CLIP. It combines a denoising-based proxy task with a novel feature denoising network for 3D pre-training. Additionally, we propose utilizing parallel noise inference to enhance the generalization of point cloud features under cross domain settings. Experiments show that our model can effectively improve the representation learning performance and adversarial robustness of the 3D learning network under zero-shot settings without adversarial training. Our code is available at https://github.com/luoshuqing2001/Dual_Denoising.

Kimi Team, Angang Du, Bohong Yin et al. · pku, tsinghua

We present Kimi-VL, an efficient open-source Mixture-of-Experts (MoE) vision-language model (VLM) that offers advanced multimodal reasoning, long-context understanding, and strong agent capabilities - all while activating only 2.8B parameters in its language decoder (Kimi-VL-A3B). Kimi-VL demonstrates strong performance across challenging domains: as a general-purpose VLM, Kimi-VL excels in multi-turn agent tasks (e.g., OSWorld), matching flagship models. Furthermore, it exhibits remarkable capabilities across diverse challenging vision language tasks, including college-level image and video comprehension, OCR, mathematical reasoning, and multi-image understanding. In comparative evaluations, it effectively competes with cutting-edge efficient VLMs such as GPT-4o-mini, Qwen2.5-VL-7B, and Gemma-3-12B-IT, while surpassing GPT-4o in several key domains. Kimi-VL also advances in processing long contexts and perceiving clearly. With a 128K extended context window, Kimi-VL can process diverse long inputs, achieving impressive scores of 64.5 on LongVideoBench and 35.1 on MMLongBench-Doc. Its native-resolution vision encoder, MoonViT, further allows it to see and understand ultra-high-resolution visual inputs, achieving 83.2 on InfoVQA and 34.5 on ScreenSpot-Pro, while maintaining lower computational cost for common tasks. Building upon Kimi-VL, we introduce an advanced long-thinking variant: Kimi-VL-Thinking-2506. Developed through long chain-of-thought (CoT) supervised fine-tuning (SFT) and reinforcement learning (RL), the latest model exhibits strong long-horizon reasoning capabilities (64.0 on MMMU, 46.3 on MMMU-Pro, 56.9 on MathVision, 80.1 on MathVista, 65.2 on VideoMMMU) while obtaining robust general abilities. Code and models are publicly accessible at https://github.com/MoonshotAI/Kimi-VL.

Kimi Team, Yifan Bai, Yiping Bao et al. · tsinghua

We introduce Kimi K2, a Mixture-of-Experts (MoE) large language model with 32 billion activated parameters and 1 trillion total parameters. We propose the MuonClip optimizer, which improves upon Muon with a novel QK-clip technique to address training instability while enjoying the advanced token efficiency of Muon. Based on MuonClip, K2 was pre-trained on 15.5 trillion tokens with zero loss spike. During post-training, K2 undergoes a multi-stage post-training process, highlighted by a large-scale agentic data synthesis pipeline and a joint reinforcement learning (RL) stage, where the model improves its capabilities through interactions with real and synthetic environments. Kimi K2 achieves state-of-the-art performance among open-source non-thinking models, with strengths in agentic capabilities. Notably, K2 obtains 66.1 on Tau2-Bench, 76.5 on ACEBench (En), 65.8 on SWE-Bench Verified, and 47.3 on SWE-Bench Multilingual -- surpassing most open and closed-sourced baselines in non-thinking settings. It also exhibits strong capabilities in coding, mathematics, and reasoning tasks, with a score of 53.7 on LiveCodeBench v6, 49.5 on AIME 2025, 75.1 on GPQA-Diamond, and 27.1 on OJBench, all without extended thinking. These results position Kimi K2 as one of the most capable open-source large language models to date, particularly in software engineering and agentic tasks. We release our base and post-trained model checkpoints to facilitate future research and applications of agentic intelligence.

Bowen Qu, Haohui Li, Wei Gao

AI-Generated Images (AGIs) have inherent multimodal nature. Unlike traditional image quality assessment (IQA) on natural scenarios, AGIs quality assessment (AGIQA) takes the correspondence of image and its textual prompt into consideration. This is coupled in the ground truth score, which confuses the unimodal IQA methods. To solve this problem, we introduce IP-IQA (AGIs Quality Assessment via Image and Prompt), a multimodal framework for AGIQA via corresponding image and prompt incorporation. Specifically, we propose a novel incremental pretraining task named Image2Prompt for better understanding of AGIs and their corresponding textual prompts. An effective and efficient image-prompt fusion module, along with a novel special [QA] token, are also applied. Both are plug-and-play and beneficial for the cooperation of image and its corresponding prompt. Experiments demonstrate that our IP-IQA achieves the state-of-the-art on AGIQA-1k and AGIQA-3k datasets. Code will be available at https://github.com/Coobiw/IP-IQA.

Bowen Qu, Xiaoyu Liang, Shangkun Sun et al.

The recent advancements in Text-to-Video Artificial Intelligence Generated Content (AIGC) have been remarkable. Compared with traditional videos, the assessment of AIGC videos encounters various challenges: visual inconsistency that defy common sense, discrepancies between content and the textual prompt, and distribution gap between various generative models, etc. Target at these challenges, in this work, we categorize the assessment of AIGC video quality into three dimensions: visual harmony, video-text consistency, and domain distribution gap. For each dimension, we design specific modules to provide a comprehensive quality assessment of AIGC videos. Furthermore, our research identifies significant variations in visual quality, fluidity, and style among videos generated by different text-to-video models. Predicting the source generative model can make the AIGC video features more discriminative, which enhances the quality assessment performance. The proposed method was used in the third-place winner of the NTIRE 2024 Quality Assessment for AI-Generated Content - Track 2 Video, demonstrating its effectiveness. Code will be available at https://github.com/Coobiw/TriVQA.

Shangkun Sun, Bowen Qu, Xiaoyu Liang et al.

Recent advances in text-driven image editing have been significant, yet the task of accurately evaluating these edited images continues to pose a considerable challenge. Different from the assessment of text-driven image generation, text-driven image editing is characterized by simultaneously conditioning on both text and a source image. The edited images often retain an intrinsic connection to the original image, which dynamically change with the semantics of the text. However, previous methods tend to solely focus on text-image alignment or have not aligned with human perception. In this work, we introduce the Text-driven Image Editing Benchmark suite (IE-Bench) to enhance the assessment of text-driven edited images. IE-Bench includes a database contains diverse source images, various editing prompts and the corresponding results different editing methods, and total 3,010 Mean Opinion Scores (MOS) provided by 25 human subjects. Furthermore, we introduce IE-QA, a multi-modality source-aware quality assessment method for text-driven image editing. To the best of our knowledge, IE-Bench offers the first IQA dataset and model tailored for text-driven image editing. Extensive experiments demonstrate IE-QA's superior subjective-alignments on the text-driven image editing task compared with previous metrics. We will make all related data and code available to the public.

Shangkun Sun, Xiaoyu Liang, Bowen Qu et al.

The advent of next-generation video generation models like \textit{Sora} poses challenges for AI-generated content (AIGC) video quality assessment (VQA). These models substantially mitigate flickering artifacts prevalent in prior models, enable longer and complex text prompts and generate longer videos with intricate, diverse motion patterns. Conventional VQA methods designed for simple text and basic motion patterns struggle to evaluate these content-rich videos. To this end, we propose \textbf{CRAVE} (\underline{C}ontent-\underline{R}ich \underline{A}IGC \underline{V}ideo \underline{E}valuator), specifically for the evaluation of Sora-era AIGC videos. CRAVE proposes the multi-granularity text-temporal fusion that aligns long-form complex textual semantics with video dynamics. Additionally, CRAVE leverages the hybrid motion-fidelity modeling to assess temporal artifacts. Furthermore, given the straightforward prompts and content in current AIGC VQA datasets, we introduce \textbf{CRAVE-DB}, a benchmark featuring content-rich videos from next-generation models paired with elaborate prompts. Extensive experiments have shown that the proposed CRAVE achieves excellent results on multiple AIGC VQA benchmarks, demonstrating a high degree of alignment with human perception. All data and code will be publicly available at https://github.com/littlespray/CRAVE.

Xiaohong Liu, Xiongkuo Min, Guangtao Zhai et al.

This paper reports on the NTIRE 2024 Quality Assessment of AI-Generated Content Challenge, which will be held in conjunction with the New Trends in Image Restoration and Enhancement Workshop (NTIRE) at CVPR 2024. This challenge is to address a major challenge in the field of image and video processing, namely, Image Quality Assessment (IQA) and Video Quality Assessment (VQA) for AI-Generated Content (AIGC). The challenge is divided into the image track and the video track. The image track uses the AIGIQA-20K, which contains 20,000 AI-Generated Images (AIGIs) generated by 15 popular generative models. The image track has a total of 318 registered participants. A total of 1,646 submissions are received in the development phase, and 221 submissions are received in the test phase. Finally, 16 participating teams submitted their models and fact sheets. The video track uses the T2VQA-DB, which contains 10,000 AI-Generated Videos (AIGVs) generated by 9 popular Text-to-Video (T2V) models. A total of 196 participants have registered in the video track. A total of 991 submissions are received in the development phase, and 185 submissions are received in the test phase. Finally, 12 participating teams submitted their models and fact sheets. Some methods have achieved better results than baseline methods, and the winning methods in both tracks have demonstrated superior prediction performance on AIGC.

Mingxing Rao, Bowen Qu, Daniel Moyer

The recovery of training data from generative models (``model inversion'') has been extensively studied for diffusion models in the data domain. The encoder/decoder pair and corresponding latent codes have largely been ignored by inversion techniques applied to latent space generative models, e.g., Latent Diffusion models (LDMs). In this work we describe two key findings: (1) The diffusion model exhibits non-uniform memorization across latent codes, tending to overfit samples located in high-distortion regions of the decoder pullback metric. (2) Even within a single latent code, different dimensions contribute unequally to memorization. We introduce a principled method to rank latent dimensions by their per-dimensional contribution to the decoder pullback metric, identifying those most responsible for memorization. Empirically, removing less-memorizing dimensions when computing attack statistics for score-based membership inference attacker significantly improves performance, with average AUROC gains of 2.7\% and substantial increases in TPR@1\%FPR (6.42\%) across diverse datasets including CIFAR-10, CelebA, ImageNet-1K, Pokémon, MS-COCO, and Flickr. This indicates stronger confidence in identifying members under extremely low false-positive tolerance. Our results highlight the overlooked influence of the auto-encoder geometry on LDM memorization and provide a new perspective for analyzing privacy risks in diffusion-based generative models.

Bowen Qu, Shangkun Sun, Xiaoyu Liang et al.

Recent advances in text-driven image editing have been significant, yet the task of accurately evaluating these edited images continues to pose a considerable challenge. Different from the assessment of text-driven image generation, text-driven image editing is characterized by simultaneously conditioning on both text and a source image. The edited images often retain an intrinsic connection to the original image, which dynamically change with the semantics of the text. However, previous methods tend to solely focus on text-image alignment or have not well aligned with human perception. In this work, we introduce the Text-driven Image Editing Benchmark suite (IE-Bench) to enhance the assessment of text-driven edited images. IE-Bench includes a database contains diverse source images, various editing prompts and the corresponding edited results from different editing methods, and nearly 4,000 samples with corresponding Mean Opinion Scores (MOS) provided by 15 human subjects. Furthermore, we introduce IE-Critic-R1, which, benefiting from Reinforcement Learning from Verifiable Rewards (RLVR), provides more comprehensive and explainable quality assessment for text-driven image editing that aligns with human perception. Extensive experiments demonstrate IE-Critic-R1's superior subjective-alignments on the text-driven image editing task compared with previous metrics. Related data and codes are available to the public.

Mingxing Rao, Bowen Qu, Daniel Moyer

Membership inference attacks (MIAs) against diffusion models have emerged as a pressing privacy concern, as these models may inadvertently reveal whether a given sample was part of their training set. We present a theoretical and empirical study of score-based MIAs, focusing on the predicted noise vectors that diffusion models learn to approximate. We show that the expected denoiser output points toward a kernel-weighted local mean of nearby training samples, such that its norm encodes proximity to the training set and thereby reveals membership. Building on this observation, we propose SimA, a single-query attack that provides a principled, efficient alternative to existing multi-query methods. SimA achieves consistently strong performance across variants of DDPM, Latent Diffusion Model (LDM). Notably, we find that Latent Diffusion Models are surprisingly less vulnerable than pixel-space models, due to the strong information bottleneck imposed by their latent auto-encoder. We further investigate this by differing the regularization hyperparameters ($β$ in $β$-VAE) in latent channel and suggest a strategy to make LDM training more robust to MIA. Our results solidify the theory of score-based MIAs, while highlighting that Latent Diffusion class of methods requires better understanding of inversion for VAE, and not simply inversion of the Diffusion process