Congcong Wang

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, 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.

Congcong Wang, Gonzalo Fiz Pontiveros, Steven Derby et al.

Despite recent advancements in Machine Learning, many tasks still involve working in low-data regimes which can make solving natural language problems difficult. Recently, a number of text augmentation techniques have emerged in the field of Natural Language Processing (NLP) which can enrich the training data with new examples, though they are not without their caveats. For instance, simple rule-based heuristic methods are effective, but lack variation in semantic content and syntactic structure with respect to the original text. On the other hand, more complex deep learning approaches can cause extreme shifts in the intrinsic meaning of the text and introduce unwanted noise into the training data. To more reliably control the quality of the augmented examples, we introduce a state-of-the-art approach for Self-Controlled Text Augmentation (STA). Our approach tightly controls the generation process by introducing a self-checking procedure to ensure that generated examples retain the semantic content of the original text. Experimental results on multiple benchmarking datasets demonstrate that STA substantially outperforms existing state-of-the-art techniques, whilst qualitative analysis reveals that the generated examples are both lexically diverse and semantically reliable.

Kimi Team, Angang Du, Bofei Gao et al. · pku, tsinghua

Language model pretraining with next token prediction has proved effective for scaling compute but is limited to the amount of available training data. Scaling reinforcement learning (RL) unlocks a new axis for the continued improvement of artificial intelligence, with the promise that large language models (LLMs) can scale their training data by learning to explore with rewards. However, prior published work has not produced competitive results. In light of this, we report on the training practice of Kimi k1.5, our latest multi-modal LLM trained with RL, including its RL training techniques, multi-modal data recipes, and infrastructure optimization. Long context scaling and improved policy optimization methods are key ingredients of our approach, which establishes a simplistic, effective RL framework without relying on more complex techniques such as Monte Carlo tree search, value functions, and process reward models. Notably, our system achieves state-of-the-art reasoning performance across multiple benchmarks and modalities -- e.g., 77.5 on AIME, 96.2 on MATH 500, 94-th percentile on Codeforces, 74.9 on MathVista -- matching OpenAI's o1. Moreover, we present effective long2short methods that use long-CoT techniques to improve short-CoT models, yielding state-of-the-art short-CoT reasoning results -- e.g., 60.8 on AIME, 94.6 on MATH500, 47.3 on LiveCodeBench -- outperforming existing short-CoT models such as GPT-4o and Claude Sonnet 3.5 by a large margin (up to +550%).

Wentao Li, Congcong Wang, Xiaoxiao Cui et al.

Open Source Intelligence (OSINT) requires the integration and reasoning of diverse multimodal data, presenting significant challenges in deriving actionable insights. Traditional approaches, including multimodal large language models (MLLMs), often struggle to infer complex contextual relationships or deliver comprehensive intelligence from unstructured data sources. In this paper, we introduce COSINT-Agent, a knowledge-driven multimodal agent tailored to address the challenges of OSINT in the Chinese domain. COSINT-Agent seamlessly integrates the perceptual capabilities of fine-tuned MLLMs with the structured reasoning power of the Entity-Event-Scene Knowledge Graph (EES-KG). Central to COSINT-Agent is the innovative EES-Match framework, which bridges COSINT-MLLM and EES-KG, enabling systematic extraction, reasoning, and contextualization of multimodal insights. This integration facilitates precise entity recognition, event interpretation, and context retrieval, effectively transforming raw multimodal data into actionable intelligence. Extensive experiments validate the superior performance of COSINT-Agent across core OSINT tasks, including entity recognition, EES generation, and context matching. These results underscore its potential as a robust and scalable solution for advancing automated multimodal reasoning and enhancing the effectiveness of OSINT methodologies.

Congcong Wang, David Lillis

In recent years, the task of mining important information from social media posts during crises has become a focus of research for the purposes of assisting emergency response (ES). The TREC Incident Streams (IS) track is a research challenge organised for this purpose. The track asks participating systems to both classify a stream of crisis-related tweets into humanitarian aid related information types and estimate their importance regarding criticality. The former refers to a multi-label information type classification task and the latter refers to a priority estimation task. In this paper, we report on the participation of the University College Dublin School of Computer Science (UCD-CS) in TREC-IS 2021. We explored a variety of approaches, including simple machine learning algorithms, multi-task learning techniques, text augmentation, and ensemble approaches. The official evaluation results indicate that our runs achieve the highest scores in many metrics. To aid reproducibility, our code is publicly available at https://github.com/wangcongcong123/crisis-mtl.

Congcong Wang, Paul Nulty, David Lillis

User-generated content (UGC) on social media can act as a key source of information for emergency responders in crisis situations. However, due to the volume concerned, computational techniques are needed to effectively filter and prioritise this content as it arises during emerging events. In the literature, these techniques are trained using annotated content from previous crises. In this paper, we investigate how this prior knowledge can be best leveraged for new crises by examining the extent to which crisis events of a similar type are more suitable for adaptation to new events (cross-domain adaptation). Given the recent successes of transformers in various language processing tasks, we propose CAST: an approach for Crisis domain Adaptation leveraging Sequence-to-sequence Transformers. We evaluate CAST using two major crisis-related message classification datasets. Our experiments show that our CAST-based best run without using any target data achieves the state of the art performance in both in-domain and cross-domain contexts. Moreover, CAST is particularly effective in one-to-one cross-domain adaptation when trained with a larger language model. In many-to-one adaptation where multiple crises are jointly used as the source domain, CAST further improves its performance. In addition, we find that more similar events are more likely to bring better adaptation performance whereas fine-tuning using dissimilar events does not help for adaptation. To aid reproducibility, we open source our code to the community.

Jun Ma, Yao Zhang, Song Gu et al.

With the unprecedented developments in deep learning, automatic segmentation of main abdominal organs seems to be a solved problem as state-of-the-art (SOTA) methods have achieved comparable results with inter-rater variability on many benchmark datasets. However, most of the existing abdominal datasets only contain single-center, single-phase, single-vendor, or single-disease cases, and it is unclear whether the excellent performance can generalize on diverse datasets. This paper presents a large and diverse abdominal CT organ segmentation dataset, termed AbdomenCT-1K, with more than 1000 (1K) CT scans from 12 medical centers, including multi-phase, multi-vendor, and multi-disease cases. Furthermore, we conduct a large-scale study for liver, kidney, spleen, and pancreas segmentation and reveal the unsolved segmentation problems of the SOTA methods, such as the limited generalization ability on distinct medical centers, phases, and unseen diseases. To advance the unsolved problems, we further build four organ segmentation benchmarks for fully supervised, semi-supervised, weakly supervised, and continual learning, which are currently challenging and active research topics. Accordingly, we develop a simple and effective method for each benchmark, which can be used as out-of-the-box methods and strong baselines. We believe the AbdomenCT-1K dataset will promote future in-depth research towards clinical applicable abdominal organ segmentation methods. The datasets, codes, and trained models are publicly available at https://github.com/JunMa11/AbdomenCT-1K.

Congcong Wang

During crisis situations, social media allows people to quickly share information, including messages requesting help. This can be valuable to emergency responders, who need to categorise and prioritise these messages based on the type of assistance being requested. However, the high volume of messages makes it difficult to filter and prioritise them without the use of computational techniques. Fully supervised filtering techniques for crisis message categorisation typically require a large amount of annotated training data, but this can be difficult to obtain during an ongoing crisis and is expensive in terms of time and labour to create. This thesis focuses on addressing the challenge of low data availability when categorising crisis messages for emergency response. It first presents domain adaptation as a solution for this problem, which involves learning a categorisation model from annotated data from past crisis events (source domain) and adapting it to categorise messages from an ongoing crisis event (target domain). In many-to-many adaptation, where the model is trained on multiple past events and adapted to multiple ongoing events, a multi-task learning approach is proposed using pre-trained language models. This approach outperforms baselines and an ensemble approach further improves performance...

Congcong Wang, Paul Nulty, David Lillis

Social media has enabled people to circulate information in a timely fashion, thus motivating people to post messages seeking help during crisis situations. These messages can contribute to the situational awareness of emergency responders, who have a need for them to be categorised according to information types (i.e. the type of aid services the messages are requesting). We introduce a transformer-based multi-task learning (MTL) technique for classifying information types and estimating the priority of these messages. We evaluate the effectiveness of our approach with a variety of metrics by submitting runs to the TREC Incident Streams (IS) track: a research initiative specifically designed for disaster tweet classification and prioritisation. The results demonstrate that our approach achieves competitive performance in most metrics as compared to other participating runs. Subsequently, we find that an ensemble approach combining disparate transformer encoders within our approach helps to improve the overall effectiveness to a significant extent, achieving state-of-the-art performance in almost every metric. We make the code publicly available so that our work can be reproduced and used as a baseline for the community for future work in this domain.

Congcong Wang, David Lillis

The Incident streams (IS) track is a research challenge aimed at finding important information from social media during crises for emergency response purposes. More specifically, given a stream of crisis-related tweets, the IS challenge asks a participating system to 1) classify what the types of users' concerns or needs are expressed in each tweet, known as the information type (IT) classification task and 2) estimate how critical each tweet is with regard to emergency response, known as the priority level prediction task. In this paper, we describe our multi-task transfer learning approach for this challenge. Our approach leverages state-of-the-art transformer models including both encoder-based models such as BERT and a sequence-to-sequence based T5 for joint transfer learning on the two tasks. Based on this approach, we submitted several runs to the track. The returned evaluation results show that our runs substantially outperform other participating runs in both IT classification and priority level prediction.

Max Allan, Jonathan Mcleod, Congcong Wang et al.

The stereo correspondence and reconstruction of endoscopic data sub-challenge was organized during the Endovis challenge at MICCAI 2019 in Shenzhen, China. The task was to perform dense depth estimation using 7 training datasets and 2 test sets of structured light data captured using porcine cadavers. These were provided by a team at Intuitive Surgical. 10 teams participated in the challenge day. This paper contains 3 additional methods which were submitted after the challenge finished as well as a supplemental section from these teams on issues they found with the dataset.

Congcong Wang, David Lillis

In this paper, we describe our approach in the shared task: COVID-19 event extraction from Twitter. The objective of this task is to extract answers from COVID-related tweets to a set of predefined slot-filling questions. Our approach treats the event extraction task as a question answering task by leveraging the transformer-based T5 text-to-text model. According to the official evaluation scores returned, namely F1, our submitted run achieves competitive performance compared to other participating runs (Top 3). However, we argue that this evaluation may underestimate the actual performance of runs based on text-generation. Although some such runs may answer the slot questions well, they may not be an exact string match for the gold standard answers. To measure the extent of this underestimation, we adopt a simple exact-answer transformation method aiming at converting the well-answered predictions to exactly-matched predictions. The results show that after this transformation our run overall reaches the same level of performance as the best participating run and state-of-the-art F1 scores in three of five COVID-related events. Our code is publicly available to aid reproducibility

Congcong Wang, Faouzi Alaya Cheikh, Azeddine Beghdadi et al.

The object sizes in images are diverse, therefore, capturing multiple scale context information is essential for semantic segmentation. Existing context aggregation methods such as pyramid pooling module (PPM) and atrous spatial pyramid pooling (ASPP) design different pooling size or atrous rate, such that multiple scale information is captured. However, the pooling sizes and atrous rates are chosen manually and empirically. In order to capture object context information adaptively, in this paper, we propose an adaptive context encoding (ACE) module based on deformable convolution operation to argument multiple scale information. Our ACE module can be embedded into other Convolutional Neural Networks (CNN) easily for context aggregation. The effectiveness of the proposed module is demonstrated on Pascal-Context and ADE20K datasets. Although our proposed ACE only consists of three deformable convolution blocks, it outperforms PPM and ASPP in terms of mean Intersection of Union (mIoU) on both datasets. All the experiment study confirms that our proposed module is effective as compared to the state-of-the-art methods.

Oleksii Sidorov, Congcong Wang, Faouzi Alaya Cheikh

In minimally invasive surgery, the use of tissue dissection tools causes smoke, which inevitably degrades the image quality. This could reduce the visibility of the operation field for surgeons and introduces errors for the computer vision algorithms used in surgical navigation systems. In this paper, we propose a novel approach for computational smoke removal using supervised image-to-image translation. We demonstrate that straightforward application of existing generative algorithms allows removing smoke but decreases image quality and introduces synthetic noise (grid-structure). Thus, we propose to solve this issue by modification of GAN's architecture and adding perceptual image quality metric to the loss function. Obtained results demonstrate that proposed method efficiently removes smoke as well as preserves perceptually sufficient image quality.

Congcong Wang, Vivek Sharma, Yu Fan et al.

Laparoscopic surgery has a limited field of view. Laser ablation in a laproscopic surgery causes smoke, which inevitably influences the surgeon's visibility. Therefore, it is of vital importance to remove the smoke, such that a clear visualization is possible. In order to employ a desmoking technique, one needs to know beforehand if the image contains smoke or not, to this date, there exists no accurate method that could classify the smoke/non-smoke images completely. In this work, we propose a new enhancement method which enhances the informative details in the RGB images for discrimination of smoke/non-smoke images. Our proposed method utilizes weighted least squares optimization framework~(WLS). For feature extraction, we use statistical features based on bivariate histogram distribution of gradient magnitude~(GM) and Laplacian of Gaussian~(LoG). We then train a SVM classifier with binary smoke/non-smoke classification task. We demonstrate the effectiveness of our method on Cholec80 dataset. Experiments using our proposed enhancement method show promising results with improvements of 4\% in accuracy and 4\% in F1-Score over the baseline performance of RGB images. In addition, our approach improves over the saturation histogram based classification methodologies Saturation Analysis~(SAN) and Saturation Peak Analysis~(SPA) by 1/5\% and 1/6\% in accuracy/F1-Score metrics.

Congcong Wang, Faouzi Alaya Cheikh, Mounir Kaaniche et al.

In laparoscopic surgery, image quality can be severely degraded by surgical smoke, which not only introduces error for the image processing (used in image guided surgery), but also reduces the visibility of the surgeons. In this paper, we propose to enhance the laparoscopic images by decomposing them into unwanted smoke part and enhanced part using a variational approach. The proposed method relies on the observation that smoke has low contrast and low inter-channel differences. A cost function is defined based on this prior knowledge and is solved using an augmented Lagrangian method. The obtained unwanted smoke component is then subtracted from the original degraded image, resulting in the enhanced image. The obtained quantitative scores in terms of FADE, JNBM and RE metrics show that our proposed method performs rather well. Furthermore, the qualitative visual inspection of the results show that it removes smoke effectively from the laparoscopic images.