Kimi-VL Technical Report

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. Developed through long chain-of-thought (CoT) supervised fine-tuning (SFT) and reinforcement learning (RL), this model exhibits strong long-horizon reasoning capabilities. It achieves scores of 61.7 on MMMU, 36.8 on MathVision, and 71.3 on MathVista while maintaining the compact 2.8B activated LLM parameters, setting a new standard for efficient multimodal thinking models. Code and models are publicly accessible at https://github.com/MoonshotAI/Kimi-VL.

Endo3R: Unified Online Reconstruction from Dynamic Monocular Endoscopic Video

Reconstructing 3D scenes from monocular surgical videos can enhance surgeon's perception and therefore plays a vital role in various computer-assisted surgery tasks. However, achieving scale-consistent reconstruction remains an open challenge due to inherent issues in endoscopic videos, such as dynamic deformations and textureless surfaces. Despite recent advances, current methods either rely on calibration or instrument priors to estimate scale, or employ SfM-like multi-stage pipelines, leading to error accumulation and requiring offline optimization. In this paper, we present Endo3R, a unified 3D foundation model for online scale-consistent reconstruction from monocular surgical video, without any priors or extra optimization. Our model unifies the tasks by predicting globally aligned pointmaps, scale-consistent video depths, and camera parameters without any offline optimization. The core contribution of our method is expanding the capability of the recent pairwise reconstruction model to long-term incremental dynamic reconstruction by an uncertainty-aware dual memory mechanism. The mechanism maintains history tokens of both short-term dynamics and long-term spatial consistency. Notably, to tackle the highly dynamic nature of surgical scenes, we measure the uncertainty of tokens via Sampson distance and filter out tokens with high uncertainty. Regarding the scarcity of endoscopic datasets with ground-truth depth and camera poses, we further devise a self-supervised mechanism with a novel dynamics-aware flow loss. Abundant experiments on SCARED and Hamlyn datasets demonstrate our superior performance in zero-shot surgical video depth prediction and camera pose estimation with online efficiency. Project page: https://wrld.github.io/Endo3R/.

dARt Vinci: Egocentric Data Collection for Surgical Robot Learning at Scale

Data scarcity has long been an issue in the robot learning community. Particularly, in safety-critical domains like surgical applications, obtaining high-quality data can be especially difficult. It poses challenges to researchers seeking to exploit recent advancements in reinforcement learning and imitation learning, which have greatly improved generalizability and enabled robots to conduct tasks autonomously. We introduce dARt Vinci, a scalable data collection platform for robot learning in surgical settings. The system uses Augmented Reality (AR) hand tracking and a high-fidelity physics engine to capture subtle maneuvers in primitive surgical tasks: By eliminating the need for a physical robot setup and providing flexibility in terms of time, space, and hardware resources-such as multiview sensors and actuators-specialized simulation is a viable alternative. At the same time, AR allows the robot data collection to be more egocentric, supported by its body tracking and content overlaying capabilities. Our user study confirms the proposed system's efficiency and usability, where we use widely-used primitive tasks for training teleoperation with da Vinci surgical robots. Data throughput improves across all tasks compared to real robot settings by 41% on average. The total experiment time is reduced by an average of 10%. The temporal demand in the task load survey is improved. These gains are statistically significant. Additionally, the collected data is over 400 times smaller in size, requiring far less storage while achieving double the frequency.

OPG-Policy: Occluded Push-Grasp Policy Learning with Amodal Segmentation

Goal-oriented grasping in dense clutter, a fundamental challenge in robotics, demands an adaptive policy to handle occluded target objects and diverse configurations. Previous methods typically learn policies based on partially observable segments of the occluded target to generate motions. However, these policies often struggle to generate optimal motions due to uncertainties regarding the invisible portions of different occluded target objects across various scenes, resulting in low motion efficiency. To this end, we propose OPG-Policy, a novel framework that leverages amodal segmentation to predict occluded portions of the target and develop an adaptive push-grasp policy for cluttered scenarios where the target object is partially observed. Specifically, our approach trains a dedicated amodal segmentation module for diverse target objects to generate amodal masks. These masks and scene observations are mapped to the future rewards of grasp and push motion primitives via deep Q-learning to learn the motion critic. Afterward, the push and grasp motion candidates predicted by the critic, along with the relevant domain knowledge, are fed into the coordinator to generate the optimal motion implemented by the robot. Extensive experiments conducted in both simulated and real-world environments demonstrate the effectiveness of our approach in generating motion sequences for retrieving occluded targets, outperforming other baseline methods in success rate and motion efficiency.

Neural Finite-State Machines for Surgical Phase Recognition

Surgical phase recognition is essential for analyzing procedure-specific surgical videos. While recent transformer-based architectures have advanced sequence processing capabilities, they struggle with maintaining consistency across lengthy surgical procedures. Drawing inspiration from classical hidden Markov models' finite-state interpretations, we introduce the neural finite-state machine (NFSM) module, which bridges procedural understanding with deep learning approaches. NFSM combines procedure-level understanding with neural networks through global state embeddings, attention-based dynamic transition tables, and transition-aware training and inference mechanisms for offline and online applications. When integrated into our future-aware architecture, NFSM improves video-level accuracy, phase-level precision, recall, and Jaccard indices on Cholec80 datasets by 2.3, 3.2, 3.0, and 4.8 percentage points respectively. As an add-on module to existing state-of-the-art models like Surgformer, NFSM further enhances performance, demonstrating its complementary value. Extended experiments on non-surgical datasets validate NFSM's generalizability beyond surgical domains. Comprehensive experiments demonstrate that incorporating NSFM into deep learning frameworks enables more robust and consistent phase recognition across long procedural videos.

Language-Model Prior Overcomes Cold-Start Items

The growth of recommender systems (RecSys) is driven by digitization and the need for personalized content in areas such as e-commerce and video streaming. The content in these systems often changes rapidly and therefore they constantly face the ongoing cold-start problem, where new items lack interaction data and are hard to value. Existing solutions for the cold-start problem, such as content-based recommenders and hybrid methods, leverage item metadata to determine item similarities. The main challenge with these methods is their reliance on structured and informative metadata to capture detailed item similarities, which may not always be available. This paper introduces a novel approach for cold-start item recommendation that utilizes the language model (LM) to estimate item similarities, which are further integrated as a Bayesian prior with classic recommender systems. This approach is generic and able to boost the performance of various recommenders. Specifically, our experiments integrate it with both sequential and collaborative filtering-based recommender and evaluate it on two real-world datasets, demonstrating the enhanced performance of the proposed approach.

Towards Robust Algorithms for Surgical Phase Recognition via Digital Twin-based Scene Representation

Purpose: Surgical phase recognition (SPR) is an integral component of surgical data science, enabling high-level surgical analysis. End-to-end trained neural networks that predict surgical phase directly from videos have shown excellent performance on benchmarks. However, these models struggle with robustness due to non-causal associations in the training set, resulting in poor generalizability. Our goal is to improve model robustness to variations in the surgical videos by leveraging the digital twin (DT) paradigm -- an intermediary layer to separate high-level analysis (SPR) from low-level processing (geometric understanding). This approach takes advantage of the recent vision foundation models that ensure reliable low-level scene understanding to craft DT-based scene representations that support various high-level tasks. Methods: We present a DT-based framework for SPR from videos. The framework employs vision foundation models to extract representations. We embed the representation in place of raw video inputs in the state-of-the-art Surgformer model. The framework is trained on the Cholec80 dataset and evaluated on out-of-distribution (OOD) and corrupted test samples. Results: Contrary to the vulnerability of the baseline model, our framework demonstrates strong robustness on both OOD and corrupted samples, with a video-level accuracy of 51.1 on the challenging CRCD dataset, 96.0 on an internal robotics training dataset, and 64.4 on a highly corrupted Cholec80 test set. Conclusion: Our findings lend support to the thesis that DT-based scene representations are effective in enhancing model robustness. Future work will seek to improve the feature informativeness, automate feature extraction, and incorporate interpretability for a more comprehensive framework.

DaRePlane: Direction-aware Representations for Dynamic Scene Reconstruction

Numerous recent approaches to modeling and re-rendering dynamic scenes leverage plane-based explicit representations, addressing slow training times associated with models like neural radiance fields (NeRF) and Gaussian splatting (GS). However, merely decomposing 4D dynamic scenes into multiple 2D plane-based representations is insufficient for high-fidelity re-rendering of scenes with complex motions. In response, we present DaRePlane, a novel direction-aware representation approach that captures scene dynamics from six different directions. This learned representation undergoes an inverse dual-tree complex wavelet transformation (DTCWT) to recover plane-based information. Within NeRF pipelines, DaRePlane computes features for each space-time point by fusing vectors from these recovered planes, then passed to a tiny MLP for color regression. When applied to Gaussian splatting, DaRePlane computes the features of Gaussian points, followed by a tiny multi-head MLP for spatial-time deformation prediction. Notably, to address redundancy introduced by the six real and six imaginary direction-aware wavelet coefficients, we introduce a trainable masking approach, mitigating storage issues without significant performance decline. To demonstrate the generality and efficiency of DaRePlane, we test it on both regular and surgical dynamic scenes, for both NeRF and GS systems. Extensive experiments show that DaRePlane yields state-of-the-art performance in novel view synthesis for various complex dynamic scenes.

Plug-and-Play Performance Estimation for LLM Services without Relying on Labeled Data

Large Language Model (LLM) services exhibit impressive capability on unlearned tasks leveraging only a few examples by in-context learning (ICL). However, the success of ICL varies depending on the task and context, leading to heterogeneous service quality. Directly estimating the performance of LLM services at each invocation can be laborious, especially requiring abundant labeled data or internal information within the LLM. This paper introduces a novel method to estimate the performance of LLM services across different tasks and contexts, which can be "plug-and-play" utilizing only a few unlabeled samples like ICL. Our findings suggest that the negative log-likelihood and perplexity derived from LLM service invocation can function as effective and significant features. Based on these features, we utilize four distinct meta-models to estimate the performance of LLM services. Our proposed method is compared against unlabeled estimation baselines across multiple LLM services and tasks. And it is experimentally applied to two scenarios, demonstrating its effectiveness in the selection and further optimization of LLM services.

Performance and Non-adversarial Robustness of the Segment Anything Model 2 in Surgical Video Segmentation

Fully supervised deep learning (DL) models for surgical video segmentation have been shown to struggle with non-adversarial, real-world corruptions of image quality including smoke, bleeding, and low illumination. Foundation models for image segmentation, such as the segment anything model (SAM) that focuses on interactive prompt-based segmentation, move away from semantic classes and thus can be trained on larger and more diverse data, which offers outstanding zero-shot generalization with appropriate user prompts. Recently, building upon this success, SAM-2 has been proposed to further extend the zero-shot interactive segmentation capabilities from independent frame-by-frame to video segmentation. In this paper, we present a first experimental study evaluating SAM-2's performance on surgical video data. Leveraging the SegSTRONG-C MICCAI EndoVIS 2024 sub-challenge dataset, we assess SAM-2's effectiveness on uncorrupted endoscopic sequences and evaluate its non-adversarial robustness on videos with corrupted image quality simulating smoke, bleeding, and low brightness conditions under various prompt strategies. Our experiments demonstrate that SAM-2, in zero-shot manner, can achieve competitive or even superior performance compared to fully-supervised deep learning models on surgical video data, including under non-adversarial corruptions of image quality. Additionally, SAM-2 consistently outperforms the original SAM and its medical variants across all conditions. Finally, frame-sparse prompting can consistently outperform frame-wise prompting for SAM-2, suggesting that allowing SAM-2 to leverage its temporal modeling capabilities leads to more coherent and accurate segmentation compared to frequent prompting.