Zijian Wu

Zhangchen Xu, Junda Chen, Yue Huang et al.

Scientific and engineering progress is fundamentally a long-horizon iterative process: proposing changes, running experiments, measuring outcomes, and continuously refining artifacts. Yet existing benchmarks for frontier models primarily evaluate either single-turn responses or short-horizon agent trajectories, failing to capture the challenges of sustained iterative improvement over extended time horizons. To address this gap, we introduce AutoLab, a new benchmark for ultra long-horizon closed-loop optimization. AutoLab consists of 36 realistic, expert-curated tasks spanning four diverse domains: system optimization, puzzle & challenge, model development, and CUDA kernel optimization. Each task begins with a correct but deliberately suboptimal baseline and challenges agents to improve it within a strict wall-clock budget. Evaluating 17 state-of-the-art models reveals the dominant predictor of success is not the quality of an agent's initial attempt, but its persistence in repeatedly benchmarking, editing, and incorporating empirical feedback. While claude-opus-4.6 exhibits strong long-horizon optimization capabilities, most frontier models, including several proprietary ones, either terminate prematurely or exhaust their budgets with minimal progress. These results underscore the importance of time awareness and persistent iteration in autonomous agents. We open-source the full benchmark, evaluation harness, and task artifacts, to accelerate research toward truly capable long-horizon agents.

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.

Yicheng Zou, Dongsheng Zhu, Lin Zhu et al.

We introduce Intern-S1-Pro, the first one-trillion-parameter scientific multimodal foundation model. Scaling to this unprecedented size, the model delivers a comprehensive enhancement across both general and scientific domains. Beyond stronger reasoning and image-text understanding capabilities, its intelligence is augmented with advanced agent capabilities. Simultaneously, its scientific expertise has been vastly expanded to master over 100 specialized tasks across critical science fields, including chemistry, materials, life sciences, and earth sciences. Achieving this massive scale is made possible by the robust infrastructure support of XTuner and LMDeploy, which facilitates highly efficient Reinforcement Learning (RL) training at the 1-trillion parameter level while ensuring strict precision consistency between training and inference. By seamlessly integrating these advancements, Intern-S1-Pro further fortifies the fusion of general and specialized intelligence, working as a Specializable Generalist, demonstrating its position in the top tier of open-source models for general capabilities, while outperforming proprietary models in the depth of specialized scientific tasks.

Zijian Wu, Jiayu Wang, Dahua Lin et al.

Recently, large language models have presented promising results in aiding formal mathematical reasoning. However, their performance is restricted due to the scarcity of formal theorem-proving data, which requires additional effort to be extracted from raw formal language corpora. Meanwhile, a significant amount of human-written formal language corpora remains underutilized. To address this issue, we propose LEAN-GitHub, a dataset consisting of large-scale formal data extracted from almost all Lean 4 repositories on GitHub. After fine-tuning InternLM-math-plus on this dataset, our model achieved accuracies of 48.8% with a single pass and 54.5% with 64 passes on the Lean 4 miniF2F test, surpassing state-of-the-art method at 52%. And it also achieves state-of-the-art on two other Lean 4 benchmarks (ProofNet and Putnam) targeting different fields/levels of math. These results demonstrate that our proposed dataset is beneficial for formal reasoning on a wide range of math topics. We open-source our model at https://GitHub. com/InternLM/InternLM-Math and our data at https://huggingface.co/ datasets/InternLM/Lean-GitHub

Zijian Wu, Lingkai Kong, Wenwei Zhang et al.

Large language models (LLMs) have achieved significant progress in solving complex reasoning tasks by Reinforcement Learning with Verifiable Rewards (RLVR). This advancement is also inseparable from the oversight automated by reliable verifiers. However, current outcome-based verifiers (OVs) are unable to inspect the unreliable intermediate steps in the long reasoning chains of thought (CoTs). Meanwhile, current process-based verifiers (PVs) have difficulties in reliably detecting errors in the complex long CoTs, limited by the scarcity of high-quality annotations due to the prohibitive costs of human annotations. Therefore, we propose the Outcome-based Process Verifier (OPV), which verifies the rationale process of summarized outcomes from long CoTs to achieve both accurate and efficient verification and enable large-scale annotation. To empower the proposed verifier, we adopt an iterative active learning framework with expert annotations to progressively improve the verification capability of OPV with fewer annotation costs. Specifically, in each iteration, the most uncertain cases of the current best OPV are annotated and then subsequently used to train a new OPV through Rejection Fine-Tuning (RFT) and RLVR for the next round. Extensive experiments demonstrate OPV's superior performance and broad applicability. It achieves new state-of-the-art results on our held-out OPV-Bench, outperforming much larger open-source models such as Qwen3-Max-Preview with an F1 score of 83.1 compared to 76.3. Furthermore, OPV effectively detects false positives within synthetic dataset, closely align with expert assessment. When collaborating with policy models, OPV consistently yields performance gains, e.g., raising the accuracy of DeepSeek-R1-Distill-Qwen-32B from 55.2% to 73.3% on AIME2025 as the compute budget scales.

Open-H-Embodiment Consortium, Nigel Nelson, Juo-Tung Chen et al.

Autonomous medical robots hold promise to improve patient outcomes, reduce provider workload, democratize access to care, and enable superhuman precision. However, autonomous medical robotics has been limited by a fundamental data problem: existing medical robotic datasets are small, single-embodiment, and rarely shared openly, restricting the development of foundation models that the field needs to advance. We introduce Open-H-Embodiment, the largest open dataset of medical robotic video with synchronized kinematics to date, spanning more than 49 institutions and multiple robotic platforms including the CMR Versius, Intuitive Surgical's da Vinci, da Vinci Research Kit (dVRK), Rob Surgical BiTrack, Virtual Incision's MIRA, Moon Surgical Maestro, and a variety of custom systems, spanning surgical manipulation, robotic ultrasound, and endoscopy procedures. We demonstrate the research enabled by this dataset through two foundation models. GR00T-H is the first open foundation vision-language-action model for medical robotics, which is the only evaluated model to achieve full end-to-end task completion on a structured suturing benchmark (25% of trials vs. 0% for all others) and achieves 64% average success across a 29-step ex vivo suturing sequence. We also train Cosmos-H-Surgical-Simulator, the first action-conditioned world model to enable multi-embodiment surgical simulation from a single checkpoint, spanning nine robotic platforms and supporting in silico policy evaluation and synthetic data generation for the medical domain. These results suggest that open, large-scale medical robot data collection can serve as critical infrastructure for the research community, enabling advances in robot learning, world modeling, and beyond.

Huaiyuan Ying, Shuo Zhang, Linyang Li et al. · pku

The math abilities of large language models can represent their abstract reasoning ability. In this paper, we introduce and open-source our math reasoning LLMs InternLM-Math which is continue pre-trained from InternLM2. We unify chain-of-thought reasoning, reward modeling, formal reasoning, data augmentation, and code interpreter in a unified seq2seq format and supervise our model to be a versatile math reasoner, verifier, prover, and augmenter. These abilities can be used to develop the next math LLMs or self-iteration. InternLM-Math obtains open-sourced state-of-the-art performance under the setting of in-context learning, supervised fine-tuning, and code-assisted reasoning in various informal and formal benchmarks including GSM8K, MATH, Hungary math exam, MathBench-ZH, and MiniF2F. Our pre-trained model achieves 30.3 on the MiniF2F test set without fine-tuning. We further explore how to use LEAN to solve math problems and study its performance under the setting of multi-task learning which shows the possibility of using LEAN as a unified platform for solving and proving in math. Our models, codes, and data are released at \url{https://github.com/InternLM/InternLM-Math}.

Hao Ding, Yuqian Zhang, Tuxun Lu et al.

Surgical data science has seen rapid advancement due to the excellent performance of end-to-end deep neural networks (DNNs) for surgical video analysis. Despite their successes, end-to-end DNNs have been proven susceptible to even minor corruptions, substantially impairing the model's performance. This vulnerability has become a major concern for the translation of cutting-edge technology, especially for high-stakes decision-making in surgical data science. We introduce SegSTRONG-C, a benchmark and challenge in surgical data science dedicated, aiming to better understand model deterioration under unforeseen but plausible non-adversarial corruption and the capabilities of contemporary methods that seek to improve it. Through comprehensive baseline experiments and participating submissions from widespread community engagement, SegSTRONG-C reveals key themes for model failure and identifies promising directions for improving robustness. The performance of challenge winners, achieving an average 0.9394 DSC and 0.9301 NSD across the unreleased test sets with corruption types: bleeding, smoke, and low brightness, shows inspiring improvement of 0.1471 DSC and 0.2584 NSD in average comparing to strongest baseline methods with UNet architecture trained with AutoAugment. In conclusion, the SegSTRONG-C challenge has identified some practical approaches for enhancing model robustness, yet most approaches relied on conventional techniques that have known, and sometimes quite severe, limitations. Looking ahead, we advocate for expanding intellectual diversity and creativity in non-adversarial robustness beyond data augmentation or training scale, calling for new paradigms that enhance universal robustness to corruptions and may enable richer applications in surgical data science.

Xiangyan Liu, Jinjie Ni, Zijian Wu et al.

Recent advances in reinforcement learning (RL) have strengthened the reasoning capabilities of vision-language models (VLMs). However, enhancing policy exploration to better scale test-time compute remains largely underexplored. In addition, VLMs continue to struggle with imperfect visual perception, which in turn affects the subsequent reasoning process. We introduce NoisyRollout, a simple yet effective data augmentation method that addresses these issues by mixing training trajectories from both clean and moderately distorted images. This approach injects perceptual diversity, encouraging better policy exploration and leading to more robust reasoning. A noise annealing schedule gradually reduces distortion strength, aiding exploration early in training while ensuring later stability. Crucially, our method is easy-to-adopt--requiring no additional training cost and no modifications to the RL objective. Extensive experiments on 2 distinct training datasets demonstrate that NoisyRollout achieves state-of-the-art performance among open-source RL-tuned models across 5 out-of-domain reasoning and perception benchmarks. Furthermore, we validate the effectiveness of NoisyRollout across model sizes (7B and 32B), data scales (from 1K to 6K) and image augmentation types (Gaussion noise and rotation), highlighting its generalizability and scalability.

Lei Bai, Zhongrui Cai, Yuhang Cao et al.

In recent years, a plethora of open-source foundation models have emerged, achieving remarkable progress in some widely attended fields, with performance being quite close to that of closed-source models. However, in high-value but more challenging scientific professional fields, either the fields still rely on expert models, or the progress of general foundation models lags significantly compared to those in popular areas, far from sufficient for transforming scientific research and leaving substantial gap between open-source models and closed-source models in these scientific domains. To mitigate this gap and explore a step further toward Artificial General Intelligence (AGI), we introduce Intern-S1, a specialized generalist equipped with general understanding and reasoning capabilities with expertise to analyze multiple science modal data. Intern-S1 is a multimodal Mixture-of-Experts (MoE) model with 28 billion activated parameters and 241 billion total parameters, continually pre-trained on 5T tokens, including over 2.5T tokens from scientific domains. In the post-training stage, Intern-S1 undergoes offline and then online reinforcement learning (RL) in InternBootCamp, where we propose Mixture-of-Rewards (MoR) to synergize the RL training on more than 1000 tasks simultaneously. Through integrated innovations in algorithms, data, and training systems, Intern-S1 achieved top-tier performance in online RL training. On comprehensive evaluation benchmarks, Intern-S1 demonstrates competitive performance on general reasoning tasks among open-source models and significantly outperforms open-source models in scientific domains, surpassing closed-source state-of-the-art models in professional tasks, such as molecular synthesis planning, reaction condition prediction, predicting thermodynamic stabilities for crystals. Our models are available at https://huggingface.co/internlm/Intern-S1.

Leheng Sheng, An Zhang, Zijian Wu et al.

Recent studies empirically reveal that large reasoning models (LRMs) can automatically allocate more reasoning strengths (i.e., the number of reasoning tokens) for harder problems, exhibiting difficulty-awareness for better task performance. While this automatic reasoning strength allocation phenomenon has been widely observed, its underlying mechanism remains largely unexplored. To this end, we provide explanations for this phenomenon from the perspective of model activations. We find evidence that LRMs pre-plan the reasoning strengths in their activations even before generation, with this reasoning strength causally controlled by the magnitude of a pre-allocated directional vector. Specifically, we show that the number of reasoning tokens is predictable solely based on the question activations using linear probes, indicating that LRMs estimate the required reasoning strength in advance. We then uncover that LRMs encode this reasoning strength through a pre-allocated directional vector embedded in the activations of the model, where the vector's magnitude modulates the reasoning strength. Subtracting this vector can lead to reduced reasoning token number and performance, while adding this vector can lead to increased reasoning token number and even improved performance. We further reveal that this direction vector consistently yields positive reasoning length prediction, and it modifies the logits of end-of-reasoning token </think> to affect the reasoning length. Finally, we demonstrate two potential applications of our findings: overthinking behavior detection and enabling efficient reasoning on simple problems. Our work provides new insights into the internal mechanisms of reasoning in LRMs and offers practical tools for controlling their reasoning behaviors. Our code is available at https://github.com/AlphaLab-USTC/LRM-plans-CoT.

Zijian Wu, Adam Schmidt, Peter Kazanzides et al.

The Segment Anything Model (SAM) is a powerful vision foundation model that is revolutionizing the traditional paradigm of segmentation. Despite this, a reliance on prompting each frame and large computational cost limit its usage in robotically assisted surgery. Applications, such as augmented reality guidance, require little user intervention along with efficient inference to be usable clinically. In this study, we address these limitations by adopting lightweight SAM variants to meet the efficiency requirement and employing fine-tuning techniques to enhance their generalization in surgical scenes. Recent advancements in Tracking Any Point (TAP) have shown promising results in both accuracy and efficiency, particularly when points are occluded or leave the field of view. Inspired by this progress, we present a novel framework that combines an online point tracker with a lightweight SAM model that is fine-tuned for surgical instrument segmentation. Sparse points within the region of interest are tracked and used to prompt SAM throughout the video sequence, providing temporal consistency. The quantitative results surpass the state-of-the-art semi-supervised video object segmentation method XMem on the EndoVis 2015 dataset with 84.8 IoU and 91.0 Dice. Our method achieves promising performance that is comparable to XMem and transformer-based fully supervised segmentation methods on ex vivo UCL dVRK and in vivo CholecSeg8k datasets. In addition, the proposed method shows promising zero-shot generalization ability on the label-free STIR dataset. In terms of efficiency, we tested our method on a single GeForce RTX 4060/4090 GPU respectively, achieving an over 25/90 FPS inference speed. Code is available at: https://github.com/wuzijian1997/SIS-PT-SAM

Adam Schmidt, Mert Asim Karaoglu, Soham Sinha et al.

Understanding tissue motion in surgery is crucial to enable applications in downstream tasks such as segmentation, 3D reconstruction, virtual tissue landmarking, autonomous probe-based scanning, and subtask autonomy. Labeled data are essential to enabling algorithms in these downstream tasks since they allow us to quantify and train algorithms. This paper introduces a point tracking challenge to address this, wherein participants can submit their algorithms for quantification. The submitted algorithms are evaluated using a dataset named surgical tattoos in infrared (STIR), with the challenge aptly named the STIR Challenge 2024. The STIR Challenge 2024 comprises two quantitative components: accuracy and efficiency. The accuracy component tests the accuracy of algorithms on in vivo and ex vivo sequences. The efficiency component tests the latency of algorithm inference. The challenge was conducted as a part of MICCAI EndoVis 2024. In this challenge, we had 8 total teams, with 4 teams submitting before and 4 submitting after challenge day. This paper details the STIR Challenge 2024, which serves to move the field towards more accurate and efficient algorithms for spatial understanding in surgery. In this paper we summarize the design, submissions, and results from the challenge. The challenge dataset is available here: https://zenodo.org/records/14803158 , and the code for baseline models and metric calculation is available here: https://github.com/athaddius/STIRMetrics

Zijian Wu, Boyao Zhou, Liangxiao Hu et al.

We present UIKA, a feed-forward animatable Gaussian head model from an arbitrary number of unposed inputs, including a single image, multi-view captures, and smartphone-captured videos. Unlike the traditional avatar method, which requires a studio-level multi-view capture system and reconstructs a human-specific model through a long-time optimization process, we rethink the task through the lenses of model representation, network design, and data preparation. First, we introduce a UV-guided avatar modeling strategy, in which each input image is associated with a pixel-wise facial correspondence estimation. Such correspondence estimation allows us to reproject each valid pixel color from screen space to UV space, which is independent of camera pose and character expression. Furthermore, we design learnable UV tokens on which the attention mechanism can be applied at both the screen and UV levels. The learned UV tokens can be decoded into canonical Gaussian attributes using aggregated UV information from all input views. To train our large avatar model, we additionally prepare a large-scale, identity-rich synthetic training dataset. Our method significantly outperforms existing approaches in both monocular and multi-view settings. Project page: https://zijian-wu.github.io/uika-page/

Zijian Wu, Suozhi Huang, Zhejian Zhou et al. · pku

Large Language Models (LLMs) have emerged as powerful tools in mathematical theorem proving, particularly when utilizing formal languages such as LEAN. A prevalent proof method involves the LLM prover iteratively constructing the proof tactic by tactic, typically following a best-first search scheme. However, this method often ignores the critical preference information inside the existing tactic trajectories, hindering the search for deeper proofs. We propose an intuitive yet effective method, which utilizes a critic model to capture the preference information and to guide the search of the prover model at runtime. Given the prover-critic framework, a large-scale expert iteration with more than 20,000 CPU days is then applied to further fine-tune the prover and the critic. The trained InternLM2.5-StepProver critic significantly boosts the performance of the prover model (59.4% to 65.9%). We also analyze the impact of the critic on various aspects of the theorem proving process during expert iteration, providing insights into its effectiveness. We open-source our models and searched proofs at https://github.com/InternLM/InternLM-Math and https://huggingface.co/datasets/internlm/Lean-Workbook.

Shuojue Yang, Zijian Wu, Chengjiaao Liao et al.

High-quality and controllable digital twins of surgical instruments are critical for Real2Sim in robot-assisted surgery, as they enable realistic simulation, synthetic data generation, and perception learning under novel poses. We present Instrument-Splatting++, a monocular 3D Gaussian Splatting (3DGS) framework that reconstructs surgical instruments as a fully controllable Gaussian asset with high fidelity. Our pipeline starts with part-wise geometry pretraining that injects CAD priors into Gaussian primitives and equips the representation with part-aware semantic rendering. Built on the pretrained model, we propose a semantics-aware pose estimation and tracking (SAPET) method to recover per-frame 6-DoF pose and joint angles from unposed endoscopic videos, where a gripper-tip network trained purely from synthetic semantics provides robust supervision and a loose regularization suppresses singular articulations. Finally, we introduce Robust Texture Learning (RTL), which alternates pose refinement and robust appearance optimization, mitigating pose noise during texture learning. The proposed framework can perform pose estimation and learn realistic texture from unposed videos. We validate our method on sequences extracted from EndoVis17/18, SAR-RARP, and an in-house dataset, showing superior photometric quality and improved geometric accuracy over state-of-the-art baselines. We further demonstrate a downstream keypoint detection task where unseen-pose data augmentation from our controllable instrument Gaussian improves performance.

Pujun Xue, Junyi Ge, Xiaotong Jiang et al.

Malocclusion is a major challenge in orthodontics, and its complex presentation and diverse clinical manifestations make accurate localization and diagnosis particularly important. Currently, one of the major shortcomings facing the field of dental image analysis is the lack of large-scale, accurately labeled datasets dedicated to malocclusion issues, which limits the development of automated diagnostics in the field of dentistry and leads to a lack of diagnostic accuracy and efficiency in clinical practice. Therefore, in this study, we propose the Oral and Maxillofacial Natural Images (OMNI) dataset, a novel and comprehensive dental image dataset aimed at advancing the study of analyzing dental images for issues of malocclusion. Specifically, the dataset contains 4166 multi-view images with 384 participants in data collection and annotated by professional dentists. In addition, we performed a comprehensive validation of the created OMNI dataset, including three CNN-based methods, two Transformer-based methods, and one GNN-based method, and conducted automated diagnostic experiments for malocclusion issues. The experimental results show that the OMNI dataset can facilitate the automated diagnosis research of malocclusion issues and provide a new benchmark for the research in this field. Our OMNI dataset and baseline code are publicly available at https://github.com/RoundFaceJ/OMNI.

Hanqi Li, Lu Chen, Da Ma et al.

Large language models have ushered in a new era of artificial intelligence research. However, their substantial training costs hinder further development and widespread adoption. In this paper, inspired by the redundancy in the parameters of large language models, we propose a novel training paradigm: Evolving Subnetwork Training (EST). EST samples subnetworks from the layers of the large language model and from commonly used modules within each layer, Multi-Head Attention (MHA) and Multi-Layer Perceptron (MLP). By gradually increasing the size of the subnetworks during the training process, EST can save the cost of training. We apply EST to train GPT2 model and TinyLlama model, resulting in 26.7\% FLOPs saving for GPT2 and 25.0\% for TinyLlama without an increase in loss on the pre-training dataset. Moreover, EST leads to performance improvements in downstream tasks, indicating that it benefits generalization. Additionally, we provide intuitive theoretical studies based on training dynamics and Dropout theory to ensure the feasibility of EST. Our code is available at https://github.com/OpenDFM/EST.

Huaiyuan Ying, Zijian Wu, Yihan Geng et al.

Large language models have demonstrated impressive capabilities across various natural language processing tasks, especially in solving mathematical problems. However, large language models are not good at math theorem proving using formal languages like Lean. A significant challenge in this area is the scarcity of training data available in these formal languages. To address this issue, we propose a novel pipeline that iteratively generates and filters synthetic data to translate natural language mathematical problems into Lean 4 statements, and vice versa. Our results indicate that the synthetic data pipeline can provide useful training data and improve the performance of LLMs in translating and understanding complex mathematical problems and proofs. Our final dataset contains about 57K formal-informal question pairs along with searched proof from the math contest forum and 21 new IMO questions. We open-source our code at https://github.com/InternLM/InternLM-Math and our data at https://huggingface.co/datasets/InternLM/Lean-Workbook.

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.

Jiawei Zhang, Zijian Wu, Zhiyang Liang et al.

Reconstructing high-fidelity, animatable 3D head avatars from effortlessly captured monocular videos is a pivotal yet formidable challenge. Although significant progress has been made in rendering performance and manipulation capabilities, notable challenges remain, including incomplete reconstruction and inefficient Gaussian representation. To address these challenges, we introduce FATE, a novel method for reconstructing an editable full-head avatar from a single monocular video. FATE integrates a sampling-based densification strategy to ensure optimal positional distribution of points, improving rendering efficiency. A neural baking technique is introduced to convert discrete Gaussian representations into continuous attribute maps, facilitating intuitive appearance editing. Furthermore, we propose a universal completion framework to recover non-frontal appearance, culminating in a 360$^\circ$-renderable 3D head avatar. FATE outperforms previous approaches in both qualitative and quantitative evaluations, achieving state-of-the-art performance. To the best of our knowledge, FATE is the first animatable and 360$^\circ$ full-head monocular reconstruction method for a 3D head avatar.

Zijian Wu, Adam Schmidt, Randy Moore et al.

Accurate and efficient surgical robotic tool pose estimation is of fundamental significance to downstream applications such as augmented reality (AR) in surgical training and learning-based autonomous manipulation. While significant advancements have been made in pose estimation for humans and animals, it is still a challenge in surgical robotics due to the scarcity of published data. The relatively large absolute error of the da Vinci end effector kinematics and arduous calibration procedure make calibrated kinematics data collection expensive. Driven by this limitation, we collected a dataset, dubbed SurgPose, providing instance-aware semantic keypoints and skeletons for visual surgical tool pose estimation and tracking. By marking keypoints using ultraviolet (UV) reactive paint, which is invisible under white light and fluorescent under UV light, we execute the same trajectory under different lighting conditions to collect raw videos and keypoint annotations, respectively. The SurgPose dataset consists of approximately 120k surgical instrument instances (80k for training and 40k for validation) of 6 categories. Each instrument instance is labeled with 7 semantic keypoints. Since the videos are collected in stereo pairs, the 2D pose can be lifted to 3D based on stereo-matching depth. In addition to releasing the dataset, we test a few baseline approaches to surgical instrument tracking to demonstrate the utility of SurgPose. More details can be found at surgpose.github.io.

Yuxin Chen, Zijian Wu, Adam Schmidt et al.

Purpose: Tissue tracking is critical for downstream tasks in robot-assisted surgery. The Sparse Efficient Neural Depth and Deformation (SENDD) model has previously demonstrated accurate and real-time sparse point tracking, but struggled with occlusion handling. This work extends SENDD to enhance occlusion detection and tracking consistency while maintaining real-time performance. Methods: We use the Segment Anything Model2 (SAM2) to detect and mask occlusions by surgical tools, and we develop and integrate into SENDD an Adaptive Multi-Flow Sparse Tracker (A-MFST) with forward-backward consistency metrics, to enhance occlusion and uncertainty estimation. A-MFST is an unsupervised variant of the Multi-Flow Dense Tracker (MFT). Results: We evaluate our approach on the STIR dataset and demonstrate a significant improvement in tracking accuracy under occlusion, reducing average tracking errors by 12 percent in Mean Endpoint Error (MEE) and showing a 6 percent improvement in the averaged accuracy over thresholds of 4, 8, 16, 32, and 64 pixels. The incorporation of forward-backward consistency further improves the selection of optimal tracking paths, reducing drift and enhancing robustness. Notably, these improvements were achieved without compromising the model's real-time capabilities. Conclusions: Using A-MFST and SAM2, we enhance SENDD's ability to track tissue in real time under instrument and tissue occlusions.

Shuojue Yang, Zijian Wu, Mingxuan Hong et al.

Real2Sim is becoming increasingly important with the rapid development of surgical artificial intelligence (AI) and autonomy. In this work, we propose a novel Real2Sim methodology, Instrument-Splatting, that leverages 3D Gaussian Splatting to provide fully controllable 3D reconstruction of surgical instruments from monocular surgical videos. To maintain both high visual fidelity and manipulability, we introduce a geometry pre-training to bind Gaussian point clouds on part mesh with accurate geometric priors and define a forward kinematics to control the Gaussians as flexible as real instruments. Afterward, to handle unposed videos, we design a novel instrument pose tracking method leveraging semantics-embedded Gaussians to robustly refine per-frame instrument poses and joint states in a render-and-compare manner, which allows our instrument Gaussian to accurately learn textures and reach photorealistic rendering. We validated our method on 2 publicly released surgical videos and 4 videos collected on ex vivo tissues and green screens. Quantitative and qualitative evaluations demonstrate the effectiveness and superiority of the proposed method.

Hongyu Liu, Xuan Wang, Yating Wang et al.

We introduce AvatarPointillist, a novel framework for generating dynamic 4D Gaussian avatars from a single portrait image. At the core of our method is a decoder-only Transformer that autoregressively generates a point cloud for 3D Gaussian Splatting. This sequential approach allows for precise, adaptive construction, dynamically adjusting point density and the total number of points based on the subject's complexity. During point generation, the AR model also jointly predicts per-point binding information, enabling realistic animation. After generation, a dedicated Gaussian decoder converts the points into complete, renderable Gaussian attributes. We demonstrate that conditioning the decoder on the latent features from the AR generator enables effective interaction between stages and markedly improves fidelity. Extensive experiments validate that AvatarPointillist produces high-quality, photorealistic, and controllable avatars. We believe this autoregressive formulation represents a new paradigm for avatar generation, and we will release our code inspire future research.

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.

Zijian Wu, Jinjie Ni, Xiangyan Liu et al.

Vision-language models (VLMs) trained via reinforcement learning with verifiable reward (RLVR) have shown notable progress in scaling test-time compute effectively. In this work, we investigate how synthesized RL data can further improve RLVR. To this end, we propose \textbf{SynthRL}-a scalable and guaranteed pipeline for automatic data scaling in reasoning-oriented RL training. SynthRL comprises three key stages: (1) selecting seed questions with appropriate distribution, (2) augmenting them into more challenging variants while preserving the original answers, and (3) a guaranteed verification stage that ensures near-perfect correctness and difficulty enhancement. Our empirical experiments demonstrate SynthRL's scalability and effectiveness. When applied to the MMK12 dataset, SynthRL synthesizes over 3.3K additional verifiable, challenging questions from approximately 8K seed samples. Models trained with our synthesized data achieve consistent gains across five out-of-domain visual math reasoning benchmarks, with a significant improvement over baseline models trained on seed data alone. Notably, detailed analysis reveals that the gains are more pronounced on the most challenging evaluation samples, highlighting SynthRL's effectiveness in eliciting deeper and more complex reasoning patterns.

Zhen Li, Duan Li, Yukai Guo et al. · tsinghua

Infographic charts are a powerful medium for communicating abstract data by combining visual elements (e.g., charts, images) with textual information. However, their visual and structural richness poses challenges for large vision-language models (LVLMs), which are typically trained on plain charts. To bridge this gap, we introduce ChartGalaxy, a million-scale dataset designed to advance the understanding and generation of infographic charts. The dataset is constructed through an inductive process that identifies 75 chart types, 440 chart variations, and 68 layout templates from real infographic charts and uses them to create synthetic ones programmatically. We showcase the utility of this dataset through: 1) improving infographic chart understanding via fine-tuning, 2) benchmarking code generation for infographic charts, and 3) enabling example-based infographic chart generation. By capturing the visual and structural complexity of real design, ChartGalaxy provides a useful resource for enhancing multimodal reasoning and generation in LVLMs.

Zijian Wu, Shuojue Yang, Yueming Jin et al.

In robot-assisted laparoscopic radical prostatectomy (RALP), the location of the instrument tip is important to register the ultrasound frame with the laparoscopic camera frame. A long-standing limitation is that the instrument tip position obtained from the da Vinci API is inaccurate and requires hand-eye calibration. Thus, directly computing the position of the tool tip in the camera frame using the vision-based method becomes an attractive solution. Besides, surgical instrument tip detection is the key component of other tasks, like surgical skill assessment and surgery automation. However, this task is challenging due to the small size of the tool tip and the articulation of the surgical instrument. Surgical instrument segmentation becomes relatively easy due to the emergence of the Segmentation Foundation Model, i.e., Segment Anything. Based on this advancement, we explore the deep learning-based surgical instrument tip detection approach that takes the part-level instrument segmentation mask as input. Comparison experiments with a hand-crafted image-processing approach demonstrate the superiority of the proposed method on simulated and real datasets.

Songyang Gao, Yuzhe Gu, Zijian Wu et al.

Large Reasoning Models (LRMs) have expanded the mathematical reasoning frontier through Chain-of-Thought (CoT) techniques and Reinforcement Learning with Verifiable Rewards (RLVR), capable of solving AIME-level problems. However, the performance of LRMs is heavily dependent on the extended reasoning context length. For solving ultra-hard problems like those in the International Mathematical Olympiad (IMO), the required reasoning complexity surpasses the space that an LRM can explore in a single round. Previous works attempt to extend the reasoning context of LRMs but remain prompt-based and built upon proprietary models, lacking systematic structures and training pipelines. Therefore, this paper introduces Intern-S1-MO, a long-horizon math agent that conducts multi-round hierarchical reasoning, composed of an LRM-based multi-agent system including reasoning, summary, and verification. By maintaining a compact memory in the form of lemmas, Intern-S1-MO can more freely explore the lemma-rich reasoning spaces in multiple reasoning stages, thereby breaking through the context constraints for IMO-level math problems. Furthermore, we propose OREAL-H, an RL framework for training the LRM using the online explored trajectories to simultaneously bootstrap the reasoning ability of LRM and elevate the overall performance of Intern-S1-MO. Experiments show that Intern-S1-MO can obtain 26 out of 35 points on the non-geometry problems of IMO2025, matching the performance of silver medalists. It also surpasses the current advanced LRMs on inference benchmarks such as HMMT2025, AIME2025, and CNMO2025. In addition, our agent officially participates in CMO2025 and achieves a score of 102/126 under the judgment of human experts, reaching the gold medal level.

Yiran Zhao, Shengyang Zhou, Zijian Wu et al.

Large Language Models (LLMs) have demonstrated impressive reasoning capabilities, but their substantial size often demands significant computational resources. To reduce resource consumption and accelerate inference, it is essential to eliminate redundant parameters without compromising performance. However, conventional pruning methods that directly remove such parameters often lead to a dramatic drop in model performance in reasoning tasks, and require extensive post-training to recover the lost capabilities. In this work, we propose a gradual compacting method that divides the compression process into multiple fine-grained iterations, applying a Prune-Tune Loop (PTL) at each stage to incrementally reduce model size while restoring performance with finetuning. This iterative approach-reminiscent of the "boiling frog" effect-enables the model to be progressively compressed without abrupt performance loss. Experimental results show that PTL can compress LLMs to nearly half their original size with only lightweight post-training, while maintaining performance comparable to the original model on reasoning tasks. Moreover, PTL is flexible and can be applied to various pruning strategies, such as neuron pruning and layer pruning, as well as different post-training methods, including continual pre-training and reinforcement learning. Additionally, experimental results confirm the effectiveness of PTL on a variety of tasks beyond mathematical reasoning, such as code generation, demonstrating its broad applicability.

Zijian Wu, Mingfeng Jiang, Zidian Lin et al.

3D Gaussian Splatting (3DGS) has recently emerged as an efficient, high-fidelity representation for real-time scene reconstruction and rendering. However, extending 3DGS to sparse-view settings remains challenging because of supervision scarcity and overfitting caused by limited viewpoint coverage. In this paper, we present CuriGS, a curriculum-guided framework for sparse-view 3D reconstruction using 3DGS. CuriGS addresses the core challenge of sparse-view synthesis by introducing student views: pseudo-views sampled around ground-truth poses (teacher). For each teacher, we generate multiple groups of student views with different perturbation levels. During training, we follow a curriculum schedule that gradually unlocks higher perturbation level, randomly sampling candidate students from the active level to assist training. Each sampled student is regularized via depth-correlation and co-regularization, and evaluated using a multi-signal metric that combines SSIM, LPIPS, and an image-quality measure. For every teacher and perturbation level, we periodically retain the best-performing students and promote those that satisfy a predefined quality threshold to the training set, resulting in a stable augmentation of sparse training views. Experimental results show that CuriGS outperforms state-of-the-art baselines in both rendering fidelity and geometric consistency across various synthetic and real sparse-view scenes. Project page: https://zijian1026.github.io/CuriGS/

Xingxing Hong, Yungong Wang, Dexin Jin et al.

Benchmarks are crucial for assessing multi-agent reinforcement learning (MARL) algorithms. While StarCraft II-related environments have driven significant advances in MARL, existing benchmarks like SMAC focus primarily on micromanagement, limiting comprehensive evaluation of high-level strategic intelligence. To address this, we introduce HLSMAC, a new cooperative MARL benchmark with 12 carefully designed StarCraft II scenarios based on classical stratagems from the Thirty-Six Stratagems. Each scenario corresponds to a specific stratagem and is designed to challenge agents with diverse strategic elements, including tactical maneuvering, timing coordination, and deception, thereby opening up avenues for evaluating high-level strategic decision-making capabilities. We also propose novel metrics across multiple dimensions beyond conventional win rate, such as ability utilization and advancement efficiency, to assess agents' overall performance within the HLSMAC environment. We integrate state-of-the-art MARL algorithms and LLM-based agents with our benchmark and conduct comprehensive experiments. The results demonstrate that HLSMAC serves as a robust testbed for advancing multi-agent strategic decision-making.

Haoyu Pan, Hongxin Lin, Zetian Feng et al.

The development of robust deep learning models for breast ultrasound (BUS) image analysis is significantly constrained by the scarcity of expert-annotated data. To address this limitation, we propose a clinically controllable generative framework for synthesizing BUS images. This framework integrates clinical descriptions with structural masks to generate tumors, enabling fine-grained control over tumor characteristics such as morphology, echogencity, and shape. Furthermore, we design a semantic-curvature mask generator, which synthesizes structurally diverse tumor masks guided by clinical priors. During inference, synthetic tumor masks serve as input to the generative framework, producing highly personalized synthetic BUS images with tumors that reflect real-world morphological diversity. Quantitative evaluations on six public BUS datasets demonstrate the significant clinical utility of our synthetic images, showing their effectiveness in enhancing downstream breast cancer diagnosis tasks. Furthermore, visual Turing tests conducted by experienced sonographers confirm the realism of the generated images, indicating the framework's potential to support broader clinical applications.

Linyi Huang, Hui Zhang, Zijian Wu et al.

Functional grasping is essential for humans to perform specific tasks, such as grasping scissors by the finger holes to cut materials or by the blade to safely hand them over. Enabling dexterous robot hands with functional grasping capabilities is crucial for their deployment to accomplish diverse real-world tasks. Recent research in dexterous grasping, however, often focuses on power grasps while overlooking task- and object-specific functional grasping poses. In this paper, we introduce FunGrasp, a system that enables functional dexterous grasping across various robot hands and performs one-shot transfer to unseen objects. Given a single RGBD image of functional human grasping, our system estimates the hand pose and transfers it to different robotic hands via a human-to-robot (H2R) grasp retargeting module. Guided by the retargeted grasping poses, a policy is trained through reinforcement learning in simulation for dynamic grasping control. To achieve robust sim-to-real transfer, we employ several techniques including privileged learning, system identification, domain randomization, and gravity compensation. In our experiments, we demonstrate that our system enables diverse functional grasping of unseen objects using single RGBD images, and can be successfully deployed across various dexterous robot hands. The significance of the components is validated through comprehensive ablation studies. Project page: https://hly-123.github.io/FunGrasp/ .