Yucheng Wu

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.

Dong Zhuo, Wenzhao Zheng, Jiahe Guo et al.

Perceiving and reconstructing 4D spatial-temporal geometry from videos is a fundamental yet challenging computer vision task. To facilitate interactive and real-time applications, we propose a streaming 4D visual geometry transformer that shares a similar philosophy with autoregressive large language models. We explore a simple and efficient design and employ a causal transformer architecture to process the input sequence in an online manner. We use temporal causal attention and cache the historical keys and values as implicit memory to enable efficient streaming long-term 4D reconstruction. This design can handle real-time 4D reconstruction by incrementally integrating historical information while maintaining high-quality spatial consistency. For efficient training, we propose to distill knowledge from the dense bidirectional visual geometry grounded transformer (VGGT) to our causal model. For inference, our model supports the migration of optimized efficient attention operator (e.g., FlashAttention) from the field of large language models. Extensive experiments on various 4D geometry perception benchmarks demonstrate that our model increases the inference speed in online scenarios while maintaining competitive performance, paving the way for scalable and interactive 4D vision systems. Code is available at: https://github.com/wzzheng/StreamVGGT.

Lingzhe Zhang, Tong Jia, Mengxi Jia et al. · tsinghua

As large language models (LLMs) grow increasingly sophisticated and pervasive, their application to various Artificial Intelligence for IT Operations (AIOps) tasks has garnered significant attention. However, a comprehensive understanding of the impact, potential, and limitations of LLMs in AIOps remains in its infancy. To address this gap, we conducted a detailed survey of LLM4AIOps, focusing on how LLMs can optimize processes and improve outcomes in this domain. We analyzed 183 research papers published between January 2020 and December 2024 to answer four key research questions (RQs). In RQ1, we examine the diverse failure data sources utilized, including advanced LLM-based processing techniques for legacy data and the incorporation of new data sources enabled by LLMs. RQ2 explores the evolution of AIOps tasks, highlighting the emergence of novel tasks and the publication trends across these tasks. RQ3 investigates the various LLM-based methods applied to address AIOps challenges. Finally, RQ4 reviews evaluation methodologies tailored to assess LLM-integrated AIOps approaches. Based on our findings, we discuss the state-of-the-art advancements and trends, identify gaps in existing research, and propose promising directions for future exploration.

Alice Rueda, Mohammed S. Hassan, Argyrios Perivolaris et al. · utoronto

Large language models (LLMs) have demonstrated remarkable capabilities in natural language understanding, reasoning, and problem-solving across various domains. However, their ability to perform complex, multi-step reasoning task-essential for applications in science, medicine, and law-remains an area of active investigation. This paper examines the reasoning capabilities of contemporary LLMs, analyzing their strengths, limitations, and potential for improvement. The study uses prompt engineering techniques on the Graduate-Level GoogleProof Q&A (GPQA) dataset to assess the scientific reasoning of GPT-4o. Five popular prompt engineering techniques and two tailored promptings were tested: baseline direct answer (zero-shot), chain-of-thought (CoT), zero-shot CoT, self-ask, self-consistency, decomposition, and multipath promptings. Our findings indicate that while LLMs exhibit emergent reasoning abilities, they often rely on pattern recognition rather than true logical inference, leading to inconsistencies in complex problem-solving. The results indicated that self-consistency outperformed the other prompt engineering technique with an accuracy of 52.99%, followed by direct answer (52.23%). Zero-shot CoT (50%) outperformed multipath (48.44%), decomposition (47.77%), self-ask (46.88%), and CoT (43.75%). Self-consistency performed the second worst in explaining the answers. Simple techniques such as direct answer, CoT, and zero-shot CoT have the best scientific reasoning. We propose a research agenda aimed at bridging these gaps by integrating structured reasoning frameworks, hybrid AI approaches, and human-in-the-loop methodologies. By critically evaluating the reasoning mechanisms of LLMs, this paper contributes to the ongoing discourse on the future of artificial general intelligence and the development of more robust, trustworthy AI systems.

Yuqi Wu, Guangya Wan, Jingjing Li et al.

Translating state-of-the-art NLP into practice often stalls at the "last mile" owing to insufficient contextualization of the target domain's knowledge, processes, and evaluation. Psychiatric differential diagnosis exemplifies this challenge: accurate assessments depend on nuanced clinical knowledge, a delicate cognitive-affective interview process, and downstream outcomes that extend far beyond benchmark accuracy. We present WiseMind, a systematic interdisciplinary contextualization framework that delivers both instrumental (diagnostic precision) and humanistic (empathy) gains. WiseMind comprises three components:(i) structured knowledge-guided proactive reasoning, which embeds DSM-5 criteria in a knowledge graph to steer questioning; (ii) a theory-informed dual-agent architecture that coordinates a "reasonable-mind" reasoning agent and an "emotional-mind" empathy agent, inspired by Dialectical Behavior Therapy; and (iii) a multi-faceted evaluation strategy covering simulated patients, user studies, clinician review, and ethical assessment. Tested on depression, anxiety, and bipolar disorder, WiseMind attains up to 84.2% diagnostic accuracy, which is comparable to human experts, while outperforming single-agent baselines in perceived empathy and trustworthiness. These results show that deep contextualization-across knowledge, process, and evaluation layers-can transform benchmark-driven NLP into clinically meaningful impact.

Yuxuan Lu, Yifan Wu, Jason Hartline et al.

Calibrated predictions can be reliably interpreted as probabilities. An important step towards achieving better calibration is to design an appropriate calibration measure to meaningfully assess the miscalibration level of a predictor. A recent line of work initiated by Haghtalab et al. [2024] studies the design of truthful calibration measures: a truthful measure is minimized when a predictor outputs the true probabilities, whereas a non-truthful measure incentivizes the predictor to lie so as to appear more calibrated. All previous calibration measures were non-truthful until Hartline et al. [2025] introduced the first perfectly truthful calibration measures for binary prediction tasks in the batch setting. We introduce a perfectly truthful calibration measure for multi-class prediction tasks, generalizing the work of Hartline et al. [2025] beyond binary prediction. We study common methods of extending calibration measures from binary to multi-class prediction and identify ones that do or do not preserve truthfulness. In addition to truthfulness, we mathematically prove and empirically verify that our calibration measure exhibits superior robustness: it robustly preserves the ordering between dominant and dominated predictors, regardless of the choice of hyperparameters (bin sizes). This result addresses the non-robustness issue of binned ECE, which has been observed repeatedly in prior work.

Jason Hartline, Lunjia Hu, Yifan Wu

Calibration requires that predictions are conditionally unbiased and, therefore, reliably interpretable as probabilities. A calibration measure quantifies how far a predictor is from perfect calibration. As introduced by Haghtalab et al. (2024), a calibration measure is truthful if it is minimized in expectation when a predictor outputs the ground-truth probabilities. Predicting the true probabilities guarantees perfect calibration, but in reality, when calibration is evaluated on a random sample, all known calibration measures incentivize predictors to lie in order to appear more calibrated. Such lack of truthfulness motivated Haghtalab et al. (2024) and Qiao and Zhao (2025) to construct approximately truthful calibration measures in the sequential prediction setting, but no perfectly truthful calibration measure was known to exist even in the more basic batch setting. We design a simple, perfectly and strictly truthful, sound and complete calibration measure in the batch setting: averaged two-bin calibration error (ATB). ATB is quadratically related to two existing calibration measures: the smooth calibration error smCal and the lower distance to calibration distCal. The simplicity in our definition of ATB makes it efficient and straightforward to compute, allowing us to give the first linear-time calibration testing algorithm, improving a result of Hu et al. (2024). We also introduce a general recipe for constructing truthful measures based on the variance additivity of independent random variables, which proves the truthfulness of ATB as a special case and allows us to construct other truthful calibration measures such as quantile-binned l_2-ECE.

Yuanyuan Wei, Yucheng Wu, Fuyang Qu et al.

Accurate molecular quantification is essential for advancing research and diagnostics in fields such as infectious diseases, cancer biology, and genetic disorders. Droplet digital PCR (ddPCR) has emerged as a gold standard for achieving absolute quantification. While computational ddPCR technologies have advanced significantly, achieving automatic interpretation and consistent adaptability across diverse operational environments remains a challenge. To address these limitations, we introduce the intelligent interpretable droplet digital PCR (I2ddPCR) assay, a comprehensive framework integrating front-end predictive models (for droplet segmentation and classification) with GPT-4o multimodal large language model (MLLM, for context-aware explanations and recommendations) to automate and enhance ddPCR image analysis. This approach surpasses the state-of-the-art models, affording 99.05% accuracy in processing complex ddPCR images containing over 300 droplets per image with varying signal-to-noise ratios (SNRs). By combining specialized neural networks and large language models, the I2ddPCR assay offers a robust and adaptable solution for absolute molecular quantification, achieving a sensitivity capable of detecting low-abundance targets as low as 90.32 copies/μL. Furthermore, it improves model's transparency through detailed explanation and troubleshooting guidance, empowering users to make informed decisions. This innovative framework has the potential to benefit molecular diagnostics, disease research, and clinical applications, especially in resource-constrained settings.

Yucheng Wu, Liyue Chen, Yu Cheng et al.

Learning representations of user behavior sequences is crucial for various online services, such as online fraudulent transaction detection mechanisms. Graph Neural Networks (GNNs) have been extensively applied to model sequence relationships, and extract information from similar sequences. While user behavior sequence data volume is usually huge for online applications, directly applying GNN models may lead to substantial computational overhead during both the training and inference stages and make it challenging to meet real-time requirements for online services. In this paper, we leverage graph compression techniques to alleviate the efficiency issue. Specifically, we propose a novel unified framework called ECSeq, to introduce graph compression techniques into relation modeling for user sequence representation learning. The key module of ECSeq is sequence relation modeling, which explores relationships among sequences to enhance sequence representation learning, and employs graph compression algorithms to achieve high efficiency and scalability. ECSeq also exhibits plug-and-play characteristics, seamlessly augmenting pre-trained sequence representation models without modifications. Empirical experiments on both sequence classification and regression tasks demonstrate the effectiveness of ECSeq. Specifically, with an additional training time of tens of seconds in total on 100,000+ sequences and inference time preserved within $10^{-4}$ seconds/sample, ECSeq improves the prediction R@P$_{0.9}$ of the widely used LSTM by $\sim 5\%$.