Selected work

Representative Papers

BabyVision: Visual Reasoning Beyond Language

Jan 10, 2026arXiv.org

This work addresses the overreliance of current multimodal large language models on linguistic priors and their consequent deficiency in foundational visual understanding—capabilities that even human infants possess—leading to markedly subpar performance on basic visual tasks. To systematically evaluate pure visual reasoning independent of language, the authors introduce BabyVision, a comprehensive benchmark comprising 388 non-linguistic visual tasks across four major categories and 22 subcategories. They further present BabyVision-Gen, a generative model tailored for this benchmark, along with an automated evaluation toolkit. Experimental results reveal that leading models, such as Gemini3-Pro-Preview (scoring 49.7), fall significantly short of adult human performance (94.1), underscoring a critical gap in foundational visual primitives and highlighting the need to advance multimodal models toward more human-like visual perception.

4 citations2 influentialRead paper

Towards Pixel-Level VLM Perception via Simple Points Prediction

Jan 27, 2026

This work proposes a novel approach to endow multimodal large language models (MLLMs) with native pixel-level perception capabilities without requiring specialized architectures. By reformulating image segmentation as a sequence generation task of coordinate points in the language space, the method leverages standard MLLM architectures and employs a two-stage training strategy: supervised fine-tuning (SFT) followed by reinforcement learning (RL) optimization using Intersection over Union (IoU) as the reward signal. This approach significantly enhances segmentation accuracy and demonstrates, for the first time, that general-purpose MLLMs inherently possess strong low-level visual perception potential. Remarkably, using only simple point prediction, the model achieves competitive or superior performance against existing complex, task-specific architectures across multiple segmentation benchmarks, offering both efficiency and broad applicability.

1 citationsRead paper

Thinking Before Running! Efficient Code Generation with Thorough Exploration and Optimal Refinement

Dec 30, 2024arXiv.org

To address high test-time computational overhead and low trial-and-error efficiency—leading to inference latency in large language model (LLM)-based code generation—this paper proposes ThinkCoder, a two-stage “explore-refine” paradigm. The first stage performs deep forward reasoning via diverse solution-space exploration and optimal-solution selection. The second stage introduces preference-driven Reinforced Self-Training (ReST), leveraging successful execution trajectories to optimize the generation policy *prior* to inference. ThinkCoder enables efficient single-LLM deployment and substantially reduces test-time computation. Experiments show that on HumanEval and MBPP, it achieves a 3.0% Pass@1 improvement over MapCoder while consuming only 6.4% of its compute cost; it outperforms AgentCoder’s 5-turn results within just 2 generations; and LLaMA2-7B reaches the performance ceiling of comparable methods using only 20% of their computational budget.

1 citationsRead paper

Kimi K3: Open Frontier Intelligence

Jul 27, 2026

This work addresses performance and efficiency bottlenecks of large language models in long-context, multimodal, and complex reasoning tasks by introducing a 2.8 trillion-parameter sparsely activated Mixture-of-Experts model with native vision capabilities and support for million-token contexts. Key innovations include Kimi Delta Attention and Attention Residuals to enhance information flow, Stable LatentMoE for robust and efficient expert routing, and a co-designed algorithm-system reinforcement learning framework featuring persistent rollouts and sandboxed states. The model achieves state-of-the-art performance across long-context encoding, agent-based tasks, knowledge-intensive question answering, reasoning, and vision benchmarks, demonstrating a 2.5× improvement in scaling efficiency over its predecessor. It surpasses existing open-source models and most closed-source counterparts, with full model weights publicly released.

0 citationsRead paper

Thinking in Scales: Accelerating Gigapixel Pathology Image Analysis via Adaptive Continuous Reasoning

May 19, 2026

This work addresses the high computational cost and inefficiency of conventional whole-slide image analysis, which relies on multiple instance learning and processes a large number of high-magnification image patches. The authors propose PathCTM, a novel model that formulates pathological diagnosis as a dynamic, sequential reasoning process. Starting from a low-magnification global view, PathCTM employs an attention mechanism to guide region pruning and adaptively switches magnification scales or terminates inference early based on prediction confidence. By integrating conditional computation, dynamic scale selection, and early stopping, the method achieves both high diagnostic accuracy and remarkable efficiency. Experimental results demonstrate that PathCTM reduces patch usage by 95.95% and inference time by 95.62% compared to baseline approaches, while preserving AUC performance without degradation.

0 citationsRead paper

Recent publications

Latest Papers

Kimi K3: Open Frontier Intelligence

Jul 27, 2026

This work addresses performance and efficiency bottlenecks of large language models in long-context, multimodal, and complex reasoning tasks by introducing a 2.8 trillion-parameter sparsely activated Mixture-of-Experts model with native vision capabilities and support for million-token contexts. Key innovations include Kimi Delta Attention and Attention Residuals to enhance information flow, Stable LatentMoE for robust and efficient expert routing, and a co-designed algorithm-system reinforcement learning framework featuring persistent rollouts and sandboxed states. The model achieves state-of-the-art performance across long-context encoding, agent-based tasks, knowledge-intensive question answering, reasoning, and vision benchmarks, demonstrating a 2.5× improvement in scaling efficiency over its predecessor. It surpasses existing open-source models and most closed-source counterparts, with full model weights publicly released.

0 citationsRead paper

Thinking in Scales: Accelerating Gigapixel Pathology Image Analysis via Adaptive Continuous Reasoning

May 19, 2026

This work addresses the high computational cost and inefficiency of conventional whole-slide image analysis, which relies on multiple instance learning and processes a large number of high-magnification image patches. The authors propose PathCTM, a novel model that formulates pathological diagnosis as a dynamic, sequential reasoning process. Starting from a low-magnification global view, PathCTM employs an attention mechanism to guide region pruning and adaptively switches magnification scales or terminates inference early based on prediction confidence. By integrating conditional computation, dynamic scale selection, and early stopping, the method achieves both high diagnostic accuracy and remarkable efficiency. Experimental results demonstrate that PathCTM reduces patch usage by 95.95% and inference time by 95.62% compared to baseline approaches, while preserving AUC performance without degradation.

0 citationsRead paper

Prefill-as-a-Service: KVCache of Next-Generation Models Could Go Cross-Datacenter

Apr 16, 2026

This work addresses the limitations of the conventional Prefill-Decode architecture, whose substantial KV Cache transfer overhead hinders cross-datacenter deployment and impedes resource elasticity and independent scaling of heterogeneous hardware. To overcome this, we propose Prefill-as-a-Service (PaaS), the first architecture to decouple and serviceify the prefill phase across datacenters. Through co-design of model and system optimizations—including hybrid attention mechanisms, KV Cache compression, selective offloading, bandwidth-aware scheduling, and cache-aware request placement—we drastically reduce inter-cluster communication costs. Evaluated on a 1-trillion-parameter mixture-of-experts model, PaaS achieves 54% and 32% higher service throughput compared to homogeneous and naive heterogeneous baselines, respectively, while requiring only modest cross-datacenter bandwidth.

0 citationsRead paper

TENT: A Declarative Slice Spraying Engine for Performant and Resilient Data Movement in Disaggregated LLM Serving

Mar 31, 2026

Existing data transmission frameworks in heterogeneous interconnect environments rely on static path selection, often leading to bandwidth underutilization, communication islands, and poor fault tolerance. This work proposes TENT, a novel transmission engine that decouples intent from physical execution through a declarative slice-spraying mechanism. TENT dynamically partitions large flows into fine-grained segments and schedules them across a unified pool of heterogeneous interconnect resources—including NVLink and Ascend UB—based on real-time link quality. This approach enables head-of-line blocking–free transmission and sub-50-millisecond transparent self-healing without application-level intervention. Experimental results on an H800 cluster demonstrate a 1.36× improvement in LLM inference throughput, a 26% reduction in P90 time-to-first-token latency, and a 20–26% acceleration in reinforcement learning parameter updates.

0 citationsRead paper

GlobeDiff: State Diffusion Process for Partial Observability in Multi-Agent Systems

Feb 17, 2026

This work addresses the challenge of cooperative decision-making in partially observable multi-agent systems, where agents’ local observations often hinder accurate inference of the global state. To overcome this limitation, we propose GlobeDiff, the first approach to leverage diffusion models for global state reconstruction in such settings. GlobeDiff formulates global state estimation as a multimodal diffusion process that generates high-fidelity approximations of the global state using only local observations, bypassing the constraints of conventional belief propagation and explicit communication protocols. The method is accompanied by theoretical guarantees on estimation error bounds. Empirical evaluations across diverse environments demonstrate that GlobeDiff significantly outperforms existing approaches, achieving both high accuracy and robustness.

0 citationsRead paper