Explore · 每周精选
发现最受关注的研究论文,追踪研究趋势,订阅感兴趣的期刊与关键词。
Jingqi Tong, Yurong Mou, Hangcheng Li +11 authors
The "Thinking with Video" paradigm enhances multimodal reasoning by integrating video generation models, demonstrated through the Video Thinking Benchmark and improved performance on both vision and text tasks.
50 篇论文 · 按点赞排序
Vladimir Arkhipkin, Vladimir Korviakov, Nikolai Gerasimenko +22 authors
Kandinsky 5.0 is a family of state-of-the-art generative models for high-resolution images and short videos, featuring model lineups with varying parameters and enhanced training techniques to achieve superior quality and performance.
Weihao Tan, Xiangyang Li, Yunhao Fang +11 authors
Lumine, a vision-language model-based agent, completes complex missions in real-time across different 3D open-world environments with human-like efficiency and zero-shot cross-game generalization.
MiroMind Team, Song Bai, Lidong Bing +51 authors
We present MiroThinker v1.0, an open-source research agent designed to advance tool-augmented reasoning and information-seeking capabilities. Unlike previous agents that only scale up model size or context length, MiroThinker explores interaction scaling at the model level, systematically training the model to handle deeper and more frequent agent-environment interactions as a third dimension of performance improvement. Unlike LLM test-time scaling, which operates in isolation and risks degradation with longer reasoning chains, interactive scaling leverages environment feedback and external information acquisition to correct errors and refine trajectories. Through reinforcement learning, the model achieves efficient interaction scaling: with a 256K context window, it can perform up to 600 tool calls per task, enabling sustained multi-turn reasoning and complex real-world research workflows. Across four representative benchmarks-GAIA, HLE, BrowseComp, and BrowseComp-ZH-the 72B variant achieves up to 81.9%, 37.7%, 47.1%, and 55.6% accuracy respectively, surpassing previous open-source agents and approaching commercial counterparts such as GPT-5-high. Our analysis reveals that MiroThinker benefits from interactive scaling consistently: research performance improves predictably as the model engages in deeper and more frequent agent-environment interactions, demonstrating that interaction depth exhibits scaling behaviors analogous to model size and context length. These findings establish interaction scaling as a third critical dimension for building next-generation open research agents, complementing model capacity and context windows.
B. Y. Yan, Chaofan Li, Hongjin Qian +2 authors
GAM, a novel framework that employs JIT compilation principles, improves memory efficiency and task completion by leveraging a lightweight memorizer and researcher in conjunction with reinforcement learning.
Sen Xu, Yi Zhou, Wei Wang +7 authors
VibeThinker-1.5B, a 1.5B-parameter model using the Spectrum-to-Signal Principle, achieves superior reasoning capabilities compared to larger models at a significantly lower cost.
Shalini Maiti, Amar Budhiraja, Bhavul Gauri +8 authors
Large Language Models (LLMs) have demonstrated remarkable capabilities across diverse domains, but their training remains resource- and time-intensive, requiring massive compute power and careful orchestration of training procedures. Model souping-the practice of averaging weights from multiple models of the same architecture-has emerged as a promising pre- and post-training technique that can enhance performance without expensive retraining. In this paper, we introduce Soup Of Category Experts (SoCE), a principled approach for model souping that utilizes benchmark composition to identify optimal model candidates and applies non-uniform weighted averaging to maximize performance. Contrary to previous uniform-averaging approaches, our method leverages the observation that benchmark categories often exhibit low inter-correlations in model performance. SoCE identifies "expert" models for each weakly-correlated category cluster and combines them using optimized weighted averaging rather than uniform weights. We demonstrate that the proposed method improves performance and robustness across multiple domains, including multilingual capabilities, tool calling, and math and achieves state-of-the-art results on the Berkeley Function Calling Leaderboard.
Nicolas Carion, Laura Gustafson, Yuan-Ting Hu +35 authors
Segment Anything Model 3 achieves state-of-the-art performance in promptable concept segmentation and tracking by leveraging a unified model architecture with decoupled recognition and localization.
Jiacheng Chen, Qianjia Cheng, Fangchen Yu +25 authors
Recent progress in large language models (LLMs) has moved the frontier from puzzle-solving to science-grade reasoning-the kind needed to tackle problems whose answers must stand against nature, not merely fit a rubric. Physics is the sharpest test of this shift, which binds symbols to reality in a fundamental way, serving as the cornerstone of most modern technologies. In this work, we manage to advance physics research by developing large language models with exceptional physics reasoning capabilities, especially excel at solving Olympiad-level physics problems. We introduce P1, a family of open-source physics reasoning models trained entirely through reinforcement learning (RL). Among them, P1-235B-A22B is the first open-source model with Gold-medal performance at the latest International Physics Olympiad (IPhO 2025), and wins 12 gold medals out of 13 international/regional physics competitions in 2024/2025. P1-30B-A3B also surpasses almost all other open-source models on IPhO 2025, getting a silver medal. Further equipped with an agentic framework PhysicsMinions, P1-235B-A22B+PhysicsMinions achieves overall No.1 on IPhO 2025, and obtains the highest average score over the 13 physics competitions. Besides physics, P1 models also present great performance on other reasoning tasks like math and coding, showing the great generalibility of P1 series.
Aleksandr Razin, Danil Kazantsev, Ilya Makarov
LUA is a lightweight module that performs super-resolution directly in the latent space of diffusion models, improving efficiency without compromising image quality.
Jinjie Ni, Qian Liu, Longxu Dou +5 authors
Diffusion language models outperform autoregressive models in low-data settings due to any-order modeling, iterative bidirectional denoising, and Monte Carlo augmentation, and maintain advantages even at scale.
Jiaru Zou, Xiyuan Yang, Ruizhong Qiu +10 authors
LatentMAS enables efficient, lossless collaboration among LLM agents in latent space, improving performance and reducing computational costs compared to text-based methods.
Jingyu Liu, Xin Dong, Zhifan Ye +6 authors
TiDAR combines diffusion and autoregressive models to achieve high throughput and quality in language generation.
Valentin Khrulkov, Andrey Galichin, Denis Bashkirov +5 authors
GigaEvo is an open-source framework for LLM-guided evolutionary computation, offering modular and concurrent tools for research and experimentation in solving complex optimization problems.
SAM 3D Team, Xingyu Chen, Fu-Jen Chu +20 authors
SAM 3D is a generative model that reconstructs 3D objects from single images using a multi-stage training framework that includes synthetic pretraining and real-world alignment, achieving high performance in human preference tests.
Yifan Jiang, Yueying Wang, Rui Zhao +4 authors
VideoP2R, a process-aware reinforcement fine-tuning framework, improves video reasoning and understanding by modeling perception and reasoning separately, achieving state-of-the-art results on multiple benchmarks.
Peng Xia, Kaide Zeng, Jiaqi Liu +5 authors
Agent0, a self-evolving framework utilizing multi-step co-evolution and tool integration, enhances LLM reasoning capabilities without human-curated data.
Tianyu Fu, Yichen You, Zekai Chen +3 authors
Think-at-Hard (TaH) dynamically refines only hard tokens in LLMs using a neural decider and LoRA, improving reasoning performance with minimal additional parameters or iterations.
Yunxin Li, Xinyu Chen, Shenyuan Jiang +9 authors
We present Uni-MoE 2.0 from the Lychee family. As a fully open-source omnimodal large model (OLM), it substantially advances Lychee's Uni-MoE series in language-centric multimodal understanding, reasoning, and generating. Based on the Qwen2.5-7B dense architecture, we build Uni-MoE-2.0-Omni from scratch through three core contributions: dynamic-capacity Mixture-of-Experts (MoE) design, a progressive training strategy enhanced with an iterative reinforcement strategy, and a carefully curated multimodal data matching technique. It is capable of omnimodal understanding, as well as generating images, text, and speech. Architecturally, our new MoE framework balances computational efficiency and capability for 10 cross-modal inputs using shared, routed, and null experts, while our Omni-Modality 3D RoPE ensures spatio-temporal cross-modality alignment in the self-attention layer. For training, following cross-modal pretraining, we use a progressive supervised fine-tuning strategy that activates modality-specific experts and is enhanced by balanced data composition and an iterative GSPO-DPO method to stabilise RL training and improve reasoning. Data-wise, the base model, trained on approximately 75B tokens of open-source multimodal data, is equipped with special speech and image generation tokens, allowing it to learn these generative tasks by conditioning its outputs on linguistic cues. Extensive evaluation across 85 benchmarks demonstrates that our model achieves SOTA or highly competitive performance against leading OLMs, surpassing Qwen2.5-Omni (trained with 1.2T tokens) on over 50 of 76 benchmarks. Key strengths include video understanding (+7% avg. of 8), omnimodallity understanding (+7% avg. of 4), and audiovisual reasoning (+4%). It also advances long-form speech processing (reducing WER by 4.2%) and leads in low-level image processing and controllable generation across 5 metrics.
Aarash Feizi, Shravan Nayak, Xiangru Jian +14 authors
GroundCUA, a large-scale desktop grounding dataset, enables the development of GroundNext models that achieve state-of-the-art performance in mapping instructions to UI elements with less training data.
Nikita Kachaev, Mikhail Kolosov, Daniil Zelezetsky +2 authors
Systematic study reveals that naive action fine-tuning degrades visual representations in Vision-Language-Action models, but targeted strategies can mitigate this and improve generalization.
Kevin Qinghong Lin, Yuhao Zheng, Hangyu Ran +5 authors
VCode introduces a benchmark for generating SVG code from images to preserve symbolic meaning, highlighting gaps in visual-centric coding and proposing VCoder to improve performance.
Haotong Lin, Sili Chen, Junhao Liew +5 authors
Depth Anything 3 (DA3) uses a plain transformer for geometry prediction from visual inputs, achieving state-of-the-art results in camera pose estimation, any-view geometry, visual rendering, and monocular depth estimation.
Yikun Wang, Zuyan Liu, Ziyi Wang +3 authors
GeoVista, an agentic model integrating tool invocation and reinforcement learning, achieves high geolocalization performance on GeoBench, outperforming open-source models and matching closed-source models.
Jing Xiong, Liyang Fan, Hui Shen +4 authors
Denoising Positional Encoding (DoPE) enhances length generalization in Transformer models by detecting and mitigating noisy frequency bands in positional embeddings, improving retrieval accuracy and reasoning stability.
Runqi Qiao, Qiuna Tan, Minghan Yang +10 authors
V-Thinker, a multimodal reasoning assistant using reinforcement learning, enhances image-interactive thinking by synthesizing datasets and aligning perception for improved performance in vision-centric tasks.
Kaichen Zhang, Keming Wu, Zuhao Yang +5 authors
OpenMMReasoner, a two-stage training approach combining supervised fine-tuning and reinforcement learning, enhances multimodal reasoning performance through rigorous data curation and improved training strategies.
Ding Chen, Simin Niu, Kehang Li +6 authors
HaluMem, a benchmark for evaluating memory hallucinations in AI systems, identifies and analyzes hallucinations across memory extraction, updating, and question answering stages using large-scale human-AI interaction datasets.
Jiayi Zhang, Yiran Peng, Fanqi Kong +12 authors
AutoEnv and AutoEnv-36 provide a standardized framework and dataset for evaluating cross-environment learning in agents, highlighting the challenges and limitations of existing learning methods.
Vladislav Pedashenko, Laida Kushnareva, Yana Khassan Nibal +5 authors
The study explores intrinsic dimension in large language models through cross-encoder analysis, linguistic features, and sparse autoencoders, revealing its independence from entropy, genre-specific stratification, and causal features related to text type.
北京市昌平区探索星信息技术及软件开发工作室
京ICP备2026059466号