文章

23 篇

研究、思考与实践

前沿之见

22 篇

实验室的研究、实验与笔记。

2026.11.03

From Pixels to Actions – Towards Native Vision-Language-Action Policies

NEO-VLA feeds pixels straight into a native vision-language backbone. Control draws on its shallow layers, and native pretraining drives the gains.

2026.09.27

Noteresearchreasoning

Symbols or Pixels?

Zero-shot language reasoning versus video generation on visual reasoning tasks.

2026.04.20

modelsmultimodalvideo

LLaVA-OneVision-2: Towards Next-Generation Perceptual Intelligence

The next generation of fully-open multimodal training — pushing the boundary of recipe transparency, native-resolution understanding, and end-to-end reproducibility.

2026.01.15

modelsmultimodal

OneVision Encoder: Codec-Aligned Sparsity as a Foundational Principle for Multimodal Intelligence

Our hypothesis: AGI is a compression problem. We introduce Codec Patchification that processes only 3.1%-25% of regions, achieving 4.1% improvement on video tasks while outperforming Qwen3-ViT and SigLIP2.

2025.12.15

modelsmultimodal

LLaVA-OneVision-1.5-RL: Unlocking Multimodal Reasoning via Lightweight Reinforcement Learning

Applying reinforcement learning post-training to enhance reasoning capabilities in multimodal models with significant improvements on STEM, coding, and reasoning tasks.

2025.11.27

modelsmultimodal

LongVT: Incentivizing "Thinking with Long Videos" via Native Tool Calling

LongVT introduces a novel paradigm that natively interleaves multimodal tool-augmented Chain-of-Thought with on-demand clip inspection over hours-long videos, enabling large multimodal models to perform more effective and reliable long-video reasoning.

2025.11.21

modelsmultimodal

OpenMMReasoner: Pushing the Frontiers for Multimodal Reasoning with an Open and General Recipe

OpenMMReasoner introduces a systematic study on constructing high-quality SFT and RL datasets for multimodal reasoning, demonstrating that both source diversity and answer diversity are crucial for building reliable supervision signals.

2025.09.30

modelsmultimodal

LLaVA-OneVision-1.5: Fully Open Framework for Democratized Multimodal Training

LLaVA-OneVision1.5 introduces a novel family of fully open-source Large Multimodal Models (LMMs) that achieves state-of-the-art performance with substantially lower cost through training on native resolution images.

2025.08.29

modelsmultimodal

LLaVA-Critic-R1: Unified Critic and Policy Model Through Reinforcement Learning

A family of generative critic VLM trained through GRPO using pairwise critic data, achieving SoTA policy performance at 7B scale while excelling at both evaluation and generation

2025.08.06

researchmultimodal

Improved MM-Search-R1: Reasoning and Action in Multimodal Search

We improve MMSearch-R1 by integrating improved reasoning capabilities into the model