arXiv cs.AI· Quanlin Chen, Xiuwei Chen, Kun Xiang, Zehua Ma, Chunjin Ye, Boqiang Guo, Xianming Liu, Xiaodan Liang·· 3 天前AI 评分22
L2T：多模态推理中自适应决定“看什么”与“想多久”
Adaptive Look, Flexible Thought: Learning How and How Long to Think in Multimodal Reasoning
AI 导读
针对现有多模态大模型快慢推理只调控文本推理长度、忽视视觉计算的问题，研究者提出统一推理框架 L2T，将工具产生的视觉状态离散化为视觉 token，使视觉感知与文本推理进入同一自回归空间。L2T 引入模态感知路由奖励选择推理模式，并用难度感知快慢策略按任务复杂度调整推理深度。在通用感知与推理密集任务上，L2T 相比显式工具调用方法保持有竞争力性能，吞吐量提升 1.7–3.6 倍。
正文 · 原文
Abstract:The introduction of long chain-of-thought (CoT) reasoning and external tools has substantially improved the performance of multimodal large language models (MLLMs) on general and reasoning tasks. However, existing fast-slow reasoning paradigms mainly regulate textual reasoning length, while visual computation introduced by external tools or visual tokens is often treated as an auxiliary component. This separation prevents MLLMs from jointly optimizing visual perception and textual reasoning. To address this issue, we propose L2T (adaptive Look, flexible Thought), a unified reasoning framework that jointly optimizes the form and depth of computation. L2T first discretizes tool-produced visual states into visual tokens, bringing visual perception and textual reasoning into a shared autoregressive space. Based on this unified representation, L2T introduces a modality-aware routing reward to select appropriate reasoning modes and a difficulty-aware fast-slow strategy to adapt reasoning depth according to task complexity. Experiments on general perception and reasoning-intensive tasks demonstrate the effectiveness of L2T. Compared with explicit tool-calling approaches, L2T maintains competitive performance while achieving 1.7--3.6 times higher throughput.
| Subjects: | Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI); Computation and Language (cs.CL) |
| Cite as: | arXiv:2610.15810 [cs.CV] |
| (or arXiv:2610.15810v1 [cs.CV] for this version) | |
| https://doi.org/10.48550/arXiv.2610.15810 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Quanlin Chen [view email]
[v1]
Sun, 11 Oct 2026 17:57:17 UTC (1,979 KB)
来源：arXiv cs.AI · arxiv.org