GeoPQA：弥合多模态大语言模型在几何推理中的视觉感知差距
GeoPQA: Bridging the Visual Perception Gap in MLLMs for Geometric Reasoning
AI总结：
本文针对多模态大语言模型在几何推理中的视觉感知瓶颈，提出 GeoPQA 基准和两阶段强化学习训练框架，先增强几何结构感知再培养推理能力，在 Qwen2.5-VL-3B-Instruct 上使几何推理和问题求解分别提升 9.7% 和 9.1%，并验证了感知基础对有效推理的重要性。
摘要：
近期强化学习（RL）的进展提升了大语言模型（LLMs）的推理能力，但对多模态大语言模型（MLLMs）的影响仍然有限。特别是在几何推理等视觉密集型任务中，MLLMs 频繁产生幻觉，导致推理不准确。我们将此归因于 MLLMs 中的感知瓶颈，该瓶颈限制了推理训练带来的收益。为了量化这一问题，我们设计了一个 Geo-Perception Question-Answering（GeoPQA）基准，针对基本几何概念和空间关系进行评测。在 GeoPQA 上的实验揭示了 MLLMs 在视觉感知方面的显著不足，这些不足制约了用于有效训练的 RL 奖励信号。为了解决这一瓶颈，我们提出了一个两阶段 RL 训练框架：首先增强对几何结构的视觉感知，然后培养推理能力。将该方法应用于 Qwen2.5-VL-3B-Instruct 后，与直接推理训练方法相比，我们的两阶段训练使几何推理性能提升了 9.7%，几何问题求解性能提升了 9.1%。我们的方法还泛化到其他视觉密集型领域，如图形理解，凸显了感知基础在有效 MLLM 推理中的重要性。
英文摘要：
Recent advancements in reinforcement learning (RL) have enhanced the reasoning abilities of large language models (LLMs), yet the impact on multimodal LLMs (MLLMs) is limited. Particularly in vision-intensive tasks like geometric reasoning, MLLMs hallucinate frequently, leading to inaccurate reasoning. We attribute this to the perceptual bottleneck in MLLMs, which caps the benefits of reasoning training. To quantify this, we design a Geo-Perception Question-Answering (GeoPQA) benchmark, targeting basic geometric concepts and spatial relationships. Experiments on GeoPQA reveal significant shortcomings of MLLMs in visual perception, which constrain RL reward signals for effective training. To address this bottleneck, we propose a two-stage RL training framework by first enhancing the visual perception of geometric structures, then fostering reasoning capabilities. Applied to Qwen2.5-VL-3B-Instruct, our two-stage training improves geometric reasoning by 9.7% and geometric problem solving by 9.1%, compared to the direct reasoning training approach. Our method also generalizes to other vision-intensive domains like figure understanding, highlighting the importance of perceptual grounding in effective MLLM reasoning.
机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。
