[论文解读] Visual Para-Thinker: Divide-and-Conquer Reasoning for Visual Comprehension

Haoran Xu, Hongyu Wang+6 more|arXiv (Cornell University)|Feb 10, 2026

Multimodal Machine Learning Applications被引用 0

一句话总结

引入 Visual Para-Thinker，这是多模态大语言模型（MLLM）首个并行推理框架，包含 Pa-Attention 与 Learnable Parallel Rotary Position Embedding (LPRoPE)，实现路径隔离、无偏向和可辨别的并行视觉推理。展示在计数、定位、细粒度感知和幻觉基准上的效率与性能提升。

ABSTRACT

Existing LLM test-time scaling laws emphasize the emergence of self-reflective behaviors through extended reasoning length. Nevertheless, this vertical scaling strategy often encounters plateaus in exploration as the model becomes locked into specific thinking pattern. By shifting from depth to parallelism, parallel thinking mitigates the narrowing of exploration. However, the extension of this paradigm to visual domain remains an open research question. In this paper, we first examine the role of visual partitioning in parallelized reasoning and subsequently propose two distinct strategies. Based on the above, we introduce Visual Para-Thinker, representing the inaugural parallel reasoning framework for MLLMs. To maintain path independence and promote diversity in reasoning, our approach integrates Pa-Attention alongside LPRoPE. Leveraging the vLLM framework, we have developed a native multimodal implementation that facilitates high-efficiency parallel processing. Empirical results on benchmark datasets such as V*, CountBench, RefCOCO, and HallusionBench confirm that Visual Para-Thinker successfully extends the benefits of parallel reasoning to the visual domain.

研究动机与目标

• 研究视觉分区如何影响视觉领域的并行推理。
• 提出 Visual Para-Thinker，成为首个面向 MLLMs 的并行推理框架。
• 引入 Pa-Attention 与 LPRoPE，确保路径隔离、无偏与可辨别性。
• 通过在 vLLM 上的原生多模实现与广泛基准测试，证明其效率与有效性。

提出的方法

• 分析视觉分区策略并提出 Block-based 与 Scan-order 分区。
• 开发 Visual Para-Thinker，采用两阶段结构：Parallel Reasoning 与 Summary。
• 在推理阶段和摘要阶段均引入 Pa-Attention 以实现推理路径的隔离。
• 集成 Learnable Parallel Rotary Position Embedding (LPRoPE) 以实现路径无偏与可辨别性。
• 在支持共享预填、并行解码和带 KV-cache 管理的摘要解码的 vLLM 上实现高效推理框架。

Figure 1 : Schematic representations of two distinct strategies for visual partitioning. (a) illustrates Block-based partitioning, while (b) shows Scan-order partitioning.

实验结果

研究问题

• RQ1 视觉分区如何影响多模态模型中的并行推理路径？
• RQ2 Pa-Attention 与 LPRoPE 能否在视觉任务中实现独立、可识别的并行推理路径？
• RQ3 与顺序或多数投票基线相比，视觉领域的并行推理是否提高了准确性并降低了幻觉？

主要发现

• Visual Para-Thinker 将并行思维扩展到视觉领域，在计数、定位和幻觉任务上取得提升。
• 采用混合的 Block-based 与 Scan-order 分区策略，结合 Pa-Attention 与 LPRoPE，实现跨路径的路径隔离、无偏与可辨别性。
• 实验表明，在更多推理路径（1、2、4 路径）下，对视觉相关任务有持续改进，且性能优于顺序或多数投票基线。
• 模型在定位方面表现出色，在 RefCOCO 系列上比多家基线有更高准确度，并在 MMVP 与 HallusionBench 上降低幻觉。
• 通过 KV-cache 重用和并行解码实现了效率提升，与顺序或多数投票方法相比，整体耗时竞争力更高、吞吐量更大。

Figure 2 : (a) illustrates the attention allocation results for Path 1 and Path 4 using the Block-based partitioning strategy during visual partitioning. The left panels present the attention maps for path 1 and path 4, while the right panels display the corresponding histograms of the spatial atten