Monet：超越图像与语言的潜视觉空间推理
Monet: Reasoning in Latent Visual Space Beyond Image and Language
把中间推理步骤从文本 token 换成连续视觉嵌入，让 MLLM 在潜视觉空间里做多步推理，并给出 VLPO 把潜嵌入纳入策略梯度更新——做多模态推理、CoT 或 RL 微调的人可以直接复用这套三阶段蒸馏 SFT + 潜变量 RL 的训练配方。
原文依据 we propose VLPO (Visual-latent Policy Optimization), a reinforcement learning method that explicitly incorporates latent embeddings into policy gradient updates
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论以图像思考已成为推进视觉推理的有效范式，它通过把视觉证据注入中间推理步骤，超越了纯文本的思维链。然而，现有方法的灵活性从根本上受限于外部工具，难以达到类人的抽象视觉思维。在本工作中，我们提出 Monet，一个训练框架，使多模态大语言模型（MLLM）能够通过生成作为中间视觉思考的连续嵌入，直接在潜视觉空间中进行推理。我们识别出训练 MLLM 进行潜视觉推理时的两个核心挑战——潜视觉对齐的高计算成本以及对潜嵌入的监督不足，并用一个三阶段的基于蒸馏的监督微调（SFT）流程加以解决。我们进一步揭示了将 GRPO 应用于潜推理的局限：它主要增强的是基于文本的推理，而非潜推理。为克服这一点，我们提出 VLPO（视觉潜变量策略优化），一种显式地将潜嵌入纳入策略梯度更新的强化学习方法。为支撑 SFT，我们构建了 Monet-SFT-125K，一个高质量的文本-图像交错思维链数据集，包含 125K 条真实世界、图表、OCR 与几何思维链。我们的模型 Monet-7B 在真实世界感知与推理基准上取得一致提升，并在具有挑战性的抽象视觉推理任务上展现出较强的分布外泛化能力。我们还实证分析了各训练组件的作用，并讨论了早期未成功的尝试，为视觉潜推理的未来发展提供洞见。
英文原文摘要 1715 字符
Thinking with images has emerged as an effective paradigm for advancing visual reasoning, extending beyond text-only chains of thought by injecting visual evidence into intermediate reasoning steps. However, existing methods fall short of human-like abstract visual thinking, as their flexibility is fundamentally limited by external tools. In this work, we introduce Monet, a training framework that enables multimodal large language models (MLLMs) to reason directly within the latent visual space by generating continuous embeddings that function as intermediate visual thoughts. We identify two core challenges in training MLLMs for latent visual reasoning--high computational cost in latent-vision alignment and insufficient supervision over latent embeddings, and address them with a three-stage distillation-based supervised fine-tuning (SFT) pipeline. We further reveal a limitation of applying GRPO to latent reasoning: it primarily enhances text-based reasoning rather than latent reasoning. To overcome this, we propose VLPO (Visual-latent Policy Optimization), a reinforcement learning method that explicitly incorporates latent embeddings into policy gradient updates.To support SFT, we construct Monet-SFT-125K, a high-quality text-image interleaved CoT dataset containing 125K real-world, chart, OCR, and geometry CoTs. Our model, Monet-7B, shows consistent gains across real-world perception and reasoning benchmarks and exhibits strong out-of-distribution generalization on challenging abstract visual reasoning tasks. We also empirically analyze the role of each training component and discuss our early unsuccessful attempts, providing insights for future developments in visual latent reasoning.