PaperlayerView exact revision on arXiv

Source-linked AI summary

Monet: Reasoning in Latent Visual Space Beyond Images and Language

Qixun Wang, Yang Shi, Yifei Wang, Yuanxing Zhang, Pengfei Wan, Kun Gai, Xianghua Ying, Yisen Wang

arXiv:2511.21395v2cs.CVcs.AI

TL;DR

Existing MLLM visual reasoning methods are limited by costly latent–visual alignment, weak supervision for latent embeddings, and reinforcement learning objectives that optimize text rather than latent reasoning. Monet addresses these gaps with a three-stage SFT pipeline and VLPO, achieving consistent gains across real-world benchmarks and strong OOD generalization on abstract visual tasks. The paper also identifies training-complexity and reward-design limitations that remain open.

  • Problem

    Prior latent-reasoning methods face expensive latent–visual alignment and weak supervision that can let models memorize following tokens instead of learning effective latent representations.

  • Method

    Monet trains MLLMs to generate latent visual thoughts through three-stage SFT and uses VLPO to incorporate latent embeddings directly into reinforcement-learning policy updates.

  • Results

    Monet substantially improves real-world perception and reasoning and exhibits strong out-of-distribution generalization on abstract visual reasoning tasks.

  • Takeaways & Limitations

    The results support latent visual reasoning as a promising direction toward more flexible and general multimodal reasoning.

  • Takeaways & Limitations

    Monet relies on a multi-stage SFT pipeline that may increase training complexity and overhead, and different reward designs remain unexplored.

Abstract

from arXiv · show

"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. Our model, data, and code are available at https://github.com/NOVAglow646/Monet.

1. Introduction

Existing image-augmented reasoning methods improve MLLM visual reasoning but remain constrained by external tools. Monet instead trains MLLMs to generate latent embeddings as intermediate visual thoughts, using specialized SFT supervision and VLPO.

  • Auxiliary images in intermediate CoT steps can improve MLLM visual reasoning, but existing approaches depend on a limited set of external tools.Methods obtain auxiliary images through region-coordinate prediction, visual tools, or executable code that modifies the input image.
  • Monet trains MLLMs to reason directly in continuous latent space by generating embeddings that serve as intermediate visual thoughts.This removes the need for explicit auxiliary images and is intended to overcome the rigidity of external tools.
  • Monet-SFT uses dual supervision and controlled attention to align latent reasoning with crucial visual features while preserving visual information.The method supervises hidden representations of text tokens associated with visual features and lets latent embeddings attend directly to auxiliary image embeddings.
  • VLPO incorporates latent embeddings into reinforcement-learning optimization, unlike GRPO, which targets text reasoning.VLPO computes policy gradients for latent embeddings by estimating their output probability.
  • Monet-SFT and VLPO outperform conventional SFT + GRPO, cropping-based methods, and prior latent visual reasoning approaches, with VLPO improving OOD generalization on unseen tasks.The framework also introduces Monet-SFT-152K, a curated image-text interleaved CoT dataset for SFT.

2. Related Work

Related work spans auxiliary-image methods and continuous latent reasoning. Auxiliary-image approaches improve visual reasoning but face generalization, supervision, deployment, and latency limitations, while latent MLLM methods remain constrained by their alignment strategies.

  • Auxiliary-image methods either emphasize content from the original image or create new visual content using tools, code interpreters, or text-to-image models.Examples include grounding, cropping, re-inputting image tokens, editing images, and generating new images.
  • Tool-specific models struggle with complex visual operations, while tool-dependent reasoning often requires extensive supervision for valid calls or executable code.The cited difficult operations include visual math, spatial reasoning, and graphic reasoning.
  • External tools or interpreters require asynchronous, multi-turn inference, complicating deployment and increasing latency.
  • Continuous latent reasoning replaces or reweights text embeddings to explore a continuous rather than discrete linguistic space and shorten reasoning chains.
  • Prior latent MLLM methods align generated embeddings with auxiliary-image embeddings, but some compress image features or focus only on cropped regions.These design choices may distort visual information or fail to encode operations over the entire image.

3. Method

Monet combines three-stage supervised fine-tuning with VLPO reinforcement learning to train MLLMs to reason through continuous latent visual embeddings. Its dataset construction filters for necessary and accurate auxiliary images and highlights visual-observation tokens for supervision.

  • Inference: During inference, Monet interleaves language reasoning with a fixed-length sequence of latent embeddings, entering and exiting latent reasoning through special tokens.The model automatically decides when to begin latent reasoning and feeds decoder representations back as subsequent input embeddings.
  • Supervised fine-tuning: The three-stage SFT pipeline warms up interleaved reasoning, aligns latent embeddings with observation-token representations under controlled attention, and then distills latent generation without auxiliary-image access.Stage 2 uses a teacher with ground-truth auxiliary images and a student whose generated latent embeddings alone access those images.
  • Dataset construction: Monet-SFT-125K filters interleaved CoTs for auxiliary-image necessity, auxiliary-image correctness, and crucial visual-observation tokens.The dataset contains 125K problems spanning real-world, document, chart, and geometry tasks with varied visual operations.
  • SFT Stage 1: Warm up: Warm-up increases the accuracy gain from auxiliary images on observation-token prediction, indicating greater use of intermediate visual features.The unadapted base model receives almost no benefit from auxiliary images, whereas the gain grows during warm-up.
  • Reinforcement learning: VLPO estimates probabilities for rollout-collected continuous latent embeddings so outcome rewards can optimize latent reasoning directly, unlike GRPO.GRPO operates only on text tokens because latent embeddings lack an explicit probability distribution.

4. Experiment

Experiments evaluate Monet across real-world perception, reasoning, and abstract out-of-distribution tasks, while ablations examine its supervision, optimization, and latent-size choices. Monet consistently improves benchmark performance, VLPO specifically strengthens latent reasoning, and test-time latent scaling is most useful when trained appropriately.

  • Main Results: Monet improves Qwen2.5-VL-7B by 4.25%–9.75% across real-world, chart, and OCR benchmarks, outperforming the evaluated baselines on most tasks.Baselines include Vanilla SFT, SFT + GRPO, Deepeyes, and LVR.
  • Main Results: Monet achieves the best performance on VisualPuzzles, demonstrating strong out-of-distribution generalization to abstract visual reasoning problems unseen during training.VisualPuzzles evaluates abstract visual reasoning rather than prior knowledge.
  • Component Analysis: Removing observation-token alignment causes a substantial performance drop, showing that dual supervision is necessary for latent reasoning.The ablations retain auxiliary-image embeddings while removing one supervision signal.
  • Component Analysis: VLPO significantly improves Monet-SFT, whereas GRPO does not consistently improve it, indicating that GRPO is not well-suited for latent reasoning.The comparison attributes the difference to VLPO’s explicit treatment of latent embeddings.
  • Component Analysis: Allowing latent-related losses to update non-latent representations sharply reduces performance, supporting latent-only backpropagation.Without this restriction, the model may minimize alignment losses through shortcut paths rather than improving latent embeddings.
  • Latent-Size Analysis: For in-distribution tasks, performance is worst at Ktest = 0, while only VLPO consistently benefits from Ktest > 0 on VisualPuzzles.This indicates that SFT alone does not induce strong out-of-distribution generalization through latent reasoning.
  • Latent-Size Analysis: On V*, HRBench, and MME-RealWorld, performance typically peaks at Ktest larger than Ktrain, and Monet-SFT + VLPO is most stable as Ktest varies.Before the peak, performance generally improves as Ktest increases.
  • Latent-Size Analysis: GRPO mainly strengthens non-latent reasoning and provides limited benefits for latent reasoning because it optimizes text tokens without directly supervising latent embeddings.With Ktest = 0 it is comparable to or better than Monet-SFT, but with Ktest > 0 it underperforms the SFT model for most settings.

5. Conclusion and Limitations

Monet combines latent-space visual reasoning, a three-stage SFT pipeline, VLPO, and the Monet-SFT-125K dataset. Experiments report stronger real-world performance and abstract-task generalization, while the framework remains limited by training complexity and unexplored reward designs.

  • Conclusion: Monet enables MLLMs to perform visual reasoning in latent space by encoding useful visual cues from auxiliary images into compact latent embeddings.VLPO explicitly optimizes these embeddings during reinforcement learning.
  • Conclusion: Monet-SFT-125K supports supervised fine-tuning, and experiments show improved real-world perception, reasoning, and out-of-distribution abstract visual reasoning.The conclusion characterizes Monet as a promising direction for flexible and general multimodal reasoning.
  • Limitations: The multi-stage SFT pipeline may increase overall training complexity and overhead, and different reward designs for latent visual reasoning remain unexplored.These are the authors’ stated limitations and open directions.

A. Additional Experimental Results

Additional results on MME-RealWorld-Lite show that Monet performs strongly across diverse domains and maintains generality across visual reasoning and perception tasks.

  • MME-RealWorld-Lite: Monet achieves consistently strong performance across MME-RealWorld-Lite data sources, including monitoring, autonomous driving, and complex diagram tasks.The result supports generality across both visual reasoning and perception.

B.1. SFT Training

SFT implementation trains the model to predict latent reasoning while enforcing latent-only gradient flow for observation alignment. Training uses memory-management constraints to avoid out-of-memory failures.

  • SFT implementation: The SFT labels train prediction of <latent> while masking latent-end and observation boundary tokens.This preserves supervision for latent-start prediction while excluding boundary markers from the labels.
  • Training constraints: SFT uses batch size 1, gradient accumulation of 16 steps, and a maximum-pixel limit to prevent out-of-memory errors.The pixel constraint applies to auxiliary images in each training sequence.
  • Latent-only backpropagation: Stage 2 replaces the observation-alignment loss with a surrogate loss that restricts gradients to generated latent representations.The construction uses stop-gradient operations to prevent parameter updates through unintended paths.
  • Latent-only backpropagation: Minimizing the surrogate observation-alignment loss is equivalent to optimizing only the generated latent embeddings.The generated latent embeddings are denoted as latent, and L is the number of language-model layers.

B.2. RL Training

RL training is implemented with EasyR1 and uses an accuracy threshold to retain informative partially successful groups. The data-construction prompts identify and preserve only visually necessary observations.

  • RL implementation: RL training uses EasyR1, an open-source framework for multimodal large language models.The framework provides the implementation basis for the paper’s reinforcement-learning training.
  • RL sampling: An accuracy threshold retains groups with non-zero accuracy below the threshold, preventing advantages from vanishing when all responses are correct or incorrect.This filtering targets groups that contain mixed or otherwise informative outcomes.
  • Training-data construction: The data-construction prompt marks solution segments that rely on visual observations with <observation> tags.The input contains a question and solution chain-of-thought text without images.
  • Training-data construction: Prompt rules preserve all markers and text while placing only indispensable image-dependent tokens inside observation tags.Information inferable from the question or earlier reasoning is excluded from the observation span.

C. Detailed Experimental Setup

Evaluation uses VLMEvalKit with rule-based and secondary model judges, while examples illustrate Monet-SFT-125K data operations and Monet-7B inference across visual and nonvisual tasks. The examples include cropping, state creation, auxiliary drawing, spatial reasoning, transformations, diagrams, OCR, commonsense, and math.

  • Evaluation setup: Evaluation uses VLMEvalKit with a maximum of 8192×28×28 visual tokens at inference time.Exact-match judging is followed by DeepSeek-V3.1 or Gemini-2.5-Pro as secondary judges.
  • SFT data examples: Monet-SFT-125K examples contain image–text interleaved CoTs with intermediate auxiliary images for visual reasoning.The illustrated operations are cropping key regions, creating new visual states, and drawing lines or bounding boxes.
  • Inference examples: Monet-7B inference examples cover 3D spatial reasoning, 2D transformation, complex diagram reasoning, commonsense QA, and fine-grained OCR.The examples show latent embeddings used before final answers, including identifying relevant image regions and transformation rules.
  • Inference examples: For pure-text math reasoning, Monet-7B relies on text-based reasoning instead of activating latent thinking mode.The reported math example is a five-day weather-forecast problem involving gains of 3 coins and fines of 2 coins.
  • Inference examples: The 2D transformation example asks the model to infer a repeated reflection rule for the number 4 and apply it to the number 5.The corresponding figure reports that Monet identifies the flipping rule by generating latent embeddings.
  • Inference examples: The OCR example asks where Joseph works, while the diagram example asks which country is the top sales country.The figures report locating key information in the image and identifying the relevant chart section.
Loading 2511.21395v2…