PaperlayerView exact revision on arXiv

Source-linked AI summary

Reasoning Within the Mind: Dynamic Multimodal Interleaving in Latent Space

Chengzhi Liu, Yuzhe Yang, Yue Fan, Qingyue Wei, Sheng Liu, Xin Eric Wang

arXiv:2512.12623v3cs.CVcs.CL

TL;DR

MLLM reasoning methods can rely on explicit step-by-step generation or external visual tools, creating grounding, interaction, and efficiency limitations. DMLR performs confidence-guided latent-token optimization with dynamic visual injection at test time, and it improves reasoning and perception across diverse tasks and architectures while remaining efficient.

  • Problem

    Existing multimodal reasoning approaches can suffer from insufficient visual grounding, unstable perception-reasoning interaction, and computational overhead.

  • Method

    DMLR refines latent think tokens through confidence-guided optimization and dynamically injects selected visual information during test-time reasoning.

  • Results

    DMLR achieves the best performance on over 95% of tasks and consistently improves reasoning and perception across diverse architectures and multimodal benchmarks.

  • Takeaways & Limitations

    Dynamic visual use guided by internal confidence provides a training-free alternative that improves both reasoning and perception while maintaining high efficiency.

Abstract

from arXiv · show

Recent advancements in Multimodal Large Language Models (MLLMs) have significantly enhanced cross-modal understanding and reasoning by incorporating Chain-of-Thought (CoT) reasoning in the semantic space. Building upon this, recent studies extend the CoT mechanism to the visual modality, enabling models to integrate visual information during reasoning through external tools or explicit image generation. However, these methods remain dependent on explicit step-by-step reasoning, unstable perception-reasoning interaction and notable computational overhead. Inspired by human cognition, we posit that thinking unfolds not linearly but through the dynamic interleaving of reasoning and perception within the mind. Motivated by this perspective, we propose DMLR, a test-time Dynamic Multimodal Latent Reasoning framework that employs confidence-guided latent policy gradient optimization to refine latent think tokens for in-depth reasoning. Furthermore, a Dynamic Visual Injection Strategy is introduced, which retrieves the most relevant visual features at each latent think token and updates the set of best visual patches. The updated patches are then injected into latent think token to achieve dynamic visual-textual interleaving. Experiments across seven multimodal reasoning benchmarks and various model architectures demonstrate that DMLR significantly improves reasoning and perception performance while maintaining high inference efficiency.

1. Introduction

Prior multimodal reasoning methods use explicit textual or visual reasoning but can suffer from weak grounding, unstable perception interaction, and overhead. DMLR instead combines confidence-guided latent optimization with dynamic visual injection, and experiments report consistent gains across architectures and tasks while maintaining efficiency.

  • Motivation: Textual-only reasoning can exhibit language bias and insufficient visual grounding, while image-based reasoning relies on external perception tools and may add instability and overhead.These limitations motivate a more integrated perception-reasoning approach.
  • Method: DMLR combines confidence-guided latent optimization with dynamic visual injection for test-time multimodal latent reasoning.Its latent think tokens are refined without additional training, while visual information is selectively integrated during reasoning.
  • Method: The framework dynamically determines whether to revisit visual information and which visual contents to inject, including the option to skip injection when confidence is sufficient.This supports selective visual-textual interleaving during latent reasoning.
  • Results: DMLR consistently outperforms other methods across diverse model architectures and multimodal tasks while maintaining high efficiency.The introduction presents this as the framework’s broad empirical outcome.

2. Related Work

Related work spans explicit textual or image-assisted reasoning and newer latent-space reasoning methods. These approaches motivate DMLR by exposing gaps in perception-reasoning alignment, tool stability, and inference efficiency.

  • Explicit Reasoning: Explicit multimodal reasoning methods separate or augment perception and reasoning through textual chains, image manipulation, or auxiliary visual tools.Examples include semantic Chain-of-Thought, local zooming, region highlighting, and generated intermediate reasoning steps.
  • Latent Reasoning: Latent reasoning methods shift inference from explicit token sequences into model representations through trained optimization or training-free manipulation of latent activations.This literature establishes latent-space reasoning as a separate direction from explicit multimodal reasoning.

3. Preliminary and Motivation

The motivation analysis finds that visual dependence is concentrated in selected reasoning steps and varies across chains, while confidence tracks reasoning quality and visual grounding. These findings support dynamic, confidence-guided visual use rather than fixed perception at every step.

  • 3.1 Dynamic Perception-Reasoning is Necessary: Visual dependence is uneven across reasoning: only a small subset of tokens strongly relies on visual features, and chains differ substantially in visual reliance.The visual dependency score compares token probabilities under original and perturbed visual inputs; higher scores indicate stronger visual dependence.
  • 3.1 Dynamic Perception-Reasoning is Necessary: Reasoning chains with stronger visual reliance consistently achieve higher accuracy than chains with weaker visual reliance.This chain-level result shows that visual use varies meaningfully across trajectories rather than following a fixed pattern.
  • 3.2 Internal Confidence Affects Multimodal Reasoning: Correct chains contain more positive confidence gains than incorrect chains, and faithful chains show stronger confidence improvement than spurious chains.Confidence gain measures the change in ground-truth answer probability after adding a reasoning step.
  • 3.2 Internal Confidence Affects Multimodal Reasoning: Hallucinated reasoning steps have lower confidence and weaker image relevancy than non-hallucinated steps, linking confidence with visual grounding quality.Non-hallucinatory steps maintain higher and more stable confidence with stronger visual alignment.

4. Methodology

DMLR performs test-time multimodal latent reasoning by optimizing learnable think tokens with confidence feedback while dynamically selecting and injecting relevant visual patches. Controlled perturbations support exploration, and the resulting latent states are decoded into outputs.

  • Problem Formulation: DMLR introduces learnable latent think tokens and optimizes their embeddings at test time using a confidence-based reward while keeping model parameters fixed.The tokens act as a mental draft and are iteratively refined toward directions that maximize the reward.
  • Dynamic Multimodal Latent Reasoning: The framework combines dynamic visual injection with confidence-guided latent optimization as its two core processes.The visual and latent optimization processes jointly support multimodal reasoning during inference.
  • Reward Formulation: Confidence is estimated from truncated entropy over top-k token probabilities, with lower entropy corresponding to higher prediction confidence.The sequence reward is defined as the complement of mean truncated entropy across latent think tokens.
  • Test-Time Latent Optimization: REINFORCE-based policy gradients update latent think tokens, while Gaussian perturbations provide controlled exploration before each optimization step.The perturbation variance controls exploration magnitude, and the latent update uses the policy-gradient objective.
  • Dynamic Visual Injection Strategy: At each iteration, DMLR resamples relevant candidate patches and retains them as the best visual patches when they improve the reward.The strategy initializes best patches from latent-token attention, compares candidate rewards, and preserves the previous best otherwise.
  • Theoretical Analysis: The theoretical analysis links confidence ascent to improved reasoning quality under positive gradient alignment and links visual injection to increased confidence through mutual information.The second theorem states I(T̂; z_v) ≥ I(T; z_v) implies J_conf(T̂) ≥ J_conf(T).

5. Experiments

Across diverse multimodal benchmarks, DMLR improves reasoning and perception while balancing accuracy with inference efficiency. Ablations show that iterative, confidence-guided optimization benefits from selective visual patches, moderate perturbations, and a moderate number of latent tokens.

  • 5.2 Main Results: DMLR achieves the best performance on over 95% of tasks and improves both mathematical and visual reasoning across model paradigms.Qwen2.5-VL-7B gains +1.5% in mathematics and +0.9% in visual reasoning, while R1-OneVision gains +4.5% and +3.45%, respectively.
  • 5.3 Ablation Study: DMLR selects relevant visual patches more stably than removing injection or injecting all patches, which improves perceptual accuracy without redundant visual information.The ablation compares strategies across multiple benchmarks and identifies best-patch injection as the stable alternative.
  • 5.3 Ablation Study: Accuracy steadily increases with more optimization iterations for both base and reasoning models, with reasoning models remaining consistently stronger and benefiting from continued refinement.The trend is reported for both reasoning and perception tasks.
  • 5.3 Ablation Study: Moderate perturbation and candidate-patch counts improve performance, whereas excessive noise or patches destabilize optimization and reduce accuracy.The selected best patch also becomes increasingly stable as iterations progress, indicating convergence toward a small relevant set.
  • 5.3 Ablation Study: Using 2–4 latent think tokens yields stable gains, but additional tokens degrade performance and make reasoning-model optimization less stable.The reported trend applies to both reasoning and perception tasks.
  • 5.4 Analysis: DMLR maintains task-relevant visual attention and separates latent think tokens from text and image embeddings while operating in a favorable accuracy-efficiency trade-off.The qualitative analyses compare DMLR with explicit CoT and characterize latent-token geometry and inference efficiency.

6. Conclusion

The paper analyzes how confidence and visual information interact during multimodal reasoning, then introduces DMLR as a training-free latent optimization framework with dynamic visual injection. Experiments show consistent gains in reasoning and perception while retaining stable, efficient inference.

  • 6. Conclusion: DMLR combines confidence-guided latent optimization with dynamic visual injection to refine multimodal reasoning at test time.The framework retrieves visual evidence only when needed and requires no additional training.
  • 6. Conclusion: Across varied tasks and architectures, DMLR consistently boosts both reasoning and perception performance as a stable, training-free alternative to other methods.The conclusion characterizes the framework as improving both task dimensions without additional training.

A.1 Datasets

The evaluation covers mathematics reasoning, visual reasoning, and multimodal composition across seven benchmarks. Dataset sizes are capped at 1000 instances.

  • A.1 Datasets: The dataset suite spans mathematics reasoning on MathVistamini, MathVisionmini, and MM Math; visual reasoning on HallusionBench and MMVP; and composition on MMStar and ScienceQA.These benchmarks cover visual mathematics, hallucination and visual-pattern reasoning, vision-dependent multimodal ability, and multimodal science questions.
  • A.1 Datasets: ScienceQA evaluates multimodal science questions with textual and visual contexts, including answer accuracy and explanation quality.The benchmark spans diverse subjects and skills.
  • A.1 Datasets: The experiments limit the maximum sample size to 1000 instances for every dataset.

A.2 Evaluation Setting

The evaluation uses a unified prompting setup with greedy decoding and task-specific reasoning prompts. Verifiable answers are extracted from boxed outputs and matched against ground truth.

  • A.2 Evaluation Setting: All models use a unified prompting setup and greedy decoding with do_sample=false unless stated otherwise.
  • A.2 Evaluation Setting: The reasoning prompt places internal reasoning in <think> tags and the final response in <answer> tags.
  • A.2 Evaluation Setting: The mathematical problem prompt requests step-by-step reasoning and a final answer enclosed in \boxed{}.
  • A.2 Evaluation Setting: Evaluation extracts content inside \boxed{} with regular expressions and matches it against verifiable ground-truth answers.

A.3 Parameters Setup

The setup fixes latent-token, patch-selection, optimization, perturbation, precision, and hardware parameters for DMLR experiments.

  • A.3 Parameters Setup: DMLR uses 4 latent think tokens and dynamically inserts visual patches into the latent stream during generation.Each latent token is followed by dynamic visual refreshing.
  • A.3 Parameters Setup: The method initially injects 2 patches, then selects 2 highest-attention patches per iteration, with at most 16 inserted patches.
  • A.3 Parameters Setup: Latent optimization runs for 15 steps with learning rate 1 × 10^-3 and perturbation magnitude σ = 0.1 with decay factor 0.95.
  • A.3 Parameters Setup: Experiments use a fixed random seed of 42, NVIDIA H100 GPUs, float32 precision, and eager attention.

B. Case Study

The case studies illustrate DMLR’s visual perception and reasoning behavior through truck-door and cube-beetle examples, including a contrast with baseline outputs.

  • B. Case Study: The qualitative analysis associates DMLR with more reliable, fine-grained visual perception and semantically coherent internal representations.
  • B. Case Study: DMLR identifies the truck-cab door as open from visible interior and misalignment, whereas the baseline judges it closed.
  • B. Case Study: DMLR’s qualitative reasoning example traces the beetle’s alternating turns around the cube until it returns to point P.
  • B. Case Study: A separate beetle explanation states that the return occurs after 6 edges, while another baseline output reports 8 edges.
Loading 2512.12623v3…