Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Why Reasoning Matters? A Survey of Advancements in Multimodal Reasoning (v1)

Jing Bi, Susan Liang, Xiaofei Zhou, Pinxin Liu, Junjia Guo, Yunlong Tang, Luchuan Song, Chao Huang, Ali Vosoughi, Guangyu Sun, Jinxi He, Jiarui Wu, Shu Yang, Daoan Zhang, Chen Chen, Lianggong Bruce Wen, Zhang Liu, Jiebo Luo, Chenliang Xu

TL;DR

The paper surveys reasoning techniques in both textual and multimodal large language models, detailing post-training policy optimization and test-time trajectory search as pathways to improved reasoning accuracy and coherence. It highlights methods ranging from RL/IL and reward alignment to adaptive inference, dynamic search, and curated datasets, emphasizing spatial-temporal modeling and robust visual grounding. Through extensive benchmarking coverage, the work identifies current gaps in multimodal reasoning and outlines practical directions for scalable, grounded reasoning frameworks that integrate cognitive science insights. The synthesis bridges theory and practice, offering guidance for future research and system design in multimodal reasoning tasks.

Abstract

Reasoning is central to human intelligence, enabling structured problem-solving across diverse tasks. Recent advances in large language models (LLMs) have greatly enhanced their reasoning abilities in arithmetic, commonsense, and symbolic domains. However, effectively extending these capabilities into multimodal contexts-where models must integrate both visual and textual inputs-continues to be a significant challenge. Multimodal reasoning introduces complexities, such as handling conflicting information across modalities, which require models to adopt advanced interpretative strategies. Addressing these challenges involves not only sophisticated algorithms but also robust methodologies for evaluating reasoning accuracy and coherence. This paper offers a concise yet insightful overview of reasoning techniques in both textual and multimodal LLMs. Through a thorough and up-to-date comparison, we clearly formulate core reasoning challenges and opportunities, highlighting practical methods for post-training optimization and test-time inference. Our work provides valuable insights and guidance, bridging theoretical frameworks and practical implementations, and sets clear directions for future research.

Why Reasoning Matters? A Survey of Advancements in Multimodal Reasoning (v1)

TL;DR

The paper surveys reasoning techniques in both textual and multimodal large language models, detailing post-training policy optimization and test-time trajectory search as pathways to improved reasoning accuracy and coherence. It highlights methods ranging from RL/IL and reward alignment to adaptive inference, dynamic search, and curated datasets, emphasizing spatial-temporal modeling and robust visual grounding. Through extensive benchmarking coverage, the work identifies current gaps in multimodal reasoning and outlines practical directions for scalable, grounded reasoning frameworks that integrate cognitive science insights. The synthesis bridges theory and practice, offering guidance for future research and system design in multimodal reasoning tasks.

Abstract

Reasoning is central to human intelligence, enabling structured problem-solving across diverse tasks. Recent advances in large language models (LLMs) have greatly enhanced their reasoning abilities in arithmetic, commonsense, and symbolic domains. However, effectively extending these capabilities into multimodal contexts-where models must integrate both visual and textual inputs-continues to be a significant challenge. Multimodal reasoning introduces complexities, such as handling conflicting information across modalities, which require models to adopt advanced interpretative strategies. Addressing these challenges involves not only sophisticated algorithms but also robust methodologies for evaluating reasoning accuracy and coherence. This paper offers a concise yet insightful overview of reasoning techniques in both textual and multimodal LLMs. Through a thorough and up-to-date comparison, we clearly formulate core reasoning challenges and opportunities, highlighting practical methods for post-training optimization and test-time inference. Our work provides valuable insights and guidance, bridging theoretical frameworks and practical implementations, and sets clear directions for future research.

Paper Structure

This paper contains 17 sections, 6 equations, 5 figures.

Table of Contents

  1. Introduction
  2. Background
  3. Post-training improvements
  4. Policy Optimization
  5. Reward Alignment
  6. Model Architecture
  7. Spatial Temporal Modeling
  8. Test-Time compute
  9. Search Strategies
  10. Adaptive Inference
  11. Reward
  12. Feedback
  13. Iterative Refinement
  14. Dataset Curation
  15. Datasets and Benchmarks
  16. ...and 2 more sections

Figures (5)

  • Figure 1: Papers on visual reasoning per quarter over the last three years, with state computed using referenced papers. (Data current to mid-March 2025.)
  • Figure 2: Framework illustrating training and inference for reasoning optimization. A virtuous cycle emerges as better policies generate improved trajectories, which in turn enhance the model through stronger supervision.
  • Figure 3: Search framework where language models explore and refine reasoning paths. Trajectories are scored using reward models, based on expected utility or final output quality, and guided by feedback, world models, and evaluators to select the most promising steps.
  • Figure 4: Comprehensive Overview of Methods and Frameworks focus on test-time compute
  • Figure 5: Comprehensive Overview of Methods and Frameworks focus on post-training improvement