Look Light, Think Heavy: What Multimodal Chain-of-Thought Reasoning Can and Cannot Do
Summary
This paper systematically evaluates multimodal Chain-of-Thought reasoning across 12 tasks, finding it selectively effective for reasoning tasks but detrimental for perception tasks, and identifies a 'Look Light, Think Heavy' pattern where visual introspection declines during reasoning.
View Cached Full Text
Cached at: 06/25/26, 05:17 AM
Paper page - Look Light, Think Heavy: What Multimodal Chain-of-Thought Reasoning Can and Cannot Do
Source: https://huggingface.co/papers/2606.22565
Abstract
Multimodal Chain-of-Thought reasoning shows selective effectiveness across different tasks, with limitations in maintaining visual introspection during reasoning processes.
Chain-of-Thought(CoT) has become a standard method for improving reasoning capabilities inlarge language models(LLMs) by eliciting step-by-step thinking, but its effectiveness inmultimodal tasksremains unclear. In this paper, we aim to systematically investigate the key question: What can multimodalChain-of-Thoughtreasoning do, and where and why does it fall short? To this end, we evaluate 12multimodal tasksacross perception and reasoning categories using both 14 non-reasoning models and 8 reasoning models. Our analysis reveals several important findings: (1) CoT is not a free lunch and should be used selectively depending on the specific requirements of each task. For perception tasks, CoT can lead to undesirable side effects, such as reduced performance invisual groundingandobject counting. In contrast, it proves effective for reasoning tasks involving mathematical, scientific, andmulti-image reasoning; (2) Compared to original models, existing open-source multimodal reasoning models often yield only marginal overall improvements, possibly due to an overemphasis onmathematical reasoningat the expense of broader capabilities; (3)Visual reasoningremains a key bottleneck for current multimodal CoT, as models exhibit a Look Light, Think Heavy pattern where verbal reflection rises and falls during reasoning, whereas visual reflection consistently diminishes. These findings suggest that while multimodal CoT handles verbal reflection relatively well, it lacks the ability to maintain deep visual introspection throughout the reasoning process.
View arXiv pageView PDFAdd to collection
Get this paper in your agent:
hf papers read 2606\.22565
Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash
Models citing this paper0
No model linking this paper
Cite arxiv.org/abs/2606.22565 in a model README.md to link it from this page.
Datasets citing this paper0
No dataset linking this paper
Cite arxiv.org/abs/2606.22565 in a dataset README.md to link it from this page.
Spaces citing this paper0
No Space linking this paper
Cite arxiv.org/abs/2606.22565 in a Space README.md to link it from this page.
Collections including this paper0
No Collection including this paper
Add this paper to acollectionto link it from this page.
Similar Articles
Chain-of-Thought Degrades Visual Spatial Reasoning Capabilities of Multimodal LLMs
Research shows Chain-of-Thought prompting harms visual-spatial reasoning in multimodal LLMs due to shortcut learning and hallucinating visual details from text alone.
See2Think: Do Multimodal Models Really Use Intermediate Visual States?
This paper introduces See2Think, a unified evaluation framework to test whether multimodal models genuinely rely on intermediate visual states during reasoning. It finds that visual reasoning is highly model- and environment-dependent, with faithful rendering being the main bottleneck.
The strange thing about LLM reasoning research: we're now trying to remove the chain-of-thought traces
The article discusses a shift in LLM reasoning research from making reasoning explicit via chain-of-thought to exploring latent reasoning that doesn't require language traces, questioning whether visibility is necessary for effective reasoning.
ThoughtFold: Folding Reasoning Chains via Introspective Preference Learning
ThoughtFold proposes a framework using introspective preference learning to reduce redundant explorations in Chain-of-Thought reasoning for Large Reasoning Models, achieving ~56% token reduction on DeepSeek-R1-Distill-Qwen-7B without accuracy loss.
Do Vision-Language Models Truly Perform Vision Reasoning? A Rigorous Study of the Modality Gap
This paper introduces CrossMath, a controlled multimodal reasoning benchmark that reveals a critical limitation in current vision-language models: they perform reasoning primarily in textual space rather than genuine vision-grounded reasoning, with visual input often degrading performance compared to text-only baselines. The authors propose fine-tuning approaches to mitigate this modality gap and improve multimodal reasoning capabilities.