AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

LLaVA-CoT: Let Vision Language Models Reason Step-by-Step

Guowei Xu, Peng Jin, Ziang Wu, Hao Li, Yibing Song, Lichao Sun, Li Yuan

ICCV 2025 · Vision-language and multimodal › Multimodal LLMs · approach: Chain-of-thought multimodal reasoning · 452 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add structured four-stage reasoning data (vs original Q&A pairs) (ablation)LLaVA-CoT · Six-benchmark averageaverage accuracy: 59.0 -> 62.4Improved
train on original Q&A pairs instead of reasoning data (ablation)Llama-3.2-11B-Vision-Instruct · MMVetaccuracy: 57.6 -> 49.9Hurt
add structured stage tags (ablation)LLaVA-CoT · Six-benchmark averageaverage accuracy: 60.9 -> 62.4Improved
add structured stage tags (ablation)LLaVA-CoT · MMBench V1.1accuracy: 74.2 -> 75.0Improved
add test-time stage-wise retracing searchLLaVA-CoT · Six-benchmark averageaverage accuracy: 62.4 -> 65.5Improved
add test-time stage-wise retracing searchLLaVA-CoT · MMStaraccuracy: 57.6 -> 62.5Improved
train on multi-task sub-skills instead of one CoT chain (ablation)LLaVA-CoT · Reasoning benchmarks averageaverage accuracy: 57.7 -> 63.1Improved
keep natural stage order instead of shuffled order (ablation)LLaVA-CoT · Reasoning benchmarks averageaverage accuracy: 58.2 -> 63.1Improved
prompt base model with structured CoT without trainingLlama-3.2-Vision-Instruct · Reasoning benchmarks averageaverage accuracy: 56.9 -> 56.9No clear effect
prompt GPT-4o with structured CoTGPT-4o-0806 · Reasoning benchmarks averageaverage accuracy: 71.8 -> 74.1Improved
apply SFT with LLaVA-CoT-100k to base modelLlama-3.2-11B-Vision-Instruct · Six-benchmark averageaverage accuracy: 56.6 -> 62.4Improved
apply SFT with LLaVA-CoT-100k to base modelLlama-3.2-11B-Vision-Instruct · MMStar math skillaccuracy: 45.2 -> 64.0Improved

Later papers that compared with this method

7 later accepted papers report LLaVA-CoT as a baseline in their results tables; 7 of them report a better number on most of the results they share with it.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy