{"paper":{"title":"LLaVA-CoT: Let Vision Language Models Reason Step-by-Step","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"By training on structured four-stage annotations, LLaVA-CoT lets vision-language models reason autonomously and outperform larger models with only 100k samples.","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Guowei Xu, Hao Li, Lichao Sun, Li Yuan, Peng Jin, Yibing Song, Ziang Wu","submitted_at":"2024-11-15T18:58:31Z","abstract_excerpt":"Large language models have demonstrated substantial advancements in reasoning capabilities. However, current Vision-Language Models (VLMs) often struggle to perform systematic and structured reasoning, especially when handling complex visual question-answering tasks. In this work, we introduce LLaVA-CoT, a large VLM designed to conduct autonomous multistage reasoning. Unlike chain-of-thought prompting, LLaVA-CoT independently engages in sequential stages of summarization, visual interpretation, logical reasoning, and conclusion generation. This structured approach enables LLaVA-CoT to achieve "},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"with only 100k training samples and test-time scaling, LLaVA-CoT not only outperforms its base model by 9.4% on a wide range of multimodal reasoning benchmarks, but also surpasses the performance of larger and even closed-source models, such as Gemini-1.5-pro, GPT-4o-mini, and Llama-3.2-90B-Vision-Instruct.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the human-provided structured reasoning annotations in the LLaVA-CoT-100k dataset faithfully capture effective multistage reasoning without introducing systematic biases or annotation artifacts that the model simply memorizes.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"LLaVA-CoT adds autonomous multistage reasoning to vision-language models, delivering 9.4% gains over its base model and outperforming larger models like Gemini-1.5-pro on reasoning benchmarks via a 100k annotated dataset and SWIRES test-time scaling.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"By training on structured four-stage annotations, LLaVA-CoT lets vision-language models reason autonomously and outperform larger models with only 100k samples.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"e493f6aa8471655f9d546fbcf7c8330b5f777cb99e5eb857e8c488adfc19b644"},"source":{"id":"2411.10440","kind":"arxiv","version":6},"verdict":{"id":"7c3f801f-376d-4a69-b65a-a4a169c90d26","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-16T11:31:07.126334Z","strongest_claim":"with only 100k training samples and test-time scaling, LLaVA-CoT not only outperforms its base model by 9.4% on a wide range of multimodal reasoning benchmarks, but also surpasses the performance of larger and even closed-source models, such as Gemini-1.5-pro, GPT-4o-mini, and Llama-3.2-90B-Vision-Instruct.","one_line_summary":"LLaVA-CoT adds autonomous multistage reasoning to vision-language models, delivering 9.4% gains over its base model and outperforming larger models like Gemini-1.5-pro on reasoning benchmarks via a 100k annotated dataset and SWIRES test-time scaling.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the human-provided structured reasoning annotations in the LLaVA-CoT-100k dataset faithfully capture effective multistage reasoning without introducing systematic biases or annotation artifacts that the model simply memorizes.","pith_extraction_headline":"By training on structured four-stage annotations, LLaVA-CoT lets vision-language models reason autonomously and outperform larger models with only 100k samples."},"references":{"count":68,"sample":[{"doi":"","year":null,"title":"https : / / opencompass","work_id":"fbaa7c31-1483-4ab8-b7d9-4fbb67f9e1d5","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"Available at: https://www","work_id":"82bd6267-6caa-4ddf-a95a-11d7c2fe5afc","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"Gpt-4o system card, 2024","work_id":"adccb5da-15dc-4b4e-af45-0758f6342e7d","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"Variational best-of-n alignment, 2024","work_id":"975889b4-94a1-4f87-808c-751080915571","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2020,"title":"Neuro-symbolic visual reasoning: Disentangling","work_id":"55532825-9fd4-4d62-aacd-bf7d8915c455","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":68,"snapshot_sha256":"253a79e483a70865339e1d8a1c2569b0a9850de6bf157d4eebdf0d057f8dcf63","internal_anchors":3},"formal_canon":{"evidence_count":2,"snapshot_sha256":"d6d3ef0647160794211088bf4ccb2eea9985534d9d19b104dbc1bb6fa3ca30ca"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}