← All notes

Summary

This paper introduces LLaVA, a Large Language and Vision Assistant. The authors use language-only GPT-4 to generate multimodal language-image instruction-following data, then instruction-tune an end-to-end model that connects a vision encoder and an LLM. The vision encoder is the pre-trained CLIP visual encoder ViT-L/14, joined to Vicuna by a simple linear layer that maps image features into the word embedding space. They collect 158K unique language-image instruction-following samples in total. When fine-tuned on Science QA, the synergy of LLaVA and GPT-4 reaches a new state-of-the-art accuracy of 92.53%. LLaVA demonstrates impressive multimodal chat abilities, and the authors make the data, model, and code publicly available. [S1]

Claims

• id: C1 claim: 'LLaVA connects a vision encoder and an LLM for visual and language understanding.' source: S1 evidence: "connects a vision encoder and an LLM"
• id: C2 claim: 'They collect 158K unique language-image instruction-following samples in total.' source: S1 evidence: "158K unique language-image instruction-following samples in total"
• id: C3 claim: 'When fine-tuned on Science QA, LLaVA plus GPT-4 reaches 92.53% accuracy.' source: S1 evidence: "the synergy of LLaVA and GPT-4 achieves a new state-of-the-art accuracy of 92.53%"

Verification

• Code available: Paper points to https://llava-vl.github.io, which returned HTTP 200 on 2026-11-11. The HTML also names the anonymized GitHub repo LLaVA-Annonymous/LLaVA (HTTP 200). The public code repo https://github.com/haotian-liu/LLaVA also returned HTTP 200. Code not run.
• Reproduced: not reproduced (2026-11-11). Code not run.
• Related work: Live notes ao-evals-multimodal-agents. This batch also has pointer notes for CLIP and Flamingo, which this recipe builds on.
• Limits: The authors discuss chatbot failures on the in-the-wild bench, including cases that need high-resolution images, wide knowledge coverage, or multilingual reading. Appendix A also flags hallucination, biases, and evaluation complexities.
• License: CC BY 4.0 on the abs page (2026-11-11; icon by-4.0.png; http://creativecommons.org/licenses/by/4.0/). Submitted 17 Apr 2023 (v1); last revised 11 Dec 2023 (v2). Comments: NeurIPS 2023 Oral. Sources: [S1] https://arxiv.org/abs/2304.08485, v2 (11 Dec 2023). checked_on: 2026-11-11.

When to apply

Use when asking how an open vision-language assistant was built by connecting a frozen CLIP encoder to a chat LLM with a linear projector and visual instruction data.

Facts

• id: F1 claim: 'LLaVA connects the CLIP ViT-L/14 vision encoder to Vicuna with a simple linear layer.' confidence: medium date: 2026-11-11 sources: [S1] license: CC BY 4.0
• id: F2 claim: 'The authors collect 158K unique language-image instruction-following samples in total.' confidence: medium date: 2026-11-11 sources: [S1] license: CC BY 4.0
• id: F3 claim: 'Fine-tuned on Science QA, the synergy of LLaVA and GPT-4 reaches 92.53% accuracy.' confidence: medium date: 2026-11-11 sources: [S1] license: CC BY 4.0

Sources

• [S1] arXiv 2304.08485, https://arxiv.org/abs/2304.08485. License: CC BY 4.0. v2 (11 Dec 2023). Accessed 2026-11-11.

License: this note is CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Summary by iinventedlight (AI-assisted). Paper is CC BY 4.0.

@misc{iil_note_6abjm38n,
  author = {{'iinventedlight (AI-assisted)'}},
  title = {{Visual Instruction Tuning}},
  howpublished = {iinventedlight Library},
  year = {2026},
  date = {2026-11-11},
  url = {https://iinventedlight.com/library/notes/2304-08485-visual-instruction-tuning/},
  note = {Permanent ID iil:note:6abjm38n. License: CC-BY-4.0.},
}

Markdown

Spotted a mistake? Use the MCP feedback tool or email hello@iinventedlight.com.

For language models and agents: start at /llms.txt | /agents.md | /.well-known/agent-library.json. This page in Markdown: /library/notes/2304-08485-visual-instruction-tuning.md.