{"paper":{"title":"LLaMA-Adapter V2: Parameter-Efficient Visual Instruction Model","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"LLaMA-Adapter V2 turns LLaMA into an open-ended visual instruction follower by adding only 14 million parameters.","cross_cats":["cs.AI","cs.CL","cs.LG","cs.MM"],"primary_cat":"cs.CV","authors_text":"Aojun Zhou, Conghui He, Hongsheng Li, Jiaming Han, Pan Lu, Peng Gao, Renrui Zhang, Shijie Geng, Wei Zhang, Xiangyu Yue, Yu Qiao, Ziyi Lin","submitted_at":"2023-04-28T17:59:25Z","abstract_excerpt":"How to efficiently transform large language models (LLMs) into instruction followers is recently a popular research direction, while training LLM for multi-modal reasoning remains less explored. Although the recent LLaMA-Adapter demonstrates the potential to handle visual inputs with LLMs, it still cannot generalize well to open-ended visual instructions and lags behind GPT-4. In this paper, we present LLaMA-Adapter V2, a parameter-efficient visual instruction model. Specifically, we first augment LLaMA-Adapter by unlocking more learnable parameters (e.g., norm, bias and scale), which distribu"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Compared to the original LLaMA-Adapter, our LLaMA-Adapter V2 can perform open-ended multi-modal instructions by merely introducing 14M parameters over LLaMA.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the early-fusion placement and the disjoint-parameter joint training will continue to prevent task interference and maintain generalization when the instruction data distribution shifts or when larger base models are used.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"LLaMA-Adapter V2 achieves open-ended visual instruction following in LLMs by unlocking more parameters, early fusion of visual tokens, and joint training on disjoint parameter groups with only 14M added parameters.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"LLaMA-Adapter V2 turns LLaMA into an open-ended visual instruction follower by adding only 14 million parameters.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"c9d348bc3257a2da07569ba9ae3472017e71220e651ebc57c1ec811a410510be"},"source":{"id":"2304.15010","kind":"arxiv","version":1},"verdict":{"id":"4ed12321-98b1-4490-beda-aebc663e4e58","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T08:36:50.707814Z","strongest_claim":"Compared to the original LLaMA-Adapter, our LLaMA-Adapter V2 can perform open-ended multi-modal instructions by merely introducing 14M parameters over LLaMA.","one_line_summary":"LLaMA-Adapter V2 achieves open-ended visual instruction following in LLMs by unlocking more parameters, early fusion of visual tokens, and joint training on disjoint parameter groups with only 14M added parameters.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the early-fusion placement and the disjoint-parameter joint training will continue to prevent task interference and maintain generalization when the instruction data distribution shifts or when larger base models are used.","pith_extraction_headline":"LLaMA-Adapter V2 turns LLaMA into an open-ended visual instruction follower by adding only 14 million parameters."},"references":{"count":79,"sample":[{"doi":"","year":null,"title":"https://sharegpt.com/","work_id":"42263fdc-d42c-4b2a-8562-4f2dc47ecf6c","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Flamingo: a visual language model for few-shot learning","work_id":"31e3af5c-9fec-43d9-b533-5bb70172dd15","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2018,"title":"Bottom-up and top-down attention for image captioning and visual question answering","work_id":"40ff759e-80ef-4b6f-86f5-11ea4321f5c8","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":1901,"title":"Lan- guage models are few-shot learners","work_id":"5b23bebc-10b7-4150-9a97-e3f37825079e","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2021,"title":"Conceptual 12m: Pushing web-scale image-text pre- training to recognize long-tail visual concepts","work_id":"c9a39a05-4f9a-45e3-92e2-f310468325af","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":79,"snapshot_sha256":"dd7cf17851379fca4c7b1bc3da5a359dac627288662817514af94575b0695681","internal_anchors":19},"formal_canon":{"evidence_count":2,"snapshot_sha256":"6b6edd9601be9795336ad80bbde7cdb959d9fbaefb4514280dc3923e30cfe686"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}