{"paper":{"title":"GRIP-VLM: Group-Relative Importance Pruning for Efficient Vision-Language Models","license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","headline":"GRIP-VLM uses reinforcement learning to optimize discrete visual token pruning in VLMs, avoiding suboptimal local minima from gradient relaxations.","cross_cats":["cs.AI"],"primary_cat":"cs.CV","authors_text":"Hao Wen, Liang Mi, Lichen Pang, Mingzhe Huang, Shansong Yang, Ting Cao, Weijun Wang, Xin Ding, Yuanchun Li, Yunxin Liu","submitted_at":"2026-05-13T11:32:03Z","abstract_excerpt":"In Vision-Language Models (VLMs), processing a massive number of visual tokens incurs prohibitive computational overhead. While recent training-aware pruning methods attempt to selectively discard redundant tokens, they largely rely on continuous-gradient relaxations. However, visual token pruning is inherently a discrete, non-convex combinatorial problem; consequently, these continuous approximations frequently trap the optimization in sub-optimal local minima, especially under aggressive compression budgets. To overcome this fundamental bottleneck, we propose GRIP-VLM, a Group-Relative Impor"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"GRIP-VLM consistently outperforms heuristic and supervised-learning baselines, achieving a superior Pareto frontier and delivering up to a 15% inference speedup at equal accuracy.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the Group Relative Policy Optimization agent can reliably discover high-quality discrete pruning masks across varying compression budgets without retraining and without the instability typical of RL on combinatorial spaces.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"GRIP-VLM applies group-relative policy optimization via reinforcement learning to prune visual tokens in VLMs, yielding up to 15% inference speedup at matched accuracy over prior methods.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"GRIP-VLM uses reinforcement learning to optimize discrete visual token pruning in VLMs, avoiding suboptimal local minima from gradient relaxations.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"0f8705bb0e12e53541017e23ee834925ef3787f53d5c3f1cdc9cafa0005510e1"},"source":{"id":"2605.13375","kind":"arxiv","version":1},"verdict":{"id":"7199ddd1-4a63-4022-bb64-fbece73cbafa","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T19:12:10.099597Z","strongest_claim":"GRIP-VLM consistently outperforms heuristic and supervised-learning baselines, achieving a superior Pareto frontier and delivering up to a 15% inference speedup at equal accuracy.","one_line_summary":"GRIP-VLM applies group-relative policy optimization via reinforcement learning to prune visual tokens in VLMs, yielding up to 15% inference speedup at matched accuracy over prior methods.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the Group Relative Policy Optimization agent can reliably discover high-quality discrete pruning masks across varying compression budgets without retraining and without the instability typical of RL on combinatorial spaces.","pith_extraction_headline":"GRIP-VLM uses reinforcement learning to optimize discrete visual token pruning in VLMs, avoiding suboptimal local minima from gradient relaxations."},"references":{"count":42,"sample":[{"doi":"","year":2023,"title":"Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models","work_id":"4819f738-f69f-49dd-8bed-404f647de63a","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"MiniGPT-4: Enhancing Vision-Language Understanding with Advanced Large Language Models","work_id":"a7e3a737-e007-42bc-be89-c4d34c5ee071","ref_index":2,"cited_arxiv_id":"2304.10592","is_internal_anchor":true},{"doi":"","year":2024,"title":"Improved baselines with visual instruction tuning","work_id":"1191d1e0-5e80-4225-a064-e1c04d70e44d","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"Visual instruction tuning.Advances in neural information processing systems, 2024","work_id":"93eddd32-4e1d-4fb1-aa72-53eb41f6d1f2","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Llama 2: Open Foundation and Fine-Tuned Chat Models","work_id":"68a5177f-d644-44c1-bd4f-4e5278c22f5d","ref_index":5,"cited_arxiv_id":"2307.09288","is_internal_anchor":true}],"resolved_work":42,"snapshot_sha256":"a3d9548cde7690ae65e1aab4675a2c375d1d4f1d743b6dabd1d24314116a5b43","internal_anchors":12},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}