{"paper":{"title":"Diet Your LLM: Dimension-wise Global Pruning of LLMs via Merging Task-specific Importance Score","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"DIET prunes LLM dimensions by merging per-task activation scores into one global mask without any training.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Jaehyung Kim, Jimyung Hong","submitted_at":"2026-03-25T06:28:58Z","abstract_excerpt":"Large language models (LLMs) have demonstrated remarkable capabilities, but their massive scale poses significant challenges for practical deployment. Structured pruning offers a promising solution by removing entire dimensions or layers, yet existing methods face critical trade-offs: task-agnostic approaches cannot adapt to task-specific requirements, while task-aware methods require costly training to learn task adaptability. We propose DIET (Dimension-wise global pruning of LLMs via merging Task-wise importance scores), a training-free structured pruning method that combines dimension-level"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"at 20% sparsity on Gemma-2 2B, DIET achieves near 10% average accuracy improvement, compared to previous state-of-the-art structured pruning methods.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"that activation magnitudes computed from only 100 samples per task are sufficient to produce a reliable global importance ranking that generalizes across unseen tasks and inputs.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"DIET prunes LLMs dimension-wise by merging task-specific activation importance scores via majority voting on 100 samples per task, yielding up to 10% accuracy gains over prior structured methods at 20% sparsity on Gemma-2 models.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"DIET prunes LLM dimensions by merging per-task activation scores into one global mask without any training.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"80943eec44744cd9029eded8a210fdac1618d561f9e108b13974ec5b9a67f168"},"source":{"id":"2603.23985","kind":"arxiv","version":3},"verdict":{"id":"f9cd5a15-2fcb-495c-9242-b8b6cf769d9e","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T00:51:10.787003Z","strongest_claim":"at 20% sparsity on Gemma-2 2B, DIET achieves near 10% average accuracy improvement, compared to previous state-of-the-art structured pruning methods.","one_line_summary":"DIET prunes LLMs dimension-wise by merging task-specific activation importance scores via majority voting on 100 samples per task, yielding up to 10% accuracy gains over prior structured methods at 20% sparsity on Gemma-2 models.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"that activation magnitudes computed from only 100 samples per task are sufficient to produce a reliable global importance ranking that generalizes across unseen tasks and inputs.","pith_extraction_headline":"DIET prunes LLM dimensions by merging per-task activation scores into one global mask without any training."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2603.23985/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"a189ac9aee3c7b82f5eb3e0bdce72037212b4b18be16d87e14f59cb4ef9734e1"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}