{"paper":{"title":"CoSpaDi: Compressing LLMs via Calibration-Guided Sparse Dictionary Learning","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"CoSpaDi replaces low-rank factorization with a sparse dictionary model that better preserves LLM accuracy at 20-40 percent compression.","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Ammar Ali, Denis Makhov, Dmitriy Shopkhoev, Magauiya Zhussip, Stamatios Lefkimmiatis","submitted_at":"2025-09-26T08:55:09Z","abstract_excerpt":"Post-training LLM compression often relies on low-rank approximations, which force all columns of a projection matrix to share a single low-dimensional subspace. We propose CoSpaDi, a training-free compression framework that replaces this single-subspace assumption with a union-of-subspaces model via sparse dictionary learning. CoSpaDi factorizes each weight matrix into a dense dictionary and column-sparse coefficients, allowing different columns to select different subsets of dictionary atoms at the same storage budget. To preserve model behavior, we use calibration activations to transform f"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Across Llama and Qwen model families, CoSpaDi consistently improves the accuracy-compression and perplexity-compression trade-offs over state-of-the-art SVD-based baselines and strong structured pruning baselines at 20-40% compression ratios.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The assumption that minimizing functional reconstruction error on a small calibration set will produce a factorization whose downstream task performance remains close to the original model without any fine-tuning or further adaptation.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"CoSpaDi introduces a training-free sparse dictionary learning framework for post-training LLM compression that optimizes functional reconstruction error via activation-derived orthonormalization and achieves improved accuracy-compression trade-offs over SVD and pruning baselines.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"CoSpaDi replaces low-rank factorization with a sparse dictionary model that better preserves LLM accuracy at 20-40 percent compression.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"f7e2648c9a5527e0cfc6f00a30dade4184bacec6357a9e3215bb0202c79939f6"},"source":{"id":"2509.22075","kind":"arxiv","version":6},"verdict":{"id":"5e9533e4-fa16-4a29-9e32-8f1fd8b16999","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-18T13:14:45.608935Z","strongest_claim":"Across Llama and Qwen model families, CoSpaDi consistently improves the accuracy-compression and perplexity-compression trade-offs over state-of-the-art SVD-based baselines and strong structured pruning baselines at 20-40% compression ratios.","one_line_summary":"CoSpaDi introduces a training-free sparse dictionary learning framework for post-training LLM compression that optimizes functional reconstruction error via activation-derived orthonormalization and achieves improved accuracy-compression trade-offs over SVD and pruning baselines.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The assumption that minimizing functional reconstruction error on a small calibration set will produce a factorization whose downstream task performance remains close to the original model without any fine-tuning or further adaptation.","pith_extraction_headline":"CoSpaDi replaces low-rank factorization with a sparse dictionary model that better preserves LLM accuracy at 20-40 percent compression."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2509.22075/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":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}