{"paper":{"title":"ReSpinQuant: Efficient Layer-Wise LLM Quantization via Subspace Residual Rotation Approximation","license":"http://creativecommons.org/licenses/by/4.0/","headline":"ReSpinQuant approximates per-layer rotation matrices with residual subspaces so that layer-wise LLM quantization accuracy can be obtained at the speed of global rotation methods.","cross_cats":["cs.AI"],"primary_cat":"cs.CV","authors_text":"Hyeonjin Kim, Hyunho Lee, Kyomin Hwang, Nojun Kwak, Sunghyun Wee, Suyoung Kim","submitted_at":"2026-04-13T07:00:26Z","abstract_excerpt":"Rotation-based Post-Training Quantization (PTQ) has emerged as a promising solution for mitigating activation outliers in the quantization of Large Language Models (LLMs). Global rotation methods achieve inference efficiency by fusing activation rotations into attention and FFN blocks, but suffer from limited expressivity as they are constrained to use a single learnable rotation matrix across all layers. To tackle this, layer-wise transformation methods emerged, achieving superior accuracy through localized adaptation. However, layer-wise methods cannot fuse activation rotation matrices into "},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"ReSpinQuant resolves such overhead by leveraging offline activation rotation fusion and matching basis using efficient residual subspace rotation. This design reconciles the high expressivity of layer-wise adaptation with only negligible inference overhead.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the residual subspace rotation approximation can capture enough of the expressivity of full per-layer transformations to match their accuracy while still permitting complete offline fusion into the model weights.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"ReSpinQuant achieves state-of-the-art accuracy in W4A4 and W3A3 LLM quantization by using efficient residual subspace rotation approximations that match layer-wise performance while retaining the inference speed of global rotation methods.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"ReSpinQuant approximates per-layer rotation matrices with residual subspaces so that layer-wise LLM quantization accuracy can be obtained at the speed of global rotation methods.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"00a25348b55197005709aa947a45c65d0e162aa27de3755e695bc3e015a0283e"},"source":{"id":"2604.11080","kind":"arxiv","version":2},"verdict":{"id":"0ed50ecb-b7c1-4d3b-a054-f171e4b552a7","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-10T16:12:12.957322Z","strongest_claim":"ReSpinQuant resolves such overhead by leveraging offline activation rotation fusion and matching basis using efficient residual subspace rotation. This design reconciles the high expressivity of layer-wise adaptation with only negligible inference overhead.","one_line_summary":"ReSpinQuant achieves state-of-the-art accuracy in W4A4 and W3A3 LLM quantization by using efficient residual subspace rotation approximations that match layer-wise performance while retaining the inference speed of global rotation methods.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the residual subspace rotation approximation can capture enough of the expressivity of full per-layer transformations to match their accuracy while still permitting complete offline fusion into the model weights.","pith_extraction_headline":"ReSpinQuant approximates per-layer rotation matrices with residual subspaces so that layer-wise LLM quantization accuracy can be obtained at the speed of global rotation methods."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2604.11080/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"}