SpinQuant: LLM Quantization with Learned Rotations

Random rotations remove activation outliers (QuaRot) but the quality you get depends heavily on WHICH random rotation you draw - some draws are far better than others.

Same computational-invariance trick as QuaRot, but the rotation is LEARNED rather than drawn at random. Searches over rotation parameterizations (Cayley-optimized orthogonal transform on the GPU/Stiefel manifold) that are identity-equivalent at full precision, optimized with gradient descent on the quantization-aware objective so as to minimize downstream quantized accuracy loss. Also introduces CPTQ (progressive end-to-end weight quantization from 16-bit down to the target width).

Paper's own claim (abstract, W4 weights + activations + KV cache): narrows the zero-shot reasoning accuracy gap to full precision to 2.9 points on LLaMA-2 7B, surpassing LLM-QAT by 19.1 points and SmoothQuant by 25.0 points at the same precision; on hard-to-quantize LLaMA-3 8B it cuts the gap to full precision by up to 45.1% relative to QuaRot. COMPRESSION RATIO: 4-bit on weights, activations and KV cache (~4x each vs FP16). QUALITY COST: 2.9 points of zero-shot accuracy - the best number in the 4-bit PTQ literature at the time, and still a 2.9-point regression. Also reports up to a 13-point spread between good and bad random rotations, which is the paper's real argument.

SLUG RENAMED 2026-10-04 (parent agent): was 'spinquant', colliding with the quantization record of the same id, which made every bare-slug cross-reference ambiguous. Inbound refs (['format_id', 'scheme', 'supported_formats']) all mean the quantization record, so none needed repointing; papers are cited via `sources`, never by bare slug. arXiv ID 2405.16406 verified via /abs/; v4 2025-02-20. Venue ICLR 2025 confirmed by the arXiv comment field and Semantic Scholar. Year = venue year 2025 (v1 was 2024-05-26). adoption=research-only: the rotation-learning step costs real gradient-descent calibration time, which is why it has not displaced the cheaper random-rotation recipes in serving. arXiv ID and title verified by fetching https://arxiv.org/abs/2405.16406 and reading the citation_title meta tag on 2026-10-03. Venue and year from the Semantic Scholar graph API batch endpoint (paper/arXiv:2405.16406) unless the record's notes say otherwise. VENUE_TRACK BACKFILL 2026-10-04 (paper-fields pass, agent w3p): venue_track 'main'. This is STRUCTURAL rather than researched: iclr has no Findings volume and no workshop track in which a paper could appear, so 'main' is the only possible honest value and no per-paper evidence was needed. The field is annotated here so that 'findings' and 'workshop' remain MEANINGFUL - a filter for findings=true must return only the papers that actually are findings, which requires every other paper to be explicitly NOT findings.