A new review says there is no single optimality claim that applies across all large language model quantization regimes. Under fixed total energy, its allocation-flexible geometric-mean coding surrogate, a simplified measure of distortion, favors concentrating energy within a group. A shared-scale AbsMax surrogate, using a peak-dependent scale for that group, instead favors flattening the energy within it. The review presents that split as the paper's central inversion.
The document is an arXiv version 1 preprint dated 25 August 2026. It surveys 200 works through a June 2026 cutoff and classifies 43 transform methods by structure, data awareness, search versus construction, and runtime cost. Published measurements were transcribed verbatim and re-verified against cited sources. The mechanism figures and the authors' own table computations were labeled illustrations of the theory rather than evaluations.
The same transform can favor opposite goals
That distinction puts the KLT result in its proper box. The KLT minimizes distortion among orthogonal transforms under a positive-definite Gaussian source, high-rate quantization, logarithmic per-coordinate allocation and a positive rate for every coordinate. The claim is precise but narrow. For a generic spectrum, the review says the KLT has no shared-scale optimality guarantee.
For the uniform-grid fixed-rate surrogate, the result reverses: flattening energy within each shared-scale group minimizes the objective, and concentration cannot lower it. The review presents a single random rotation as a data-free way to flatten a vector even when the locations of outliers are unknown. It describes this as a method for handling unknown outliers, not as a promise that every group can be made exactly flat.
Rounding adds another constraint. Error-feedback rounding can reduce coupled rounding error, but it cannot by itself remove the dynamic-range wall imposed by a shared peak-dependent scale. That makes rounding and transforms complementary in the paper's framework, with possible overlap rather than a substitute for controlling the shared peak.
WUSH gets a more limited mathematical endorsement. Under its stochastic per-group AbsMax model, WUSH is optimal for FP-AbsMax and near-optimal for INT-AbsMax on zero-mean Gaussian or Laplacian data among invertible block-diagonal transforms. For arbitrary distributions, the paper gives a weaker factor-d guarantee. The review does not extend that model-specific result to the realized deployed objective or to all formats and distributions.
Number formats can reverse the result
The format can change even the direction of a reported comparison. On a tested Llama-3-8B layer, the single-random-Hadamard condition had INT8 reconstruction error lower by about 1.6 effective bits, but FP8 reconstruction error higher by about 0.2. The review identifies this as an ideal-FP analysis tied to that source and matrix, rather than a universal rule for floating-point deployment.
A separate author-computed exercise used mean SQNR, or signal-to-quantization-noise ratio, for a 4-bit AbsMax quantizer. The authors computed it on 12,288 real 128-channel activation groups derived from 768 training tokens, with 16 groups per token, comparing no rotation with a per-group randomized Hadamard. Mean SQNR ran from 13.1 to 19.3 dB for INT4, a 6.2 dB gain, and from 16.6 to 19.4 dB for FP4, a 2.8 dB gain. The survey labels these calculations illustrations rather than evaluations.
An MXFP4 comparison further complicates any simple rule. Block-diagonal Hadamard helped under both rounders, while fixed global QuaRot was 6.4 points below plain per-block RTN and 7.7 points below GPTQ. The contrast was measured but left unexplained in the survey.
The survey also reports a near-substitute relationship between rotation and per-coordinate allocation in one high-rate setting. On real Llama-3-8B covariances, GPTQ was within about 0.1 bit of WaterSIC under entropy-coded high-rate accounting. The review cautions that the margin was not inherited by a fixed-length deployed kernel, so the comparison does not establish worst-case equivalence.
A framework, not a universal recipe
The practical message is therefore narrower than a universal recipe. The flatness theorem is for a uniform grid, KLT optimality assumes its Gaussian high-rate setting, WUSH's claim is tied to its stochastic AbsMax model, and the activation calculations are illustrations rather than deployment evaluations. The review identifies open questions including a guarantee for the true deployed extreme-value objective, joint optimization of transforms and error-feedback rounding, and a quantitative explanation of the MXFP4 global-versus-block-confined gap.
Paper data and sources
Original title: Transforms for LLM Quantization: The Great Inversion and Format Co-Design
Authors: Ehsan Jokar
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text