AI Paper Atlas

Research map / Architectures and efficiency / Quantization

FPTQuant: Function-Preserving Transforms for LLM Quantization

Boris van Breugel, Yelysei Bondarenko, Paul N. Whatmough, Markus Nagel

ICML 2026 · Architectures and efficiency › Quantization · approach: Post-training quantization of LLMs · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add per-token residual scaling Sn to the other transforms (ablation)FPTQuant, Llama-3.2 3B it W4A4KV4 · WikiText-2perplexity: 12.78 -> 12.57Improved
add pre-RoPE transform Tk (ablation)FPTQuant, Llama-3.2 3B it W4A4KV4 · 5-shot MMLUaccuracy: 35.46 -> 38.33Improved
add multihead value transform Tv (ablation)FPTQuant, Llama-3.2 3B it W4A4KV4 · WikiText-2perplexity: 12.78 -> 11.84Improved
use all three novel transforms together instead of baseline transforms only (ablation)FPTQuant, Llama-3.2 3B it W4A4KV4 · 5-shot MMLUaccuracy: 35.46 -> 44.64Improved
apply dynamic per-token scaling Sn inside FFN (ablation)Llama-3.2 3B it first blocks · FFN intermediate activation kurtosiskurtosis: 2278 -> 544 (-76.1%)Improved
FPTQuant vs SpinQuant and FlatQuant in the hardest static setting (all attention/MLP activations quantized)Llama-3.2 3B it W4A4KV4 · WikiText-2 setting (iii)perplexity: FlatQuant 18.60 -> FPTQuant 17.17Improved
quantize queries and keys to 4 bit with the mergeable Pre-RoPE transform instead of non-mergeable transformsFPTQuant vs FlatQuant · WikiText-2 setting (ii), W4A4KV4perplexity: FlatQuant 12.30 vs FPTQuant 13.99Hurt
FPTQuant instead of FlatQuant on very challenging W4A4KV4 static quantizationLlama-3 8B · zero-shot CSR avgaccuracy: FlatQuant 61.43 vs FPTQuant 52.96Hurt
use FPTQuant mergeable transforms instead of FlatQuant at W4A8KV8Llama-2 7B static · WikiText-2perplexity: FlatQuant 6.46 -> FPTQuant 5.85Improved
use student-teacher JSD loss instead of next-token prediction loss for end-to-end training (ablation)FPTQuant and baselines · generalizationgeneralizationImproved
add locally optimized (Lp norm) transform initialization before end-to-end training (ablation)FPTQuant · end-to-end trainingquantization errorImproved
compare FPTQuant with FlatQuant speed on static INT4 prefillFPTQuant · single transformer block INT4 prefillspeedup: 2.8-3.9x over FP16; 15-29% faster than FlatQuantImproved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy