ARCHead: Activation-Metric Residual Correction for Large Language Model Output Heads

arXiv:2608.02703 · cs.CL, cs.LG · Submitted 2026-08-03 · Read on arXiv

Şuayp Talha Kocabay, Talha Rüzgar Akkuş, Kamer Ali Yuksel

cs.CL, cs.LG

Submitted: 2026-08-03

Comments: 13 pages, 4 figures. Submitted to ACL Rolling Review (ARR). Code: https://github.com/suayptalha/archead

Code: https://github.com/suayptalha/archead

License: http://creativecommons.org/licenses/by/4.0/

The gist: Weight-only quantization substantially reduces the storage of large language model (LLM) transformer blocks, but practical backends often retain the final language-modeling head (LM-head) in BF16 or

Terminology

Abstract

Weight-only quantization substantially reduces the storage of large language model (LLM) transformer blocks, but practical backends often retain the final language-modeling head (LM-head) in BF16 or FP16. Quantizing this projection naively can strongly perturb the vocabulary-logit distribution. We present ARCHead, a packed LM-head compressor that combines a quantized low-rank core, group-wise INT4 residuals, and a low-rank correction fitted in an activation-derived metric. ARCHead stores no dense BF16 head and reduces persistent LM-head storage by 3.7-3.9x. On Qwen3-8B-Base, it uses 25.6% of BF16 head storage while attaining 1.007 relative perplexity; storage-matched naive INT4 yields 1.14-1.16. Replacing the BF16 head left by AWQ or bitsandbytes adds only 0.006-0.007 cross-entropy, with less than 2% throughput change in our measurements. ARCHead therefore complements block quantizers by compressing the large output projection they can leave untouched. Code is available at https://github.com/suayptalha/archead.

Sources

Related papers