AI Paper Atlas

Research map / Architectures and efficiency / Quantization

DictLLM: Post-training Compression of Large Language Model with Dictionary Kernels

Jinho Park, Se Young Chun, Mingoo Seok

NeurIPS 2026 · Architectures and efficiency › Quantization · approach: KV cache quantization · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes DictLLM, a post-training LLM compression method that decomposes weight matrices within grouped layers into a shared dictionary of kernels for mobile deployment.

Problem

Deploying LLMs on mobile devices under compute and memory constraints; post-training compression is needed to trade off performance against resource use.

Approach

Decomposes weight matrices within grouped layers into a shared dictionary of kernels (dictionary-based compression), applied post-training without retraining.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use DictLLM instead of Basis Sharing at 2x compressionBasis Sharing · LLaMA-7B WikiText-2perplexity: 19.99 -> 10.59Improved
prune only the coefficients, keep dictionary dense (ablation)DictLLM · LLaMA-7B WikiText-2 at 1.43xperplexity: 9.25 -> 7.83Improved
prune the dictionary instead of coefficients (ablation)DictLLM · LLaMA-7B WikiText-2 at 1.43xperplexity: 7.83 -> 9.04Hurt
use dictionary-aware Hessian lazy-update pruning instead of magnitude pruning (ablation)DictLLM · LLaMA-7B WikiText-2perplexity: 47.43 -> 9.89Improved
use lazy update instead of full Hessian inversion (ablation)DictLLM · LLaMA-7B WikiText-2pruning time: 13.49 h -> 55.6 min, same perplexity 9.89Improved
add dictionary refit after coefficient pruning (ablation)DictLLM · LLaMA-7B WikiText-2perplexity: 11.53 -> 10.59Improved
add dictionary refit at higher sparsity (ablation)DictLLM · LLaMA-7B WikiText-2perplexity reduction: 1.04 PPL at 50% sparsity vs 0.06 PPL at 20%Improved
quantize both dictionary and coefficients to NVFP4 versus NVFP4/INT8 (ablation)DictLLM · LLaMA-7B WikiText-2compression gain and perplexity: 16.05 -> 16.82 ppl; 2.60x -> 3.56xMixed
add NVFP4 quantization after factorization, pruning and refit (ablation)DictLLM · LLaMA-7Baverage zero-shot accuracy: 0.43 -> 0.40; 2x -> 7x compressionHurt
add post-compression LoRA fine-tuningDictLLM · LLaMA-7B WikiText-2perplexity: up to 1.9 points lowerImproved
increase calibration set size from 32 to 256 (ablation)DictLLM and Basis Sharing · LLaMA-7B WikiText-2 at 1.43xperplexityImproved

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