AI Paper Atlas

Research map / Architectures and efficiency / Quantization

LFQ: Logit-aware Final-block Quantization for Boosting the Generation Quality of Low-Bit Quantized LLMs

Jung Hyun Lee, June Yong Yang, Jungwook Choi, Eunho Yang

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 LFQ (cross-entropy on final block)FlexRound W4 Qwen2.5-7B-Instruct · IFEval greedyaccuracy (%): 69.50 -> 71.35Improved
add LFQBlock-AP W3g128 Qwen2.5-7B-Instruct · IFEval greedyaccuracy (%): 61.00 -> 63.77Improved
add LFQFlexRound W4 L1-Qwen-7B-Max · AIME 2024 greedyaccuracy (%): 30.00 -> 43.33Improved
add LFQFlexRound W4 DeepSeek-R1-Distill-Llama-8B · AIME 2024 greedyaccuracy (%): 16.67 -> 26.67Improved
add LFQFlexRound W4g128 Qwen3-30B-A3B-Instruct-2507 · AIME 2025 pass@1accuracy (%): 57.29 -> 58.75Improved
add LFQOmniQuant W3g128 Qwen2.5-14B-Instruct · MATH500 greedyaccuracy (%): 70.4 -> 69.8Hurt
add LFQFlexRound W4 Qwen2.5-7B-Instruct · MMLUaccuracy (%): 72.50 -> 72.48No clear effect
add LM head to final-block objective (MSE kept) (ablation)FlexRound Llama-3.1-8B-Instruct W4 · IFEval greedyaccuracy (%): 70.24 -> 71.53Improved
switch from MSE to cross-entropy with LM head (ablation)OmniQuant Llama-3.1-8B-Instruct W4 · GSM8K greedyaccuracy (%): 78.32 -> 79.76Improved
switch from MSE to cross-entropy with LM head (ablation)FlexRound Qwen2.5-7B-Instruct W4 · MATH500 greedyaccuracy (%): 71.4 -> 73.4Improved
apply LFQ to top 2 blocks instead of only final block (ablation)FlexRound Llama-3.1-8B-Instruct · (IFEval+GSM8K)/2accuracy (%): 77.13 -> 76.78 (k=2 vs k=3)No clear effect
add LFQ on the final block (vs skipping last block) (ablation)FlexRound k=2 Llama-3.1-8B-Instruct · (IFEval+GSM8K)/2accuracy (%): 76.17 -> 77.13Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 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