AI Paper Atlas

Research map / Architectures and efficiency / Quantization

LLaVA-FA: Learning Fourier Approximation for Compressing Large Multimodal Models

Pengcheng Zheng, Chaoning Zhang, Jiarong Mo, GuoHui Li, Jiaquan Zhang, Jiahao Zhang, Sihan Cao, Sheng Zheng, Caiyan Qin, Guoqing Wang, Yang Yang

ICLR 2026 · Architectures and efficiency › Quantization · approach: KV cache quantization · 15 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
raise LoRA rank from 64 to 256 (ablation)LLaVA-FA-2B · Multimodal suite averageaverage accuracy: 58.9 -> 60.8 (+1.9%)Improved
raise LoRA rank from 64 to 256 (ablation)LLaVA-FA-1B · Multimodal suite averageaverage accuracy: 55.3 -> 56.6 (+1.3%)Improved
raise amplitude and phase bit-widths from 2 to 4 (ablation)LLaVA-FA-2B · Multimodal suite averageaverage accuracy: 56.3 -> 60.9Improved
raise amplitude and phase bit-widths from 2 to 4 (ablation)LLaVA-FA-2B · Object HalBenchhallucination rate: 18.7 -> 11.2Improved
raise bit-widths from 4 to 6 each (ablation)LLaVA-FA-2B · Multimodal suite averageaverage accuracy: 60.9 -> 61.2No clear effect
raise bit-widths from 4 to 6 each (compression cost) (ablation)LLaVA-FA-2B · compression ratiocompression ratio: 11.2 -> 8.9Hurt
use Fourier-domain complex low-rank approximation instead of spatial real (ablation)Qwen-2.5-7B query layers · Spectrum-decay experimentFrobenius error at rank 256: 0.031 -> 0.018Improved
use Fourier-domain complex low-rank approximation instead of spatial real (ablation)Qwen-2.5-7B query layers · Spectrum-decay experimentmin rank for 1% error: 312 -> 217Improved
compress with Fourier approximation versus Imp-2B, FLOPs and TTFTImp-2B · 576 image tokensFLOPs (T): 2.14 -> 1.93Improved
compress with Fourier approximation versus Imp-2B, latencyImp-2B · 576 image tokenslatency (ms): 36.2 -> 30.6Improved
compare LLaVA-FA-2B with the larger LLaVA-1.5-7B on average accuracyLLaVA-FA-2B · Multimodal suite averageaverage accuracy: 62.1 vs 60.8Mixed
compare LLaVA-FA-1B with SPHINX-Tiny using 5M instead of 15M samplesSPHINX-Tiny · Multimodal suite averageaverage accuracy: 49.4 vs 49.2No clear effect

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