AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

Efficient Mismatch-Tolerant Coding for Model-Driven Compression

Aviv Adler, Jennifer Tang

ICML 2026 · Architectures and efficiency › RNNs and normalization · approach: Distributed LLM inference systems · 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
replace uniform bins and loose buffers (PMATIC V1) with tight bound, non-uniform bins and message-optimized parameters (V2)PMATIC · Emma, eps=0.02compression ratio: 0.2113 -> 0.0892Improved
replace V1 with V2 on Qwen 2.5 WikipediaPMATIC · Wikipedia, eps=0.02compression ratio: 0.2297 -> 0.1349Improved
replace V1 with V2 on Mistral HamletPMATIC · Hamlet, eps=0.02compression ratio: 0.3447 -> 0.2269Improved
increase robustness eps from 0.02 to 1.0 (ablation)PMATIC with Llama 3.1 · Emmacompression ratio: 0.0893 -> 0.2932Hurt
add PMATIC robustness (eps=0.02) over plain arithmetic codingLlama 3.1 compression · Wikipediacompression ratio: 0.0815 -> 0.1349Hurt
use PMATIC instead of arithmetic coding under real cross-machine non-determinismLLM compression · Real mismatch decoding (M4 vs M2)correctly decoded filesImproved
raise eps from 0.0001 to 0.003 (ablation)PMATIC with Llama 3.1 · Emma token error ratetoken error rate: 0.007758 -> 0.000610Improved
add synthetic uniform noise within [-eps, eps] to decoder logits (ablation)PMATIC · all files, same devicedecoding successNo 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