AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Prototype Transformer: Towards Language Model Architectures Interpretable by Design

Yordan Yordanov, Matteo Forasassi, Bayar Menzat, Ruizhi Wang, Chang Qi, Markus Kaltenberger, Amine M'Charrak, Tommaso Salvatori, Thomas Lukasiewicz

ICML 2026 · Architectures and efficiency › Attention and transformers · approach: Efficient language transformers · 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 self-attention with prototype mixer (ProtoT)LLaMA-3 style transformer · FineWeb-Edu large-scaletest perplexity: 25.8 -> 29.5Hurt
use ProtoT instead of DeltaNet linear attentionlinear-compute LM · FineWeb-Edu large-scaletest perplexity: 31.5 -> 29.5Improved
increase context length from 256 to 2048 with fixed model size (ablation)ProtoT · FineWeb-Edu default sizetest perplexity: 80.7 (1024) -> 83.0 (2048)Mixed
increase hidden size to 512 (ablation)ProtoT · FineWeb-Edu ctx=2048test perplexity: 83.0 -> 70.2Improved
increase layers to 12 (ablation)ProtoT · FineWeb-Edu ctx=2048test perplexity: 83.0 -> 71.8Improved
increase prototypes from 32 to 64 (ablation)ProtoT · FineWeb-Edu ctx=2048test perplexity: 83.0 -> 76.1Improved
use ProtoT instead of LLaMA self-attention for generationlarge-scale LM · LLM-judge text generationElo: 975.18 -> 1021.24Improved
use ProtoT instead of LLaMA for GLUE fine-tuninglarge-scale LM · GLUE averageGLUE score: 71.6 -> 67.6Hurt
use prototype write gate features instead of LLaMA SAE featuresinterpretability analysis · LLM-judged disentanglementdisentanglement score: 5.91 -> 6.52Improved
disrupt female prototype L9 P7 by re-initialization (ablation)ProtoT · women word probabilitydelta probability: -16.60%Hurt
disrupt control prototype L9 P2 (ablation)ProtoT · women word probabilitydelta probabilityNo clear effect
prototype interventions across concepts (ablation)ProtoT · FineWeb validationperplexity change: Max +5.58% (COVID)No 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