AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Learned Meta-Tokens for Language Modeling

Alok Shah, Khush Gupta, Keshav Ramji, Pratik Chaudhari

ICLR 2026 · Architectures and efficiency › Attention and transformers · approach: Attention head mechanisms

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 meta-tokens and meta-attention in pretrainingGPT-2 124M with RoPE · PG19perplexity: 16.13 -> 14.79Improved
add meta-tokens and meta-attentionGPT-2 APE · List Recall (eval 512)accuracy: 3.6 -> 98.7Improved
add meta-tokens and meta-attentionGPT-2 APE · Segment Counting (eval 512)accuracy: 25.0 -> 80.9Improved
add meta-tokens and meta-attentionGPT-2 APE · Copying (eval 512)accuracy: 7.8 -> 98.5Improved
use RoPE instead of APE with meta-tokensMeta-attention model · Segment Counting (eval 1024)accuracy: 33.3 -> 66.7Improved
increase fine-tune length from 4k to 8k (YaRN)Meta-attention model · List Recall at 8kaccuracy: 3.5 -> 100.0Improved
zero positional encoding at meta-tokens (ablation)Meta + APE (train 128) · List Recallaccuracy: 100.0 -> 99.3No clear effect
zero positional encoding at meta-tokens (ablation)Meta + APE (train 512) · List Pointer extra-hardaccuracy: 11.1% -> 50.0%Improved
zero positional encoding at meta-tokens (ablation)Meta + APE (train 128) · Segment Countingaccuracy: 77.4 -> 63.1Hurt
zero positional encoding at meta-tokens (ablation)Meta + APE (train 256) · Segment Countingaccuracy: 83.3 -> 88.1Improved
keep token embedding at meta-tokens (versus zeroing it) (ablation)Meta + APE (train 128) · List Recallaccuracy: 17.4 -> 100.0Improved
keep token embedding at meta-tokens (versus zeroing it) (ablation)Meta + RoPE (train 128) · Copyingaccuracy: 6.9 -> 100.0Improved

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