AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Predicting the Order of Upcoming Tokens Improves Language Modeling

Zayd Muhammad Kawakibi Zuhri, Erland Hilman Fuadi, Alham Fikri Aji

ICML 2026 · Architectures and efficiency › Attention and transformers · approach: Language model pretraining and induction

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 TOP auxiliary loss to NTP340M-7B LMs · 9 standard NLP benchmarksaccuracy / perplexity: 7B Lambada acc 55.89 -> 57.03Improved
add MTP auxiliary heads to NTP7B LM · standard NLP benchmarksaccuracy: 7B Lambada acc 55.89 -> 53.13Hurt
use DS-MTP instead of MTPLMs · standard NLP benchmarksaccuracyNo clear effect
add TOP loss then continue training on math1.8B LM · GSM8Kaccuracy (%): 39.20 -> 45.64Improved
add TOP loss then continue training on math7B LM · MATHaccuracy (%): 21.74 -> 20.40Hurt
add DS-MTP then continue training on math/code1.8B LM · GSM8Kaccuracy (%): 39.20 -> 2.65Hurt
add TOP loss on code continued training7B LM · MBPP pass@64accuracy (%): 49.20 -> 54.00Improved
add TOP loss on star graph pathfinding8-layer transformer · star graph G(5,5)test accuracy (%): 0.1 -> 100Improved
use more MTP heads (4 instead of 2) (ablation)MTP · star graph G(3,5)test accuracy (%): 59.0 -> 100Improved
use TOP head for self-speculative decoding vs MTP headsTOP · self-speculative decodingaccepted tokens per pass: 7B math: MTP 2.49, TOP 1.55Hurt
vary TOP window size from 4 to 4096 (ablation)TOP 340M · Lambada accuracyaccuracy (%): 36.35 (NTP) -> 36.95-38.68Improved
raise TOP loss weight alpha toward 0.9 (ablation)TOP 340M · Lambada accuracyaccuracy (%): 37.07 (alpha 0.5) -> 38.39 (alpha 0.9)Improved

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