AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

NITP: Next Implicit Token Prediction for LLM Pre-training

Xiangdong Zhang, Debing Zhang, Shaofeng Zhang, Xiaohan Qin, Yu Cheng, Junchi Yan

ICML 2026 · Architectures and efficiency › RNNs and normalization · approach: Layer-wise analysis of language models · 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
add NITP auxiliary loss to NTP pre-training1.9bA0.3b MoE · 13 benchmarksaverage score: 26.65 -> 27.46Improved
add NITP auxiliary loss to NTP pre-training3bA0.5b MoE · 13 benchmarksaverage score: 31.37 -> 33.49Improved
add NITP auxiliary loss to NTP pre-training9bA1b MoE · MMLU-Proaccuracy: 15.29 -> 21.00Improved
add NITP auxiliary loss to NTP pre-trainingdense 0.5B-3B · 7 benchmarksaverage score: 31.91 -> 33.70 (2B)Improved
add NITP loss, frozen hidden states3B MoE · MTEB 25 tasksoverall score: 39.24 -> 41.56Improved
add NITP loss (validation loss check)9B MoE · Pile validationcross-entropy: 1.841 vs. 1.840No clear effect
use shallow L4 target instead of deep L14 (ablation)NITP 3B MoE · ablation suiteaverage score: 22.16 -> 23.58Improved
use shallow L4 target instead of middle L8 (ablation)NITP 3B MoE · ablation suiteaverage score: 21.22 -> 23.58Improved
predict next token (t+1) implicit target instead of current step (ablation)NITP 3B MoE · ablation suiteaverage score: 18.75 -> 23.58Improved
use cosine loss instead of MSE (ablation)NITP 3B MoE · ablation suiteaverage score: 19.38 -> 23.58Improved
use cosine loss instead of Smooth L1 (ablation)NITP 3B MoE · ablation suiteaverage score: 23.24 -> 23.58Improved
use cosine loss instead of KL distillation (ablation)NITP 3B MoE · ablation suiteaverage score: 22.88 -> 23.58Improved

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