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
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 changed | Where | Measured | Effect |
|---|---|---|---|
| add NITP auxiliary loss to NTP pre-training | 1.9bA0.3b MoE · 13 benchmarks | average score: 26.65 -> 27.46 | Improved |
| add NITP auxiliary loss to NTP pre-training | 3bA0.5b MoE · 13 benchmarks | average score: 31.37 -> 33.49 | Improved |
| add NITP auxiliary loss to NTP pre-training | 9bA1b MoE · MMLU-Pro | accuracy: 15.29 -> 21.00 | Improved |
| add NITP auxiliary loss to NTP pre-training | dense 0.5B-3B · 7 benchmarks | average score: 31.91 -> 33.70 (2B) | Improved |
| add NITP loss, frozen hidden states | 3B MoE · MTEB 25 tasks | overall score: 39.24 -> 41.56 | Improved |
| add NITP loss (validation loss check) | 9B MoE · Pile validation | cross-entropy: 1.841 vs. 1.840 | No clear effect |
| use shallow L4 target instead of deep L14 (ablation) | NITP 3B MoE · ablation suite | average score: 22.16 -> 23.58 | Improved |
| use shallow L4 target instead of middle L8 (ablation) | NITP 3B MoE · ablation suite | average score: 21.22 -> 23.58 | Improved |
| predict next token (t+1) implicit target instead of current step (ablation) | NITP 3B MoE · ablation suite | average score: 18.75 -> 23.58 | Improved |
| use cosine loss instead of MSE (ablation) | NITP 3B MoE · ablation suite | average score: 19.38 -> 23.58 | Improved |
| use cosine loss instead of Smooth L1 (ablation) | NITP 3B MoE · ablation suite | average score: 23.24 -> 23.58 | Improved |
| use cosine loss instead of KL distillation (ablation) | NITP 3B MoE · ablation suite | average score: 22.88 -> 23.58 | Improved |
Closest papers by meaning
- Implicit Optimization Bias of Next-token Prediction in Linear Models (NeurIPS 2024 · similarity 0.90)
- Reinforced Fast Weights via Next-Sequence Prediction (NeurIPS 2026 · similarity 0.89)
- Better & Faster Large Language Models via Multi-token Prediction (ICML 2024 · similarity 0.88)
- On the Generalization Ability of Next-Token-Prediction Pretraining (ICML 2025 · similarity 0.88)
- Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization (ICLR 2025 · similarity 0.88)
- Beyond Next Token Prediction: Patch-Level Training for Large Language Models (ICLR 2025 · similarity 0.88)
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.
