Pre-Training Curriculum for Multi-Token Prediction in Language Models (2025.acl-long)
Copied to clipboard
| Challenge: | Multi-token prediction (MTP) is a pre-training objective for language models . prior work has shown that smaller language models struggle with the MTP objective . |
| Approach: | They propose a curriculum learning strategy that uses multiple prediction heads to predict the next tokens at each prediction step. |
| Outcome: | The proposed curriculum improves performance and output quality while retaining the benefits of self-speculative decoding. |
Similar Papers
MTP-RL: Acceleration of Reinforcement Learning Rollouts with Policy-Aligned Multi-Token Prediction (2026.findings-acl)
Copied to clipboard
| Challenge: | Reinforcement learning (RL) is widely applied to boost the performance of pretrained models, yet its training efficiency is severely constrained by rollout generation. |
| Approach: | They propose a framework that accelerates the rollout phase for diverse models by equipping a pipeline to equip the multi-layer parameter-sharing MTP for all models and an advantage-aware MTP optimization strategy. |
| Outcome: | The proposed framework achieves stable growth of acceptance length during RL training, and also accelerates RL rollouts, achieving an average 23.1%–55.3% reduction in rollout time compared to baselines. |
VocalNet: Speech LLMs with Multi-Token Prediction for Faster and High-Quality Generation (2025.emnlp-main)
Copied to clipboard
| Challenge: | Experimental results show VocalNet outperforms existing open-source speech LLMs despite limited training data. |
| Approach: | They propose a scalable and model-agnostic training framework and a novel multi-token prediction paradigm for speech generation. |
| Outcome: | The proposed model outperforms open-source speech LLMs while outperforming existing open-sourced models. |
Toward Consistent World Models with Multi-Token Prediction and Latent Semantic Enhancement (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods to learn internal world models rely on one-step supervision . however, standard MTP suffers from structural hallucinations . |
| Approach: | They propose a method which anchors predictions to ground-truth hidden state trajectories. |
| Outcome: | The proposed method bridges the gap between discrete tokens and continuous state representations, reducing structural hallucinations, and improving robustness to perturbations. |
Beyond Tokens: Concept-Level Training Objectives for LLMs (2026.eacl-short)
Copied to clipboard
| Challenge: | Large language models (LLMs) are trained with a surprisingly narrow objective: predicting the next token in a sequence. |
| Approach: | They propose a shift from token-level to concept-level prediction where concepts group multiple surface forms of the same idea. |
| Outcome: | The proposed model improves on human-level models on diverse NLP benchmarks. |
Beyond Random Sampling: Efficient Language Model Pretraining via Curriculum Learning (2026.eacl-long)
Copied to clipboard
| Challenge: | Curriculum learning has improved efficiency across machine learning domains, but remains underexplored for language model pretraining. |
| Approach: | They present a systematic investigation of curriculum learning in LLM pretraining . they use vanilla curriculum learning, pacing-based sampling, and interleaved curricula . |
| Outcome: | The proposed framework accelerates convergence in early and mid-training phases, reducing training steps by 18-45% to reach baseline performance. |
Sequence-level Large Language Model Training with Contrastive Preference Optimization (2025.findings-naacl)
Copied to clipboard
| Challenge: | a new method to improve the performance of large language models requires a small computational cost. |
| Approach: | They propose a CPO procedure that can inject sequence-level information into the model at any training stage without expensive human labeled data. |
| Outcome: | The proposed objective surpasses the next token prediction in terms of win rate in instruction-following and text generation tasks. |
Token-wise Curriculum Learning for Neural Machine Translation (2021.findings-emnlp)
Copied to clipboard
| Challenge: | Existing curriculum learning approaches to Neural Machine Translation (NMT) require sampling sufficient amounts of “easy” samples from training data at the early stage of training. |
| Approach: | They propose a token-wise curriculum learning approach that creates sufficient amounts of easy samples from training data. |
| Outcome: | The proposed approach outperforms baselines on five language pairs on low-resource languages. |
On-the-fly Cross-lingual Masking for Multilingual Pre-training (2023.acl-long)
Copied to clipboard
| Challenge: | In multilingual pre-training, multilingual models only learn cross-linguality implicitly from isomorphic spaces formed by overlapping different language spaces due to the lack of explicit cross-linguistic forward pass. |
| Approach: | They propose a dynamic token-wise masking scheme for multilingual pre-training that uses a special token [C]x to replace a random token in the input sentence. |
| Outcome: | The proposed model improves the performance of UNMT models on De, Ro, Ne En. |
Reinforcement Learning on Pre-Training Data (2026.acl-long)
Copied to clipboard
Siheng Li, Kejiao Li, Zenan Xu, Guanhua Huang, Kun Li, Haoyuan Wu, null Wujiajia, Zihao Zheng, Chenchen Zhang, Kun Shi, Xue Gong, Qi Yi, Ruibin Xiong, Tingqiang Xu, Yuhao Jiang, Jianfeng Yan, Yuyuan Zeng, Guanghui Xu, Jinbao Xue, Zhijiang xu, Zheng Fang, Shuai LI, Qibin Liu, Xiaoxue Li, Zhuoyu Li, Yangyu Tao, Fei Gao, Cheng Jiang, Bochao Wang, Kai Liu, Jianchen Zhu, Wai Lam, Bo Zhou, Di Wang
| Challenge: | Recent progress in large language models is driven by scaling of training compute through pre-training with nexttoken prediction (NTP) or post-training (RL) Pre-training using NTP enables models to acquire extensive knowledge and skills from general data, but it suffers from data inefficiency and catastrophic forgetting in continual learning settings. |
| Approach: | They propose to scale training compute through pre-training with next-token prediction (NTP) or post-training by scaling reinforcement learning (RL) to improve learning from general data. |
| Outcome: | Experiments on multiple benchmarks and models show that the proposed approach improves continual pre-training and provides a strong foundation for post-training on Qwen3-8B-Base. |
Your Pretrained Model Tells the Difficulty Itself: A Self-Adaptive Curriculum Learning Paradigm for Natural Language Understanding (2025.acl-srw)
Copied to clipboard
| Challenge: | Existing curriculum learning approaches rely on manually defined difficulty metrics which may not accurately reflect the model’s own perspective. |
| Approach: | They propose a self-adaptive curriculum learning paradigm that prioritizes fine-tuning examples based on difficulty scores predicted by pre-trained language models (PLMs) they evaluate four datasets covering binary and multi-class classification tasks. |
| Outcome: | The proposed model leads to faster convergence and improved performance compared to standard random sampling. |