# Reasoning Bias of Next Token Prediction Training

arXiv: https://arxiv.org/abs/2502.02007 · HTML: https://mlsanity.com/paper/2502.02007/

## Key facts

- Cited by 3 papers in the map (data through 14 Oct 2026).
- Placed in Test-time compute and efficient thinking, one of 17 directions in LLM reasoning.
- Builds on 22 of its 67 references that are in the map.
- First posted on arXiv 4 Feb 2025.

## Placed in

- LLM reasoning / [Test-time compute and efficient thinking](https://mlsanity.com/lr/test-time-compute-and-efficient-thinking/)

## Builds on (most cited first, 22 of 22)

- [Training Large Language Models to Reason in a Continuous Latent Space](https://mlsanity.com/paper/2412.06769/) (2024)
- [Phi-4 Technical Report](https://mlsanity.com/paper/2412.08905/) (2024)
- [Reducing Transformer Depth on Demand with Structured Dropout](https://mlsanity.com/paper/1909.11556/) (2019)
- [Towards AI-Complete Question Answering: A Set of Prerequisite Toy Tasks](https://mlsanity.com/paper/1502.05698/) (2015)
- [From Explicit CoT to Implicit CoT: Learning to Internalize CoT Step by Step](https://mlsanity.com/paper/2405.14838/) (2024)
- [Rho-1: Not All Tokens Are What You Need](https://mlsanity.com/paper/2404.07965/) (2024)
- [InternLM-Math: Open Math Large Language Models Toward Verifiable Reasoning](https://mlsanity.com/paper/2402.06332/) (2024)
- [NEFTune: Noisy Embeddings Improve Instruction Finetuning](https://mlsanity.com/paper/2310.05914/) (2023)
- [What Matters in Transformers? Not All Attention is Needed](https://mlsanity.com/paper/2406.15786/) (2024)
- [Power-law escape rate of SGD](https://mlsanity.com/paper/2105.09557/) (2021)
- [DropAttention: A Regularization Method for Fully-Connected Self-Attention Networks](https://mlsanity.com/paper/1907.11065/) (2019)
- [Initialization is Critical to Whether Transformers Fit Composite Functions by Reasoning or Memorizing](https://mlsanity.com/paper/2405.05409/) (2024)
- [Multi-Step Deductive Reasoning Over Natural Language: An Empirical Study on Out-of-Distribution Generalisation](https://mlsanity.com/paper/2207.14000/) (2022)
- [A Law of Next-Token Prediction in Large Language Models](https://mlsanity.com/paper/2408.13442/) (2024)
- [LogicInference: A New Dataset for Teaching Logical Inference to seq2seq Models](https://mlsanity.com/paper/2203.15099/) (2022)
- [A Diffusion Theory For Deep Learning Dynamics: Stochastic Gradient Descent Exponentially Favors Flat Minima](https://mlsanity.com/paper/2002.03495/) (2020)
- [UniDrop: A Simple yet Effective Technique to Improve Transformer without Extra Cost](https://mlsanity.com/paper/2104.04946/) (2021)
- [Stochastic Modified Equations and Dynamics of Dropout Algorithm](https://mlsanity.com/paper/2305.15850/) (2023)
- [Anchor function: a type of benchmark functions for studying language models](https://mlsanity.com/paper/2401.08309/) (2024)
- [Next-token prediction capacity: general upper bounds and a lower bound for transformers](https://mlsanity.com/paper/2405.13718/) (2024)
- [DropKey](https://mlsanity.com/paper/2208.02646/) (2022)
- [Pretrained Transformers Do not Always Improve Robustness](https://mlsanity.com/paper/2210.07663/) (2022)

## Built on by (most cited first, 3 of 3)

- [Language Models, Graph Searching, and Supervision Adulteration: When More Supervision is Less and How to Make More More](https://mlsanity.com/paper/2503.10542/) (2025)
- [On Next-Token Prediction in LLMs: How End Goals Determine the Consistency of Decoding Algorithms](https://mlsanity.com/paper/2505.11183/) (2025)
- [d-OPD: Future-Aware On-Policy Distillation for Block Diffusion Language Models](https://mlsanity.com/paper/2609.35362/) (2026)

## Abstract (arXiv)

Since the inception of Large Language Models (LLMs), the quest to efficiently train them for superior reasoning capabilities has been a pivotal challenge. The dominant training paradigm for LLMs is based on next token prediction (NTP). Alternative methodologies, called Critical Token Prediction (CTP), focused exclusively on specific critical tokens (such as the answer in Q&A dataset), aiming to reduce the overfitting of extraneous information and noise. Contrary to initial assumptions, our research reveals that despite NTP's exposure to noise during training, it surpasses CTP in reasoning ability. We attribute this counterintuitive outcome to the regularizing influence of noise on the training dynamics. Our empirical analysis shows that NTP-trained models exhibit enhanced generalization and robustness across various benchmark reasoning datasets, demonstrating greater resilience to perturbations and achieving flatter loss minima. These findings illuminate that NTP is instrumental in fostering reasoning abilities during pretraining, whereas CTP is more effective for finetuning, thereby enriching our comprehension of optimal training strategies in LLM development.