The Imitation Game: Turing Machine Imitator is Length Generalizable Reasoner

TL;DR

This paper introduces TAIL, a method that enhances the length generalization of large language models by imitating Turing Machine processes through synthetic chain-of-thought data, leading to improved reasoning over longer sequences.

Contribution

The paper proposes TAIL, a novel approach that synthesizes Turing Machine-like reasoning data to significantly improve LLMs' length generalization capabilities on diverse algorithmic tasks.

Findings

TAIL outperforms previous methods on synthetic datasets

Qwen2.5-7B shows improved performance with TAIL

Turing Machine concepts are crucial for length generalization

Abstract

Length generalization, the ability to solve problems of longer sequences than those observed during training, poses a core challenge of Transformer-based large language models (LLM). Although existing studies have predominantly focused on data-driven approaches for arithmetic operations and symbolic manipulation tasks, these approaches tend to be task-specific with limited overall performance. To pursue a more general solution, this paper focuses on a broader case of reasoning problems that are computable, i.e., problems that algorithms can solve, thus can be solved by the Turing Machine. From this perspective, this paper proposes Turing MAchine Imitation Learning (TAIL) to improve the length generalization ability of LLMs. TAIL synthesizes chain-of-thoughts (CoT) data that imitate the execution process of a Turing Machine by computer programs, which linearly expands the reasoning steps into atomic states to alleviate shortcut learning and explicit memory fetch mechanism to reduce the difficulties of dynamic and long-range data access in elementary operations. To validate the reliability and universality of TAIL, we construct a challenging synthetic dataset covering 8 classes of algorithms and 18 tasks. Without bells and whistles, TAIL significantly improves the length generalization ability as well as the performance of Qwen2.5-7B on various tasks using only synthetic data, surpassing previous methods and DeepSeek-R1. The experimental results reveal that the key concepts in the Turing Machine, instead of the thinking styles, are indispensable for TAIL for length generalization, through which the model exhibits read-and-write behaviors consistent with the properties of the Turing Machine in their attention layers. This work provides a promising direction for future research in the learning of LLM reasoning from synthetic data.