Long Chain-of-Thought Fine-tuning via Understanding-to-Reasoning Transition

Chenxin An, Zhihui Xie, Xiaonan Li, Ming Zhong, Shansan Gong, Lei Li, Jun Zhang, Jingjing Xu, Lingpeng Kong2025 · DOI

Abstract

Researchers have introduced Understanding-to-Reasoning Transition (URT), a fine-tuning framework that teaches Large Language Models (LLMs) to generate long "Chain-of-Thought" (CoT) reasoning by first teaching them to understand it. By gradually shifting the model's task from interpreting provided reasoning steps to generating them from scratch, URT significantly boosts mathematical reasoning performance while preserving the model's ability to handle long-context inputs—a skill often lost in traditional reasoning-model training.

The Reasoning-Context Paradox

The recent "o1" era of AI has shifted focus toward models that "think" before they speak, generating thousands of tokens of internal reasoning to solve complex problems. However, this progress has encountered two major obstacles:

  1. The Reasoning-Context Paradox: Models trained specifically to generate long reasoning chains often become significantly worse at understanding long inputs. Because they are typically trained on short prompts paired with long outputs, they lose their "long-context" comprehension skills.
  2. The Learning Gap: Jumping straight from a short prompt to a 10,000-token reasoning chain is a massive leap. Standard Supervised Fine-Tuning (SFT) often fails to bridge the gap between simple pattern matching and complex, multi-step logical deduction.

To address this, the authors developed a training method that leverages an LLM's existing long-context understanding to "bootstrap" its reasoning capabilities.

Methods: The URT Framework

The core innovation is URT (Understanding-to-Reasoning Transition), a sequence-level curriculum learning framework. Instead of immediately asking the model to generate a full reasoning trace $y$ from a prompt $x$, the process is decomposed into a transition from "understanding" to "reasoning."

The Curriculum Shift

The framework divides a long reasoning chain into a prefix ($y_{1:t}$) and a suffix ($y_{t+1:T}$):

  • The Understanding Phase: Early in training, the model is given a large portion of the reasoning chain in the prompt. It only needs to generate the final steps to reach the answer.
  • The Reasoning Phase: As training progresses, the amount of reasoning provided in the prompt is gradually reduced. This forces the model to generate more of the logical chain itself.
Figure 4: Accuracy improves significantly as more reasoning steps are provided in the input, justifying the "understanding" phase as a foundation for reasoning.

This acts as "training wheels" for logic. By exposing the model to diverse prompt lengths as reasoning steps move from input to output, the model naturally maintains its long-context comprehension while learning to reason.

Experimental Setup and Datasets

The researchers utilized several key datasets and benchmarks to validate URT:

  • Training Data: The STILL2 dataset (approx. 4.9K problems), containing long CoT traces distilled from teacher models like QwQ-32B-preview.
  • Reasoning Benchmarks: Standard evaluations including MATH500, AIME24 (Olympiad-level math), and GPQA Diamond (PhD-level science).
  • Long-Context Benchmarks: The RULER benchmark, which tests "needle-in-a-haystack" and variable tracing tasks.
  • CoT-U (Chain-of-Thought Understanding): A new benchmark introduced by the authors to evaluate if a model can actually use a provided, human-verified reasoning path to reach a correct answer, rather than ignoring the context.

Key Results and Insights

The URT-SFT method consistently outperformed standard SFT across all model sizes (8B, 14B, and 32B).

Figure 2: Performance comparison across multiple benchmarks shows URT-trained models (like Qwen2.5-32B) competitive with or exceeding state-of-the-art reasoning models.

Performance Highlights

  • AIME24: The Qwen2.5-32B model trained with URT achieved 56.7% accuracy, surpassing its own teacher model (QwQ-32B-preview at 50.0%) and standard SFT (43.3%).
  • GPQA Diamond: The 32B model reached 61.6%, a significant jump from the base model's 43.4%.
  • Long-Context Preservation: On the RULER benchmark (32K context), the URT-trained Qwen model scored 81.5, significantly higher than the standard reasoning variant (68.9). This proves that URT protects the model's ability to "read" while it learns to "think."

Understanding Precedes Generation

A critical finding was that a model's ability to solve a problem is closely tied to how many reasoning steps it is given. By starting with "understanding" tasks, the model learns the structure of a good argument before it has to invent one. Furthermore, because URT moves tokens from output to input, it naturally creates a training set with varied prompt lengths, preventing the model from overfitting to the "short prompt, long response" pattern.

Conclusion and Future Work

URT demonstrates that long-form reasoning and long-context understanding are not mutually exclusive. While the method shows high training efficiency and is orthogonal to Reinforcement Learning (RL), some limitations remain. Performance still declines when inputs exceed 64K tokens, and the current study is heavily weighted toward mathematical reasoning. Future work will focus on expanding high-quality, non-mathematical long-prompt data to help models master broader forms of logical deduction.

Summary updated

BibTeX
@misc{an2025long,
  author = {Chenxin An and Zhihui Xie and Xiaonan Li and Ming Zhong and Shansan Gong and Lei Li and Jun Zhang and Jingjing Xu and Lingpeng Kong},
  title = {{Long Chain-of-Thought Fine-tuning via Understanding-to-Reasoning Transition}},
  year = {2025},
  doi = {10.18653/v1/2025.emnlp-main.1751},
}
Figure