SynAdapt: Learning Adaptive Reasoning in Large Language Models via Synthetic Continuous Chain-of-Thought

ACL ARR 2026 January Submission1267 Authors

Keywords: Large Language Models, Efficient Reasoning, Implicit Reasoning

Abstract:

While Chain-of-Thought (CoT) reasoning improves model performance, it incurs significant time costs due to the generation of discrete CoT tokens (DCoT).
Continuous CoT (CCoT) offers a more efficient alternative, but existing CCoT methods are hindered by indirect fine-tuning, limited alignment, or inconsistent targets.
To address these limitations, we propose ***SynAdapt***, an efficient reasoning framework that generates **Synthetic CCoTs** as alignment targets, enabling more effective CCoT learning while reducing generation length.
Moreover, our *SynAdapt* introduces a difficulty classifier that leverages CCoTs to identify hard questions for adaptive re-thinking, achieving improved performance.
Extensive experimental results across various benchmarks from different difficulty levels strongly demonstrate the effectiveness of our method, achieving the best accuracy-efficiency trade-off.

Paper Type: Long

Research Area: LLM Efficiency

Research Area Keywords: LLM Efficiency, efficient models

Contribution Types: NLP engineering experiment, Approaches low compute settings-efficiency

Languages Studied: English

Submission Number: 1267