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

2025-08-01

AI Summary

Existing continuous chain-of-thought (CCoT) methods improve reasoning efficiency but suffer from inaccurate alignment, objective inconsistency, and limited capability on complex problems. To address these issues, this paper proposes SynAdapt—a novel framework comprising three key components: (1) a synthetic CCoT generation mechanism that directly constructs high-quality, differentiable alignment targets; (2) a difficulty classifier jointly modeling CCoT representations and contextual features to dynamically trigger adaptive rethinking; and (3) a hybrid training strategy integrating discrete CoT distillation with end-to-end fine-tuning for efficient optimization. Evaluated across multi-level reasoning benchmarks, SynAdapt consistently outperforms state-of-the-art CoT and CCoT baselines, achieving superior accuracy–efficiency trade-offs. Ablation studies and cross-dataset experiments further validate its effective alignment, strong generalization, and practical applicability.

Technology Category

• Cognitive Modeling & Cognitive Systems: Adaptive Behavior
• Natural Language Processing: Code Generation / Program Synthesis from Natural Language
• Multiagent Systems: Adversarial Agents

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

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 hampered by indirect fine-tuning, limited alignment, or inconsistent targets. To overcome these limitations, we propose extit{SynAdapt}, an innovative efficient reasoning framework. Specifically, extit{SynAdapt} generates the synthetic CCoT to serve as a precise and effective alignment target for LLMs. This synthetic CCoT explicitly guides the LLM to learn CCoT and derive accurate answers directly. Furthermore, relying solely on CCoT is insufficient for solving hard questions. To address this, extit{SynAdapt} integrates a difficulty classifier that leverages both question context and CCoT to identify hard questions. CCoT can effectively help identify hard questions after some brief reasoning. We then adaptively prompt the LLM to re-think these hard questions for 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.

Problem

• Improving efficiency of continuous Chain-of-Thought reasoning
• Enhancing alignment and accuracy in synthetic reasoning targets
• Adaptively addressing hard questions via difficulty classification

Innovation

• Generates synthetic CCoT for precise alignment
• Integrates difficulty classifier for hard questions
• Adaptively prompts LLM to re-think hard questions