ThoughtFold: Folding Reasoning Chains via Introspective Preference Learning
The paper introduces ThoughtFold, a framework designed to improve the efficiency of Large Reasoning Models (LRMs) by addressing redundant explorations in long reasoning chains. By employing introspective preference learning, ThoughtFold penalizes unnecessary steps and encourages more concise reasoning paths. Experimental results indicate that this approach significantly reduces token usage while maintaining high accuracy.
- ▪ThoughtFold aims to mitigate redundant explorations in long reasoning chains of LRMs.
- ▪The framework utilizes fine-grained preference learning to optimize reasoning efficiency.
- ▪Experiments show a 56% reduction in token usage for DeepSeek-R1-Distill-Qwen-7B while preserving state-of-the-art accuracy.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2606.03503 (cs) [Submitted on 2 Jun 2026] Title:ThoughtFold: Folding Reasoning Chains via Introspective Preference Learning Authors:Ziyan Liu, Xueda Shen, Yuzhe Gu, Songyang Gao, Kuikun Liu, Guangran Cheng, Chengqi Lyu, Dahua Lin, Wenwei Zhang, Kai Chen View a PDF of the paper titled ThoughtFold: Folding Reasoning Chains via Introspective Preference Learning, by Ziyan Liu and 9 other authors View PDF Abstract:Large Reasoning Models (LRMs) have achieved remarkable progress thanks to Reinforcement Learning with Verifiable Rewards (RLVR) on Chain-of-Thoughts (CoTs).
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
