Chain of Draft: Thinking Faster by Writing LessSilei Xu, Wenhao Xie, Lingxiao Zhao et al.
Large Language Models (LLMs) have demonstrated remarkable performance in solving complex reasoning tasks through mechanisms like Chain-of-Thought (CoT) prompting, which emphasizes verbose, step-by-step reasoning. However, humans typically employ a more efficient strategy: drafting concise intermediate thoughts that capture only essential information. In this work, we propose Chain of Draft (CoD), a novel paradigm inspired by human cognitive processes, where LLMs generate minimalistic yet informative intermediate reasoning outputs while solving tasks. By reducing verbosity and focusing on critical insights, CoD matches or surpasses CoT in accuracy while using as little as only 7.6% of the tokens, significantly reducing cost and latency across various reasoning tasks. Our code and data are available at https://github.com/sileix/chain-of-draft.
Harnessing Negative Signals: Reinforcement Distillation from Teacher Data for LLM ReasoningShuyao Xu, Cheng Peng, Jiangxuan Long et al.
Recent advances in model distillation demonstrate that data from advanced reasoning models (e.g., DeepSeek-R1, OpenAI's o1) can effectively transfer complex reasoning abilities to smaller, efficient student models. However, standard practices employ rejection sampling, discarding incorrect reasoning examples -- valuable, yet often underutilized data. This paper addresses the critical question: How can both positive and negative distilled reasoning traces be effectively leveraged to maximize LLM reasoning performance in an offline setting? To this end, We propose Reinforcement Distillation (REDI), a two-stage framework. Stage 1 learns from positive traces via Supervised Fine-Tuning (SFT). Stage 2 further refines the model using both positive and negative traces through our proposed REDI objective. This novel objective is a simple, reference-free loss function that outperforms established methods like DPO and SimPO in this distillation context. Our empirical evaluations demonstrate REDI's superiority over baseline Rejection Sampling SFT or SFT combined with DPO/SimPO on mathematical reasoning tasks. Notably, the Qwen-REDI-1.5B model, post-trained on just 131k positive and negative examples from the open Open-R1 dataset, achieves an 83.1% score on MATH-500 (pass@1). Its performance matches or surpasses that of DeepSeek-R1-Distill-Qwen-1.5B (a model post-trained on 800k proprietary data) across various mathematical reasoning benchmarks, establishing a new state-of-the-art for 1.5B models post-trained offline with openly available data.
2.7CLSep 18, 2025
Controlling Language Difficulty in Dialogues with Linguistic FeaturesShuyao Xu, Wenguang Wang, Handong Gao et al.
Large language models (LLMs) have emerged as powerful tools for supporting second language acquisition, particularly in simulating interactive dialogues for speaking practice. However, adapting the language difficulty of LLM-generated responses to match learners' proficiency levels remains a challenge. This work addresses this issue by proposing a framework for controlling language proficiency in educational dialogue systems. Our approach leverages three categories of linguistic features, readability features (e.g., Flesch-Kincaid Grade Level), syntactic features (e.g., syntactic tree depth), and lexical features (e.g., simple word ratio), to quantify and regulate text complexity. We demonstrate that training LLMs on linguistically annotated dialogue data enables precise modulation of language proficiency, outperforming prompt-based methods in both flexibility and stability. To evaluate this, we introduce Dilaprix, a novel metric integrating the aforementioned features, which shows strong correlation with expert judgments of language difficulty. Empirical results reveal that our approach achieves superior controllability of language proficiency while maintaining high dialogue quality.