Chain of Draft: Thinking Faster by Writing Less

2025/02/25 by Shijing Xu, Xu, Silei, Wenhao Xie +4 · 17 voices · 56 citations
Computer Science · Medicine · #Artificial Intelligence in Healthcare and Education #Computation and Language (cs.CL) #FOS: Computer and information sciences #I.2.7 #Multimodal Machine Learning Applications #Topic Modeling

paper · pdf · doi:10.48550/arxiv.2502.18600

openalex publication_date 2025/02/25 · openalex created_date 2025/10/12 · openalex updated_date 2026/07/28

Abstract

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.

Cited by

• ThinkGen: Generalized Thinking for Visual Generation
• Understanding Tone-Dependent Inference Cost in Large Language Models
• DeepLook: Deeper Thinking with Lookahead
• When Reasoning Meets Its Laws
• Asynchronous Reasoning: Training-Free Interactive Thinking LLMs
• Information-Dense Reasoning for Efficient and Auditable Security Alert Triage
• LLM-Cave: A benchmark and light environment for large language models reasoning and decision-making system
• Focused Chain-of-Thought: Efficient LLM Reasoning via Structured Input Information
• DRAFT-RL: Multi-Agent Chain-of-Draft Reasoning for Reinforcement Learning-Enhanced LLMs
• Reuse, Don't Recompute: Efficient Large Reasoning Model Inference via Memory Orchestration
• ChainV: Atomic Visual Hints Make Multimodal Reasoning Shorter and Better
• Learning to Think Fast and Slow for Visual Language Models
• TokenSqueeze: Performance-Preserving Compression for Reasoning LLMs
• EcoAlign: An Economically Rational Framework for Efficient LVLM Alignment
• From Efficiency to Adaptivity: A Deeper Look at Adaptive Reasoning in Large Language Models
• Investigating CoT Monitorability in Large Reasoning Models
• Efficient Reasoning via Thought-Training and Thought-Free Inference
• DTS: Enhancing Large Reasoning Models via Decoding Tree Sketching
• MR-Align: Meta-Reasoning Informed Factuality Alignment for Large Reasoning Models
• SheetBrain: A Neuro-Symbolic Agent for Accurate Reasoning over Complex and Large Spreadsheets
• Training-Free Spectral Fingerprints of Voice Processing in Transformers
• Video Reasoning without Training
• Putting on the Thinking Hats: A Survey on Chain of Thought Fine-tuning from the Perspective of Human Reasoning Mechanism
• ThinkPilot: Steering Reasoning Models via Automated Think-prefixes Optimization
• Are Large Reasoning Models Interruptible?
• Revisiting Model Interpolation for Efficient Reasoning
• Merlin's Whisper: Enabling Efficient Reasoning in LLMs via Black-box Adversarial Prompting
• One Token Embedding Is Enough to Deadlock Your Large Reasoning Model
• Concise Reasoning in the Lens of Lagrangian Optimization
• Adaptive Dual Reasoner: Large Reasoning Models Can Think Efficiently by Hybrid Reasoning
• Mitigating Overthinking through Reasoning Shaping
• Logit Arithmetic Elicits Long Reasoning Capabilities Without Training
• On the Provable Performance Guarantee of Efficient Reasoning Models
• DeepPrune: Parallel Scaling without Inter-trace Redundancy
• Upfront Chain-of-Thought: A Cooperative Framework for Chain-of-Thought Compression
• Gold-Switch: Training-Free Superposition of Slow- and Fast- Thinking LLMs
• The Valley of Code Reasoning: Scaling Knowledge Distillation of Large Language Models
• Explore Briefly, Then Decide: Mitigating LLM Overthinking via Cumulative Entropy Regulation
• Beyond Token Length: Step Pruner for Efficient and Accurate Reasoning in Large Language Models
• RoRecomp: Enhancing Reasoning Efficiency via Rollout Response Recomposition in Reinforcement Learning
• Dynamic Policy Induction for Adaptive Prompt Optimization: Bridging the Efficiency-Accuracy Gap via Lightweight Reinforcement Learning
• HiPO: Hybrid Policy Optimization for Dynamic Reasoning in LLMs
• From Long to Lean: Performance-aware and Adaptive Chain-of-Thought Compression via Multi-round Refinement
• Reinforcement Learning-Guided Chain-of-Draft for Token-Efficient Code Generation
• StyleBench: Evaluating thinking styles in Large Language Models
• Parallel Thinking, Sequential Answering: Bridging NAR and AR for Efficient Reasoning
• Early Stopping Chain-of-thoughts in Large Language Models
• Meta-R1: Empowering Large Reasoning Models with Metacognition
• BudgetThinker: Empowering Budget-aware LLM Reasoning with Control Tokens
• A Comprehensive Survey on Trustworthiness in Reasoning with Large Language Models
• Mini-Omni-Reasoner: Token-Level Thinking-in-Speaking in Large Speech Models
• Promoting Efficient Reasoning with Verifiable Stepwise Reward
• Pruning Long Chain-of-Thought of Large Reasoning Models via Small-Scale Preference Optimization
• Train Long, Think Short: Curriculum Learning for Efficient Reasoning
• Don't Overthink It: A Survey of Efficient R1-style Large Reasoning Models
• SynAdapt: Learning Adaptive Reasoning in Large Language Models via Synthetic Continuous Chain-of-Thought

Discussions

• Chain of Draft: Thinking Faster by Writing Less [hn, 4 points, 0 comments]
• "Chain of Draft," a possible efficiency improvement over the chain-of-thought process used in existing leading-edge AI models. They also provide their code and data. Seems like something to keep one's [bsky, 2 points, 0 comments]
• Chain of Draft: Thinking Faster by Writing Less [hn, 2 points, 0 comments]
• Chain of Draft: Thinking Faster by Writing Less [hn, 2 points, 0 comments]
• Chain of Draft: Thinking Faster by Writing Less [hn, 2 points, 0 comments]
• 💰 Significantly lowers computational costs for complex reasoning 🔄 Works with leading #LLMs like GPT-4o and Claude 3.5 Sonnet 🧮 Excels in arithmetic, common-sense, and symbolic reasoning benchmarks [bsky, 2 points, 0 comments]
• Chain of Draft (PDF) https://arxiv.org/pdf/2502.18600 #AI #research [bsky, 2 points, 0 comments]
• Chain-of-Draft surpasses CoT with only 7.6% of tokens [hn, 1 points, 0 comments]
• Chain of Drafts is a promising prompting strategy successfully applied to frozen models like Claude 3.7. My own *anecdotal* test show *qualitative* improvements in responses supporting the paper’s f [bsky, 1 points, 0 comments]
• Link to paper: buff.ly/43aAwY2 #LLM #AIEfficiency #PromptEngineering #MachineLearning [bsky, 1 points, 0 comments]
• This is The Way. Think more with fewer tokens, use smaller models, have agents collaborate. Iterate on real world feedback. arxiv.org/abs/2502.18600 [bsky, 1 points, 0 comments]
• Chain of Draft: Thinking Faster by Writing Less [hn, 1 points, 0 comments]
• ps : ça me fait penser à cette autre recherche, elle sur l'idée de pousser les IA de raisonnement à être moins verbeuses dans leurs CoT (ici sans réduire le nombre d'étapes, mais la taille de chacune [bsky, 0 points, 1 comments]
• Chain of Draft: Thinking Faster by Writing Less https://arxiv.org/abs/2502.18600 [bsky, 0 points, 0 comments]
• Chain-of-Thought (CoT) prompting emphasizes verbose, step-by-step reasoning. However, humans typically employ a more efficient strategy: drafting concise intermediate thoughts. By reducing verbosity a [bsky, 0 points, 0 comments]
• Chain of Draft: Thinking Faster by Writing Less arxiv.org/abs/2502.18600 [bsky, 0 points, 0 comments]
• https://arxiv.org/abs/2502.18600 この海外論文では、LLMにおける推論タスクにおいて、Chain-of-Thoughtプロンプトよりも効率的なChain of Draftという新しいパラダイムを提案。 CoDは、人間が思考する際の重要な情報のみを抽出するという戦略に着想を得て、LLMに最小限の情報で推論させるものです。 CoDは、CoTと同等以上の精度を維持しながら [bsky, 0 points, 0 comments]

Related

• Sketch-of-Thought: Efficient LLM Reasoning with Adaptive Cognitive-Inspired Sketching
• Focused Chain-of-Thought: Efficient LLM Reasoning via Structured Input Information
• Efficient Reasoning for LLMs through Speculative Chain-of-Thought
• Efficient Reasoning Models: A Survey
• SCOUT: Teaching Pre-trained Language Models to Enhance Reasoning via Flow Chain-of-Thought