Research map / Large language models / Reasoning and chain of thought
Aware First, Think Less: Dynamic Boundary Self-Awareness Drives Significant Gains in Reasoning Efficiency in Large Language Models
Qiguang Chen, Dengyun Peng, Jinhao Liu, Huikang Su, Jiannan Guan, Libo Qin, Wanxiang Che
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a dynamic boundary self-awareness approach that lets LLMs judge problem difficulty themselves to reduce redundant long chain-of-thought reasoning while keeping accuracy.
Problem
Long chain-of-thought reasoning in LLMs produces substantial redundancy, hurting computational efficiency and latency; existing efficiency methods rely on human-defined difficulty priors that do not align with the model's own capability.
Approach
Instead of human-defined difficulty priors, the model learns a dynamic boundary of its own capability so it becomes aware of when to answer directly versus think longer, allocating reasoning length per problem.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Boundary Self-Awareness Alignment (ablation) | DR. SAF · Average over five benchmarks | token efficiency: 8.04 -> 13.65 | Improved |
| add Adaptive Length Management instead of simple length penalty (ablation) | DR. SAF · Average over five benchmarks | accuracy: 67.54 -> 75.28 | Improved |
| add Adaptive Length Management instead of simple length penalty (ablation) | DR. SAF · Average over five benchmarks | token efficiency: 11.96 -> 13.65 | Improved |
| add Boundary Preservation Mechanism (ablation) | DR. SAF · Average over five benchmarks | accuracy: 67.87 -> 75.28 | Improved |
| add Boundary Preservation Mechanism (ablation) | DR. SAF · Average over five benchmarks | token efficiency: 13.65 -> 13.65 | No clear effect |
| add Boundary Preservation Mechanism (ablation) | DR. SAF · BigGSM out-of-domain | accuracy: 54.8 -> 60.7 | Improved |
| apply DR. SAF to R1-Distill-Qwen3-8B | R1-Distill-Qwen3-8B · GSM8K | token efficiency: 4.41 -> 17.72 | Improved |
| apply DR. SAF to R1-Distill-Qwen3-8B | R1-Distill-Qwen3-8B · AMC | accuracy: 83.5 -> 95.6 | Improved |
| use DR. SAF instead of AdaptThink on 7B | R1-Distill-Qwen2.5-7B · GSM8K | token efficiency: 29.45 -> 54.38 | Improved |
| use DR. SAF instead of AdaptThink on 7B | R1-Distill-Qwen2.5-7B · GSM8K | accuracy: 91.0 -> 88.1 | Hurt |
| use DR. SAF instead of Length-Penalty | R1-Distill-Qwen2.5-7B · AIME24 | accuracy: 51.9 -> 50.6 | Mixed |
Experimental setup
Builds on: long chain-of-thought
Closest papers by meaning
- Longer Context, Deeper Thinking: Uncovering the Role of Long-Context Ability in Reasoning (NeurIPS 2025 · similarity 0.90)
- Think Deep, Not Just Long: Measuring LLM Reasoning Effort via Deep-Thinking Tokens (ICML 2026 · similarity 0.89)
- Don’t Think Longer, Think Wisely: Optimizing Thinking Dynamics for Large Reasoning Models (NeurIPS 2025 · similarity 0.89)
- LEAD: Length-Efficient Adaptive and Dynamic Reasoning for Large Language Models (NeurIPS 2026 · similarity 0.89)
- A Theoretical Study on Bridging Internal Probability and Self-Consistency for LLM Reasoning (NeurIPS 2025 · similarity 0.89)
- MILES: Modular Instruction Memory with Learnable Selection for Self-Improving LLM Reasoning (NeurIPS 2026 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
