Research map / Architectures and efficiency / Attention and transformers
Dynamic Thinking-Token Selection for Efficient Reasoning in Large Reasoning Models
Zhenyuan Guo, Tong Chen, Wenlong Meng, Chen GONG, Xin Yu, Chengkun Wei, Wenzhi CHEN
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add local window tokens (ablation) | DynTS (R1-Llama) · Six reasoning benchmarks (average) | Pass@1: 59.0 -> 61.9 | Improved |
| add question tokens (ablation) | DynTS (R1-Llama) · Six reasoning benchmarks (average) | Pass@1: 39.8 -> 61.9 | Improved |
| add critical thinking tokens (ablation) | DynTS (R1-Llama) · Six reasoning benchmarks (average) | Pass@1: 59.6 -> 61.9 | Improved |
| add local window tokens (ablation) | DynTS (R1-Qwen) · Six reasoning benchmarks (average) | Pass@1: 61.6 -> 65.3 | Improved |
| add question tokens (ablation) | DynTS (R1-Qwen) · Six reasoning benchmarks (average) | Pass@1: 62.6 -> 65.3 | Improved |
| add critical thinking tokens (ablation) | DynTS (R1-Qwen) · Six reasoning benchmarks (average) | Pass@1: 63.5 -> 65.3 | Improved |
| select thinking tokens by predicted importance instead of randomly (ablation) | DynTS (R1-Llama) · Six reasoning benchmarks (average) | Pass@1: 42.2 -> 61.9 | Improved |
| select thinking tokens by predicted importance instead of bottom-k (ablation) | DynTS (R1-Qwen) · Six reasoning benchmarks (average) | Pass@1: 57.6 -> 65.3 | Improved |
| select thinking tokens by predicted importance instead of randomly (ablation) | DynTS (R1-Qwen) · Six reasoning benchmarks (average) | Pass@1: 61.9 -> 65.3 | Improved |
| retaining random or bottom thinking tokens instead of removing all thinking tokens (ablation) | DynTS (R1-Llama) · Six reasoning benchmarks (average) | Pass@1: 59.6 -> 39.8 | Hurt |
| replace R-KV with DynTS | R1-Llama KV compression · Six reasoning benchmarks (average) | Pass@1: 59.6 -> 61.9 | Improved |
| replace SnapKV with DynTS | R1-Qwen KV compression · Six reasoning benchmarks (average) | Pass@1: 63.2 -> 65.3 | Improved |
Closest papers by meaning
- BeaconKV: Key-Value Cache Compression Guided by Beacon Queries for Efficient Large Reasoning Model Inference (ICML 2026 · similarity 0.91)
- Tracing the Traces: Latent Temporal Signals for Efficient and Accurate Reasoning (ICLR 2026 · similarity 0.91)
- ThinKV: Thought-Adaptive KV Cache Compression for Efficient Reasoning Models (ICLR 2026 · similarity 0.90)
- Think or Not? Exploring Thinking Efficiency in Large Reasoning Models via an Information-Theoretic Lens (NeurIPS 2025 · similarity 0.90)
- Which Heads Matter for Reasoning? RL-Guided KV Cache Compression (ICML 2026 · similarity 0.90)
- Don’t Think Longer, Think Wisely: Optimizing Thinking Dynamics for Large Reasoning Models (NeurIPS 2025 · similarity 0.90)
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.
