Research map / Large language models / Reasoning and chain of thought
Efficient Reasoning for Large Reasoning Language Models via Certainty-Guided Reflection Suppression
Read the paperCheck your idea against 108,967 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 CGRS, a training-free decoding method that probes tentative answers at paragraph breaks, computes an entropy-based certainty score, and probabilistically suppresses reflection trigger tokens (e.g., "Wait", "But") to cut overthinking in reasoning models.
Problem
Large reasoning language models produce long chain-of-thought outputs with redundant reflection (triggered by words like "Wait" and "Alternatively") after they already reach correct answers, inflating token usage and inference cost.
Approach
Estimate confidence by probing tentative final answers at "\n\n" checkpoints and computing normalized entropy-based certainty; then suppress reflection-trigger token logits with probability proportional to certainty above a threshold, so reflection is cut only when the model is confident.
Key results
- On DeepSeek-R1-Distill-Qwen-7B, CGRS reduces average length by 41.9% (AMC23 and average) with average accuracy 65.2 matching Vanilla 66.8 within ~1.6 points.
- On Qwen3-14B, CGRS achieves 25.4% average length reduction with average accuracy 80.8 vs Vanilla 80.1, while NoThinking gives 77.3% reduction but drops accuracy 17.8 points.
- On QwQ-32B, CGRS achieves 30.5% average length reduction with accuracy 80.8 vs Vanilla 80.2, whereas baselines achieve only minor reductions.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| certainty-guided reflection trigger suppression (CGRS) | Qwen3-14B · AVG over AIME24, AMC23, MATH500, GPQA-D | Acc and length reduction: 25.4% length reduction, avg Acc 80.8 vs 80.1 Vanilla | Helped |
| certainty-guided reflection trigger suppression (CGRS) | DeepSeek-R1-Distill-Qwen-7B · AVG over four benchmarks | Acc and length reduction: 41.9% length reduction, avg Acc 65.2 vs 66.8 Vanilla | Helped |
| certainty-guided reflection trigger suppression (CGRS) | QwQ-32B · AVG over four benchmarks | length reduction with accuracy: 30.5% LR, avg Acc 80.8 vs 80.2 Vanilla | Helped |
| certainty-guided reflection trigger suppression (CGRS) | Qwen3-4B, Qwen3-8B, Qwen3-32B, DeepSeek-R1-Distill-Llama-8B · AVG over four benchmarks | length reduction (LR) with average accuracy: LR 38.1% (4B), 29.3% (8B), 31.2% (32B), 18.5% (Llama-8B); Acc 72.5, 75.9, 79.2, 64.7 vs Vanilla 74.5, 75.3, 82.2, 61.7 | Mixed |
| certainty-guided reflection trigger suppression (CGRS) | DeepSeek-R1-Distill-Qwen-7B · AMC23 | Acc vs length: Acc 88.3 vs 87.5; Len 3406 vs 5861 (41.9% LR) | Helped |
| fixed reflection-trigger suppression probability p=0.25 (vs certainty-guided) (ablation) | DeepSeek-R1-Distill-Qwen-7B · AMC23 | Acc: 81.7 vs 88.3 (certainty-guided) | Hurt |
| fixed reflection-trigger suppression probability p=1.0 (all triggers suppressed) (ablation) | DeepSeek-R1-Distill-Qwen-7B · AMC23 | Acc and Len: Acc 76.7 vs 88.3; Len 2373 vs 3406 | Hurt |
| confidence threshold delta lowered from 0.9 to 0.1 (ablation) | CGRS on DeepSeek-R1-Distill-Qwen-7B · AMC23 | Acc: 88.3 -> 72.5 | Hurt |
| certainty-guided suppression with fixed-interval reflection vs threshold 0.9 (higher threshold δ=0.9 vs 0.1) (ablation) | CGRS on DeepSeek-R1-Distill-Qwen-7B · AMC23 | LR: LR 41.9% at δ=0.9 vs 53.7% at δ=0.1 | Hurt |
| NoThinking prompting | DeepSeek-R1-Distill-Qwen-7B · MATH500 | Acc: LR 65.4% but accuracy drop about 9.4 points | Hurt |
| DEER early exit | Qwen3-14B · AVG over four benchmarks | Acc: 74.4 vs 80.1 Vanilla | Hurt |
Surprising findings
- Baseline methods achieve only minor length reductions on QwQ-32B because they rely on </think> boundaries that QwQ-32B often continues reflecting past.
- Fixed suppression probabilities reduce length more but cause large accuracy drops (e.g., p=1.0 gives 76.7% accuracy vs 87.5% vanilla on AMC23), motivating the certainty-guided schedule.
Limitations
- Certainty probing at every \n\n checkpoint adds extra generation cost not fully quantified in the paper.
- Accuracy drops of up to about 3 points on some individual benchmark cases (e.g., AIME24 for some models) and sensitivity to threshold delta.
- Trigger set built from DeepSeek-R1-Distill-Qwen-7B traces on AIME24 and AMC23 and tokenizer-specific token IDs, limiting portability.
Experimental setup
Datasets: AIME24 AMC23
Benchmarks: AIME24 AMC23 MATH500 GPQA-D
Metrics: pass@1 accuracy average output token length length reduction ratio
Baselines: Vanilla TALE NoThinking Dynasor DEER
Builds on: Dynasor DEER TALE NoThinking
Closest papers by meaning
- Mitigating Overthinking in Large Reasoning Language Models via Reasoning Path Deviation Monitoring (NeurIPS 2026 · similarity 0.91)
- Stop Unnecessary Reflection: Training LRMs for Efficient Reasoning with Adaptive Reflection and Length Coordinated Penalty (ICLR 2026 · similarity 0.91)
- Training Language Models to Reason Efficiently (NeurIPS 2025 · similarity 0.90)
- THE PATH OF LEAST RESISTANCE: GUIDING LLM REASONING TRAJECTORIES WITH PREFIX CONSENSUS (ICLR 2026 · similarity 0.89)
- Dynamic Early Exit in Reasoning Models (ICLR 2026 · similarity 0.89)
- Reasoning Path Compression: Compressing Generation Trajectories for Efficient LLM Reasoning (NeurIPS 2025 · 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.