Research map / Large language models / Reasoning and chain of thought
Reasoning Models Better Express Their Confidence
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.
Benchmarks six reasoning models against non-reasoning counterparts on verbalized confidence calibration and shows slow-thinking CoT behaviors drive calibration gains, reproducible via in-context slow thinking in non-reasoning models.
Problem
LLMs often express overconfidence when verbalizing confidence; it is unclear whether reasoning models with extended chain-of-thought express their confidence more accurately and what in the CoT causes this.
Approach
Reasoning models' slow-thinking CoT (exploring alternatives, backtracking, verifying) lets them adjust verbalized confidence dynamically, so calibration improves progressively along the CoT; this is isolated by ablations and by inducing slow thinking via in-context examples in non-reasoning models.
Key results
- Reasoning models achieve strictly better calibration than non-reasoning counterparts in 33 of 36 settings across calibration metrics.
- Calibration of reasoning models improves steadily along the CoT (significant slopes, p<0.05), while non-reasoning models show no consistent or sometimes worsening trend.
- Removing non-linear reasoning from R1-Distill-Qwen-32B CoT raises ECE from 0.042 to 0.161 on TriviaQA; prompting non-reasoning models with slow-thinking exemplars improves their ECE (e.g., GLM-4-32B ECE 0.150 to 0.029 on TriviaQA).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use reasoning models instead of non-reasoning counterparts | six reasoning-model pairs · Across six datasets and 36 model-dataset settings | confidence calibration | Helped |
| Use Thinking Mode rather than Non-thinking Mode | Qwen3-32B · TriviaQA | ECE: 0.207 -> 0.063 | Helped |
| Use Thinking Mode rather than Non-thinking Mode | Qwen3-32B · NonambigQA | ECE: 0.403 -> 0.226 | Helped |
| Add slow-thinking few-shot examples | Qwen2.5-32B-Inst. · TriviaQA | ECE: 0.135 -> 0.091 | Helped |
| Add slow-thinking few-shot examples | Qwen2.5-32B-Inst. · NonambigQA | ECE: 0.309 -> 0.269 | Helped |
| Add slow-thinking few-shot examples | GLM-4-32B-0414 · TriviaQA | ECE: 0.150 -> 0.029 | Helped |
| Add slow-thinking few-shot examples | GLM-4-32B-0414 · NonambigQA | ECE: 0.338 -> 0.126 | Helped |
| Add slow-thinking few-shot examples | EXAONE-32B-Inst. · TriviaQA | ECE: 0.118 -> 0.096 | Helped |
| Add slow-thinking few-shot examples | EXAONE-32B-Inst. · NonambigQA | ECE: 0.326 -> 0.279 | Helped |
| Add explicit Confidence Reasoning (ablation) | R1-Distill-Qwen-32B · TriviaQA | ECE: 0.042 -> 0.063 | No clear effect |
| Add Epistemic Markers (ablation) | R1-Distill-Qwen-32B · TriviaQA | ECE: 0.109 -> 0.042 | Helped |
| Add Epistemic Markers (ablation) | R1-Distill-Qwen-32B · TriviaQA | AUROC: 0.837 -> 0.782 | Hurt |
Surprising findings
- Non-reasoning models have their lowest Brier score with no CoT, and calibration worsens by the end of their CoT.
- Removing epistemic markers from reasoning traces degrades ECE but improves AUROC, because the model becomes overconfident yet keeps discriminative ability.
Limitations
- Reasoning models still rarely express confidence below 55%, showing room for improvement.
- Multiple-choice format provides implicit cues that inflate AUROC for some non-reasoning models.
- Ablations rely on GPT-4.1-based CoT rewriting with manual inspection of only 100 samples per setting.
Experimental setup
Datasets: TriviaQA NonambigQA SuperGPQA MMLU-Pro
Benchmarks: TriviaQA NonambigQA SuperGPQA MMLU-Pro
Metrics: ECE Brier score AUROC accuracy
Baselines: Qwen2.5-32B-Instruct GLM-4-32B-0414 EXAONE-3.5-32B Qwen3 Non-thinking Two-turn prompting Top-K prompting Multi-step prompting
Builds on: chain-of-thought verbalized confidence estimation budget forcing
Closest papers by meaning
- LYNX: Learning Dynamic Exits for Confidence-Controlled Reasoning (NeurIPS 2026 · similarity 0.90)
- Deep Hidden Cognition Facilitates Reliable Chain-of-Thought Reasoning (AAAI 2026 · similarity 0.90)
- When More is Less: Understanding Chain-of-Thought Length in LLMs (ICLR 2026 · similarity 0.90)
- Are Reasoning LLMs Robust to Interventions on their Chain-of-Thought? (ICLR 2026 · similarity 0.90)
- Decoding the Critique Mechanism in Large Reasoning Models (NeurIPS 2026 · similarity 0.90)
- Facts Don't Speak Louder than Words: The Behavioral Essence of Long-CoT Distillation (NeurIPS 2026 · 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.