Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes
Optimal Bayesian Stopping for Efficient Inference of Consistent LLM Answers
Jingkai Huang, Will Ma, Zhengyuan Zhou
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 |
|---|---|---|---|
| aggregate to L=3 instead of exact posterior (ablation) | L-aggregated Bayesian stopping · Synthetic (K=8, 1-δ=0.99) | number of generations: 18.13 -> 18.07 | No clear effect |
| use L=3 instead of coarsest L=2 (ablation) | L-aggregated Bayesian stopping · Synthetic (K=8, 1-δ=0.99) | number of generations: 22.43 -> 18.07 | Improved |
| increase L from 3 to 4 (ablation) | L-aggregated Bayesian stopping · Synthetic (K=8) | posterior computation time (ms): 14.2 -> 37.4 | Hurt |
| increase L from 3 to 4 (ablation) | L-aggregated Bayesian stopping · Synthetic (K=8, 1-δ=0.99) | number of generations: 18.07 -> 18.11 | No clear effect |
| use L=3 instead of L=2 at 1-δ=0.9 (ablation) | L-aggregated Bayesian stopping · Synthetic (K=8, 1-δ=0.9) | number of generations: 10.74 -> 10.12 | Improved |
| add known-prior Bayesian stopping (L=3) instead of prior-free ASC | L-aggregated Bayesian stopping · Synthetic (K=8, 1-δ=0.9) | number of generations: 16.72 -> 10.12 | Improved |
| use known-prior L=3 instead of ASC | L-aggregated Bayesian stopping · CommonsenseQA, Qwen-2.5-72B (1-δ=0.95) | number of generations: 5.28 -> 3.38 | Improved |
| use uncertain-prior L=3 instead of ASC | L-aggregated Bayesian stopping · CommonsenseQA, Qwen-2.5-72B (1-δ=0.95) | number of generations: 5.28 -> 1.00 | Improved |
| use uncertain-prior L=3 instead of ASC | L-aggregated Bayesian stopping · CommonsenseQA, LLaMA-3.1-405B (1-δ=0.95) | answer accuracy: 90.4% -> 88.6% | Hurt |
| use uncertain-prior L=3 instead of ASC | L-aggregated Bayesian stopping · CommonsenseQA, Qwen-2.5-72B (1-δ=0.95) | answer accuracy: 87.2% -> 86.3% | Hurt |
| use uncertain-prior L=3 instead of L=2 (ablation) | L-aggregated Bayesian stopping · CommonsenseQA, LLaMA-3.1-405B (1-δ=0.95) | number of generations: 5.63 -> 4.58 | Improved |
| use known-prior L=3 instead of L=2 (ablation) | L-aggregated Bayesian stopping · CommonsenseQA, LLaMA-3.1-405B (1-δ=0.95) | number of generations: 6.92 -> 6.50 | Improved |
Closest papers by meaning
- Statistical Early Stopping for Reasoning Models (ICML 2026 · similarity 0.90)
- CITE: Anytime Valid Statistical Inference in LLM Self-Consistency (NeurIPS 2026 · similarity 0.89)
- Sample Smart, Not Hard: Correctness-First Decoding for Better Reasoning in LLMs (ICLR 2026 · similarity 0.88)
- Representation Consistency for Accurate and Coherent LLM Answer Aggregation (NeurIPS 2025 · similarity 0.88)
- From Drift to Coherence: Stabilizing Beliefs in LLMs (ICML 2026 · similarity 0.87)
- Evaluating LLMs When They Do Not Know the Answer: Statistical Evaluation of Mathematical Reasoning via Comparative Signals (ICML 2026 · similarity 0.87)
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.
