Research map / Large language models / Reasoning and chain of thought
Best-of-Infinity: Asymptotic Performance of Test-Time Compute
Junpei Komiyama, Daisuke Oba, Masafumi Oyamada
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.
Studies best-of-N with majority-vote selection in the limit N→∞ (best-of-∞) and proposes an adaptive generation scheme that chooses N based on answer agreement to approach that limit efficiently.
Problem
Best-of-N test-time scaling with majority-vote selection requires an effectively infinite test-time budget to reach its asymptotic performance.
Approach
Analyzes the N→∞ limit of majority-vote best-of-N (best-of-∞) and proposes adaptively choosing N according to answer agreement among sampled responses, so generation stops once the vote is stable.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use adaptive Bayes-factor sampling instead of fixed N in majority voting | GPT-OSS-20B majority voting · MATH500 | accuracy vs samples: average N=3 matches fixed N=10; average N≈10 matches fixed N=100 | Improved |
| use MILP-weighted ensemble of multiple LLMs instead of the best single LLM | best-of-infinity majority voting · AIME2025 | best-of-infinity accuracy: 90.0 -> 93.3 | Improved |
| ensemble of five LLMs with optimized weights instead of any single LLM | majority voting · GPQA-Diamond | accuracy | Improved |
| learn ensemble weights from AIME2024 and apply to AIME2025 | three-model ensembles · AIME2025 (transfer) | matches or exceeds best single model: 106 of 165 combinations (64.2%) | Mixed |
| select answer by majority voting instead of reward models or LLM-as-a-judge | GPT-OSS-20B Bo5 · AIME2025 | accuracy: 85.42 vs 82.92 (judge), 80.00 (best reward model) | Improved |
| select answer by self-certainty instead of random selection | GPT-OSS-20B Bo5 · AIME2025 | accuracy: 76.25 -> 75.83 | No clear effect |
| use more samples N in majority voting (ablation) | GPT-OSS-20B · four reasoning benchmarks | accuracy | Improved |
| learn weights from only five training problems (ablation) | ensemble weight MILP · AIME2025 | best-of-infinity accuracy | Improved |
Experimental setup
Builds on: best-of-n sampling self-consistency
Closest papers by meaning
- Provable Scaling Laws for the Test-Time Compute of Large Language Models (NeurIPS 2025 · similarity 0.89)
- Scalable Best-of-N Selection for Large Language Models via Self-Certainty (NeurIPS 2025 · similarity 0.88)
- Inference Scaling Laws: An Empirical Analysis of Compute-Optimal Inference for LLM Problem-Solving (ICLR 2025 · similarity 0.88)
- Majority of the Bests: Improving Best-of-N via Bootstrapping (NeurIPS 2025 · similarity 0.88)
- Are More LLM Calls All You Need? Towards the Scaling Properties of Compound AI Systems (NeurIPS 2024 · similarity 0.87)
- On the Limits of Test-Time Compute: Sequential Reward Filtering for Better Inference (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.
