Saved in:
Bibliographic Details
Main Authors: Aghazadeh, Ehsan, Ghasemi, Ahmad, Beyhaghi, Hedyeh, Pishro-Nik, Hossein
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2511.02603
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866909887088295936
author Aghazadeh, Ehsan
Ghasemi, Ahmad
Beyhaghi, Hedyeh
Pishro-Nik, Hossein
author_facet Aghazadeh, Ehsan
Ghasemi, Ahmad
Beyhaghi, Hedyeh
Pishro-Nik, Hossein
contents Large language models (LLMs) are often queried multiple times at test time, with predictions aggregated by majority vote. While effective, this self-consistency strategy (arXiv:2203.11171) requires a fixed number of calls and can fail when the correct answer is rare. We introduce Confidence-Guided Early Stopping (CGES), a Bayesian framework that forms posteriors over candidate answers using scalar confidence signals derived from token probabilities or reward models. CGES adaptively halts sampling once the posterior mass of a candidate exceeds a threshold. We provide theoretical guarantees for both perfectly calibrated confidences and realistic noisy confidence signals. Across five reasoning benchmarks, CGES reduces the average number of model calls by about 69 percent (for example, from 16.0 to 4.9) while matching the accuracy of self-consistency within 0.06 percentage points.
format Preprint
id arxiv_https___arxiv_org_abs_2511_02603
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle CGES: Confidence-Guided Early Stopping for Efficient and Accurate Self-Consistency
Aghazadeh, Ehsan
Ghasemi, Ahmad
Beyhaghi, Hedyeh
Pishro-Nik, Hossein
Computation and Language
Large language models (LLMs) are often queried multiple times at test time, with predictions aggregated by majority vote. While effective, this self-consistency strategy (arXiv:2203.11171) requires a fixed number of calls and can fail when the correct answer is rare. We introduce Confidence-Guided Early Stopping (CGES), a Bayesian framework that forms posteriors over candidate answers using scalar confidence signals derived from token probabilities or reward models. CGES adaptively halts sampling once the posterior mass of a candidate exceeds a threshold. We provide theoretical guarantees for both perfectly calibrated confidences and realistic noisy confidence signals. Across five reasoning benchmarks, CGES reduces the average number of model calls by about 69 percent (for example, from 16.0 to 4.9) while matching the accuracy of self-consistency within 0.06 percentage points.
title CGES: Confidence-Guided Early Stopping for Efficient and Accurate Self-Consistency
topic Computation and Language
url https://arxiv.org/abs/2511.02603