When Are Likely Answers Right? On Sequence Probability and Correctness in LLMs

Forum topic · 小凯 · 2026-06-27

Summary

This arXiv paper (2606.27359) by Johannes Zenn and Jonas Geiping investigates a fundamental question underlying LLM decoding methods: when does sequence probability—the conditional probability of a continuation given a prompt—actually align with correctness? Many decoding methods for large language models work by shifting probability mass toward outputs that are more likely under the model, either locally at the token level or globally at the sequence level. The success of such approaches therefore depends on whether higher probability correlates with answers being factually or logically correct. The authors set out to quantify this relationship across different decoding methods and models, providing empirical analysis of when likely answers are also right ones. This work is relevant to researchers studying decoding strategies, calibration, and the reliability of likelihood-based selection in language model inference.

Paper Overview

Research Area: Machine Learning Authors: Johannes Zenn, Jonas Geiping Published: 2026-06-27 arXiv: 2606.27359

Abstract

Many decoding methods for large language models can be understood as shifting probability mass toward outputs that are more likely under the model, either locally at the token level or globally at the sequence level. Therefore, their success depends on a fundamental question: when does sequence probability, that is, the conditional probability of a continuation given a prompt, actually align with correctness? In this paper, we set out to quantify this relationship across decoding methods, models...

Why This Matters

Decoding techniques that bias generation toward higher-probability outputs implicitly assume that probability tracks correctness. This paper directly examines that assumption, quantifying the alignment between sequence probability and answer correctness across decoding methods and models. Understanding when likely answers are right has implications for decoding strategy design, answer selection, and calibration in LLM inference.

---

*Auto-collected on 2026-06-27*