Hot or Cold? Adaptive Temperature Sampling for Code Generation with Large Language Models

TL;DR

This paper introduces Adaptive Temperature sampling, a novel decoding strategy for code generation with large language models that dynamically adjusts sampling temperature based on token difficulty, improving performance.

Contribution

It is the first systematic study of decoding strategies tailored for code generation, proposing a dynamic temperature adjustment method based on token difficulty analysis.

Findings

AdapT sampling outperforms existing decoding strategies.

Challenging tokens mainly occur at the beginning of code blocks.

Dynamic temperature adjustment improves code generation quality.

Abstract

Recently, Large Language Models (LLMs) have shown impressive abilities in code generation. However, existing LLMs' decoding strategies are designed for Natural Language (NL) generation, overlooking the differences between NL and programming languages (PL). Due to this oversight, a better decoding strategy for code generation remains an open question. In this paper, we conduct the first systematic study to explore a decoding strategy specialized in code generation. With an analysis of loss distributions of code tokens, we find that code tokens can be divided into two categories: challenging tokens that are difficult to predict and confident tokens that can be easily inferred. Among them, the challenging tokens mainly appear at the beginning of a code block. Inspired by the above findings, we propose a simple yet effective method: Adaptive Temperature (AdapT) sampling, which dynamically adjusts the temperature coefficient when decoding different tokens. We apply a larger temperature when sampling for challenging tokens, allowing LLMs to explore diverse choices. We employ a smaller temperature for confident tokens avoiding the influence of tail randomness noises. We apply AdapT sampling to LLMs with different sizes and conduct evaluations on two popular datasets. Results show that AdapT sampling significantly outperforms state-of-the-art decoding strategy.