Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Contrastive Decoding Mitigates Score Range Bias in LLM-as-a-Judge

Yoshinari Fujinuma

TL;DR

The paper analyzes the reliability of LLMs as judges for direct assessment, identifying score range bias as a pervasive issue that appears across model sizes and families. It proposes contrastive decoding, formulating the adjusted score as $\log p_{main} - \lambda \log p_{assist}$ with $p_{assist}$ computed via a temperature-controlled softmax, and uses a hyperparameter $\lambda$ to align logit distributions across models. The authors demonstrate that this approach yields robust mitigation across different score ranges, achieving up to 11.3% relative improvement in Spearman correlation with human judgments on summarization tasks. These findings broaden the viable scoring ranges beyond the standard 1-5 scale and illuminate how model-family biases interact with evaluation, along with practical considerations such as modest increases in inference-time compute and the influence of assistant-model choice.

Abstract

Large Language Models (LLMs) are commonly used as evaluators in various applications, but the reliability of the outcomes remains a challenge. One such challenge is using LLMs-as-judges for direct assessment, i.e., assigning scores from a specified range without any references. We first show that this challenge stems from LLM judge outputs being associated with score range bias, i.e., LLM judge outputs are highly sensitive to pre-defined score ranges, preventing the search for optimal score ranges. We also show that similar biases exist among models from the same family. We then mitigate this bias through contrastive decoding, achieving up to 11.3% relative improvement on average in Spearman correlation with human judgments across different score ranges.

Contrastive Decoding Mitigates Score Range Bias in LLM-as-a-Judge

TL;DR

The paper analyzes the reliability of LLMs as judges for direct assessment, identifying score range bias as a pervasive issue that appears across model sizes and families. It proposes contrastive decoding, formulating the adjusted score as with computed via a temperature-controlled softmax, and uses a hyperparameter to align logit distributions across models. The authors demonstrate that this approach yields robust mitigation across different score ranges, achieving up to 11.3% relative improvement in Spearman correlation with human judgments on summarization tasks. These findings broaden the viable scoring ranges beyond the standard 1-5 scale and illuminate how model-family biases interact with evaluation, along with practical considerations such as modest increases in inference-time compute and the influence of assistant-model choice.

Abstract

Large Language Models (LLMs) are commonly used as evaluators in various applications, but the reliability of the outcomes remains a challenge. One such challenge is using LLMs-as-judges for direct assessment, i.e., assigning scores from a specified range without any references. We first show that this challenge stems from LLM judge outputs being associated with score range bias, i.e., LLM judge outputs are highly sensitive to pre-defined score ranges, preventing the search for optimal score ranges. We also show that similar biases exist among models from the same family. We then mitigate this bias through contrastive decoding, achieving up to 11.3% relative improvement on average in Spearman correlation with human judgments across different score ranges.
Paper Structure (26 sections, 1 equation, 3 figures, 9 tables)

This paper contains 26 sections, 1 equation, 3 figures, 9 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. LLM Judge Tasks
  4. LLM Judge Biases
  5. Contrastive Decoding for Mitigating LLM Judge Biases
  6. Experiments
  7. Setup
  8. Task and Metrics
  9. Score Scale and Ranges
  10. Models
  11. Dataset
  12. Reveal and Mitigate Score Range Biases
  13. Similar score range biases exist across models from the same family
  14. Contrastive Decoding is a Robust Mitigation Strategy across Different Score Ranges
  15. Does Assistant Model Choice Matter for Bias Mitigation?
  16. ...and 11 more sections

Figures (3)

  • Figure 1: Overview of score range bias in 2-4 range and how contrastive decoding mitigates it through canceling out similar bias across models from the same family.
  • Figure 2: Coherence score distribution in 2-6 score range with greedy decoding, contrastive decoding, and human annotations. The greedy decoding outputs from both Llama 8B and 3B models are highly skewed towards outputting score of 4 and Qwen2.5 3B and 7B models are outputting score of 2 showing similar biases are encoded in these models.
  • Figure 3: Logit distribution of the first output token in 0-4 score range. For 3B and 7B models, the logit of Score 2 is the highest while the logit of Score 3 gets higher and becomes closer to human judgements as the model size increase from 3B through 14B.