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

1. Problem Statement

Large Language Models (LLMs) are increasingly used as automated judges ("LLM-as-a-Judge") to evaluate model outputs, particularly in direct assessment tasks where a model assigns a numerical score to a single output without a reference. While pairwise comparison (choosing between two outputs) has shown higher reliability, direct assessment remains a critical bottleneck.

The paper identifies a specific, previously under-explored bias termed Score Range Bias.

• Definition: LLM judges exhibit inconsistent correlations with human judgments depending on the specific numerical range assigned to the evaluation scale (e.g., 0–4 vs. 1–5 vs. 2–6), even when the semantic meaning of the scale remains identical.
• Manifestation: Models tend to favor specific scores within a range regardless of the actual content quality. For instance, Llama models might cluster heavily around a score of "4" in a 2–6 range, while Qwen models might cluster around "2" in a 0–4 range.
• Family Bias: This bias is not random; models from the same family (e.g., Llama-3-3B and Llama-3-8B) encode similar biases, suggesting a shared architectural or training artifact.

2. Methodology

The authors propose a mitigation strategy using Contrastive Decoding, adapted from prior work on open-ended text generation.

Core Mechanism

The method employs two models:

1. Main Model ($p_{main}$): The primary judge generating the score.
2. Assistant Model ($p_{asst}$): A smaller model from the same family, hypothesized to share similar biases but with different magnitudes.

The final output probability is adjusted by subtracting the weighted log-probability of the assistant model from the main model:
$$\text{Adjusted Logit} = \log p_{main} - \lambda \log p_{asst}$$
Where:

• $\lambda$ is a hyperparameter controlling the magnitude of the subtraction.
• The authors introduce a temperature $t$ to normalize the logit distributions, as different model sizes (e.g., 3B vs. 14B) produce logits with different scales.

Hypothesis

Since models from the same family share similar score range biases (e.g., both favoring a specific number), using a smaller model as an "assistant" allows the system to cancel out these shared biases. The subtraction effectively removes the systematic error common to the family, leaving a signal closer to human judgment.

Experimental Setup

• Task: Summarization evaluation (Coherence, Relevance, Consistency) using the SummEval dataset (1,600 summaries).
• Models:
  • Llama-3 Family: Main (8B), Assistants (3B, 1B).
  • Qwen-2.5 Family: Main (14B, 7B), Assistant (3B).
• Score Ranges: Evaluated across four 5-point Likert scales: 0–4, 1–5, 2–6, and 3–7.
• Metrics: Pearson, Spearman, and Kendall correlations against human annotations.

3. Key Contributions

1. Discovery of Score Range Bias: The paper empirically demonstrates that LLM judges are highly sensitive to the specific numerical range used for scoring. A model that performs well on a 1–5 scale may perform significantly worse on a 2–6 scale, despite the scales being semantically equivalent.
2. Family-Shared Bias Analysis: Through logit distribution analysis (Figure 3), the authors show that models within the same family encode similar biases (e.g., Qwen models consistently favoring score "2" in the 0–4 range).
3. Contrastive Decoding Mitigation: The authors successfully adapt contrastive decoding to cancel out these shared biases. By subtracting the assistant's biased logits from the main model, the resulting score distribution aligns more closely with human annotations.
4. Robustness Across Ranges: The method stabilizes performance, allowing LLM judges to be reliable across arbitrary score ranges, not just the standard 1–5.

4. Results

The experiments demonstrate significant improvements in correlation with human judgments:

• Performance Gains:
  • Llama-3 Family: Contrastive decoding (8B-3B) achieved up to 6.7% relative improvement in Spearman correlation on average across ranges. In the challenging 2–6 range, it significantly outperformed greedy decoding.
  • Qwen-2.5 Family: Contrastive decoding (14B-3B) achieved an 11.7% relative improvement in Spearman correlation on average.
• Stability: While greedy decoding performance fluctuated wildly depending on the score range (e.g., Qwen-14B dropped from ~0.47 Spearman in 1–5 to ~0.31 in 2–6), contrastive decoding maintained stable, high correlations across all ranges.
• Assistant Model Selection:
  • Smaller assistants (e.g., 3B) generally worked better than larger ones (e.g., 7B).
  • Using a larger assistant (14B-7B) degraded performance, particularly in the 1–5 range, suggesting that larger assistants may penalize correct logits from the main model rather than just the bias.
• Parsing Reliability: Contrastive decoding also improved the rate of valid numeric output, reducing parsing failures significantly compared to smaller greedy models.

5. Significance and Implications

• Reliability in Evaluation: This work addresses a critical flaw in automated evaluation pipelines. If an LLM judge's performance depends on the arbitrary choice of a score range (0–4 vs. 1–5), benchmarking results are unreliable. This method makes evaluations robust to such hyperparameter choices.
• Cost-Effective Scaling: The method leverages smaller, cheaper models (assistants) to improve the performance of larger, more expensive models (main judges), offering a practical way to enhance evaluation quality without training new models.
• Beyond 1–5 Scales: The findings suggest that the standard 1–5 scale is not inherently "correct" but rather a convention that some models happen to align with. Contrastive decoding unlocks the potential to use diverse scoring ranges tailored to specific application needs without sacrificing accuracy.
• Limitations: The approach increases inference compute (requiring two forward passes), though this can be mitigated if speculative decoding is already in use. Experiments were limited to English and summarization tasks, though the bias mechanism is likely generalizable.

In conclusion, the paper provides a rigorous analysis of a subtle but impactful bias in LLM judges and offers a simple, effective decoding strategy to neutralize it, significantly improving the reliability of automated evaluation systems.