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

AI summary

Overview

• Research area: Evaluation methodology for large language models, specifically the "LLM-as-a-judge" paradigm for direct assessment of summarization quality, and contrastive decoding as a bias-mitigation technique.
• Technical level: Intermediate. Readers should be comfortable with correlation metrics (Pearson, Spearman, Kendall), logit-level decoding, and instruct-tuned model families.
• Scope: The paper documents a previously unexamined bias ("score range bias") in LLM judges, shows it is shared across model sizes within a family, and demonstrates that contrastive decoding reduces it, reporting up to 11.7% average relative improvement in Spearman correlation with human judgments across score ranges.

What This Paper Is About

When an LLM is asked to rate a summary on a 5-point scale, its agreement with human raters changes depending on how that scale is written — for example, whether the options are 0-4, 1-5, 2-6, or 3-7 — even though all four represent the same 5-point scale. The paper names this phenomenon score range bias, shows it appears across Llama-3 and Qwen-2.5 model families and sizes, and tests whether contrastive decoding (subtracting the logits of a smaller "assistant" model from a larger "main" model) can cancel the bias out, since models in the same family appear to share it.

Key Contributions

1. The paper identifies and characterizes score range bias in LLM judges used for direct assessment: judge outputs are highly sensitive to the pre-defined score range, and correlations with human judgments shift accordingly, across different model sizes and across the Llama-3 and Qwen-2.5 families.
2. It shows that similar score range biases are shared across models from the same family, both in the distribution of emitted scores and in the first-output-token logit distribution, with the bias toward a particular score shrinking but persisting as model size grows from 3B to 14B.
3. It adapts contrastive decoding with an added scaling hyperparameter (λ) to align logit distributions between the main and assistant models, and demonstrates this as a mitigation that produces more stable correlations across score ranges.
4. It reports the mitigation effect as up to 11.7% relative improvement on average in Spearman correlation with human judgments across different score ranges, with additional results on a second dataset (BigGen-Bench) and across three summarization dimensions (coherence, relevance, consistency).

Main Findings

• Score range bias is real and large: For Llama-3 and Qwen-2.5 judges on summarization, correlation with human annotations varies substantially by range. For example, Qwen2.5-14B achieves Spearman .452 on the 0-4 range but only .313 on 2-6, and Llama 3.1-8B achieves .426 on 3-7 versus .257 on 2-6. Qwen2.5-3B and Llama 3.2-1B collapse to near-zero values in some ranges (Llama 3.2-1B: .000 on both 1-5 and 2-6).

• Skewed score distributions: In the 2-6 range, greedy decoding from Llama 8B and 3B models is highly skewed toward emitting a score of 4, while Qwen2.5 3B and 7B models skew toward a score of 2. Contrastive decoding moves these distributions closer to the human annotation distribution.

• Biases are shared within a family at the logit level: In the 0-4 range, the first-output-token logit for Score 2 is highest for Qwen2.5 3B, 7B, and 14B, while the most frequent human annotation is Score 3. The bias shrinks as model size increases from 3B to 14B but remains at 14B.

• Logit magnitudes differ across sizes: Maximum logits are approximately 25 for the 3B model, approximately 30 for the 7B model, and approximately 34 for the 14B model. This motivates including the λ hyperparameter to align logit distributions between main and assistant models.

• Contrastive decoding improves average correlation: Averaged across all score ranges on coherence, Llama 3.1-8B rises from Spearman .344 to .375 with the 1B assistant (a reported 6.7% relative improvement, .330 → .352) and Qwen2.5-14B rises from .392 to .438 with the 3B assistant (a reported 11.7% relative improvement). The paper states the maximal improvement is observed in the 2-6 score range.

• Larger assistant models can hurt: With Llama-3 assistants, the 1B assistant reaches an average Spearman of .375 versus .352 for the 3B assistant. In an ablation using Qwen2.5-7B as the assistant for the Qwen2.5-14B main model, the 1-5 range degrades sharply (Spearman .282 versus .470 with the 3B assistant), which the paper interprets as a larger assistant penalizing correct logits from the main model.

• The 1-5 range is not universally best for greedy decoding: For greedy decoding, Qwen-2.5 family models and Llama 3.2-3B show their highest correlation in the 1-5 range (7B, 1-5: Spearman .370; 14B, 1-5: .468), while Llama 3.1-8B is the exception, with its highest correlation in the 3-7 range.

• Invalid output rates differ sharply by model: Qwen2.5-3B produces roughly 98% invalid outputs across all score ranges (98.2%, 98.7%, 97.6%, 97.6%), while Qwen2.5-7B produces 0% and Qwen2.5-14B produces 0% except for 4.6% below-range outputs in the 3-7 range. Llama 3.2-3B fails on 17.4% to 24.0% of outputs, while Llama 3.1-8B has 0% invalid outputs in all ranges. Contrastive decoding on Llama yields 0.3% invalid outputs in the 2-6 range.

• Results extend to a second dataset: On BigGen-Bench, Llama 3.1-8B with the 1B assistant improves average Pearson from .372 to .376, Spearman from .349 to .352, and Kendall from .294 to .297. Qwen2.5-14B with the 3B assistant improves Pearson from .532 to .550, Spearman from .500 to .517, and Kendall from .435 to .449.

• Note on a reported figure: For the 2-6 range, the text states a contrastive Pearson of .310 compared with .168 for Llama 3.2-3B and .270 for Llama 3.1-8B, whereas Table 1 lists .302 for Contra. (8B-3B), .008 for Llama 3.2-3B, and .262 for Llama 3.1-8B in that range. The two sets of numbers do not match.

Methodology in Plain English

The study uses direct assessment: an LLM judge reads a news article and a candidate summary and assigns a single numeric score. The evaluation task is coherence of the summary, with additional experiments on relevance and consistency. Each prompt template is instantiated with a different score range — 0-4, 1-5, 2-6, or 3-7 — but all ranges are 5-point Likert scales, drawn from a prompt used in prior work. If the judge's output cannot be parsed into a number, it is assigned the lowest score in the range; if a parsed number exceeds the maximum, it is clamped to the highest score in the range.

Judge outputs are compared with human annotation scores using Pearson, Spearman, and Kendall correlations, with 95% confidence intervals from bootstrap testing. The dataset is SummEval, containing 100 news articles, each paired with 16 summaries with human annotation scores, for 1600 summaries total; 10% of the news articles are held out as a development set for grid search.

To mitigate the bias, the authors apply contrastive decoding: instead of using only the main model's next-token probability, they subtract a temperature-scaled, λ-weighted assistant model probability, log p_main − λ log p_asst, where the assistant's probability is computed with temperature t. They add the λ term specifically to handle the differing logit magnitudes across model sizes in a family. Hyperparameters λ ∈ {0.01, 0.1, 0.5, 1.0} and t ∈ {0.5, 1.0, 2.0, 3.0, 4.0, 5.0} are selected by grid search on the development set, and the chosen values differ per main model, assistant model, score range, and evaluation dimension (tables of settings are provided in the appendix).

Models used: Llama-3.1-8B-Instruct as main with Llama-3.2-3B-Instruct and Llama-3.2-1B-Instruct as assistants; Qwen-2.5-14B-Instruct and Qwen-2.5-7B-Instruct as mains with Qwen-2.5-3B-Instruct as assistant. All experiments ran on NVIDIA A100 GPUs.

Why This Matters

Impact on research. The paper argues that score range is an uncontrolled variable in LLM-as-a-judge research: two studies using "the same 5-point scale" may obtain different correlations purely from how the scale is written. It also broadens the catalog of judge biases beyond self-enhancement bias and family enhancement bias, and connects bias analysis to contrastive decoding, which had primarily been used for other purposes.

Real-world applications:

• Evaluation and benchmarking pipelines: Teams that score model outputs with LLM judges need to know whether their reported numbers depend on an arbitrary prompt choice rather than actual quality.
• Training data generation: Preference data or reward signals produced by LLM judges inherit any range sensitivity, which can propagate into downstream model behavior.
• Summarization and content quality assurance: Automatic quality scoring for news summarization or similar content can shift systematically with the rubric's numeric range.
• Leaderboard and rubric design: The finding that the standard 1-5 range is not uniformly best motivates testing alternative ranges when building evaluation rubrics.

Industry relevance. Running two models per judgment increases test-time compute, but the paper points out that a main-plus-assistant model setup is already common in speculative decoding, so contrastive decoding can potentially be added without an extra forward pass in such deployments. The practical takeaway for teams shipping LLM judges is that mitigation is achievable at moderate cost.

Future Directions

• Explain assistant-model size effects. The paper notes its own ablation does not explain why smaller assistants sometimes outperform larger ones, beyond hypothesizing that larger assistants penalize correct main-model logits.
• Extend beyond 14B parameters. Experiments were limited to models up to 14B due to computational budget constraints, leaving open whether the bias and mitigation behave the same at larger scales.
• Broaden task and language coverage. The work covers summarization only, and only English; the authors state the experiments do not exploit English-specific linguistic knowledge but acknowledge the language limitation as a potential risk.
• Apply to judge-specialized models. The authors leave models finetuned specifically for judging, such as Prometheus, to future work, since multiple model sizes are not available for contrastive decoding and those models are tuned toward the 1-5 score range.
• Use the framework to expand beyond the 1-5 range. The paper proposes its score range bias analysis as a general test that can be applied to arbitrary models, potentially unlocking score ranges other than the standard 1-5.

Target Audience

Researchers and practitioners building or auditing LLM-as-a-judge evaluation systems, especially those working on summarization evaluation, reward modeling, or automatic quality scoring. It is also relevant to benchmarking teams who need to justify rubric design choices, and to anyone interested in decoding-time interventions for controlling model behavior. Readers without background in correlation metrics or token-level decoding will need to consult the cited prior work on contrastive decoding and on LLM judge evaluation first.

Authors’ 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. Focusing on summarization, 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…