TLDR: A new research paper identifies and mitigates ‘score range bias’ in Large Language Models (LLMs) used as evaluators. This bias causes LLM judge outputs to be overly sensitive to the defined scoring scale. By employing a technique called contrastive decoding, which leverages similar biases across models from the same family, the researchers achieved up to an 11.3% relative improvement in correlation with human judgments, making LLM evaluations more robust and reliable across different score ranges.
Large Language Models (LLMs) are increasingly used as automated evaluators, or “judges,” in various applications, offering a scalable and cost-effective way to assess outputs. However, ensuring the reliability of these AI judges remains a significant challenge. A recent research paper by Yoshinari Fujinuma from Cantina Labs sheds light on a newly identified issue: “score range bias” in LLM-as-a-judge systems. This bias means that an LLM’s assigned scores are highly sensitive to the specific numerical range provided for evaluation, hindering the search for optimal scoring scales.
The paper, titled Contrastive Decoding Mitigates Score Range Bias in LLM-as-a-Judge, reveals that this score range bias is not isolated but is consistently observed across different LLM families and sizes, such as Llama-3 and Qwen-2.5. For instance, Llama family models might disproportionately favor a score of 4, while Qwen models might lean towards a score of 2, regardless of the actual quality of the content being judged. This inherent sensitivity makes it difficult to trust LLMs for direct assessment tasks, where they assign scores without external references.
Understanding the Bias and Its Mitigation
The core of the problem lies in how LLMs interpret and assign scores within a given range. The research found that even models from the same family exhibit similar patterns of this score range bias. To address this, the paper proposes a mitigation strategy using “contrastive decoding.” This technique modifies the LLM’s output by involving two models: a main model and an assistant model. The final adjusted score is calculated by subtracting a weighted probability from the assistant model’s output from the main model’s output. This process effectively cancels out similar biases encoded across models from the same family.
Experiments focused on the summarization task, a common application for LLM judges. The researchers used the SummEval dataset and evaluated LLM performance across different 5-point Likert scales (e.g., 0-4, 1-5, 2-6, 3-7). The results were compelling: contrastive decoding significantly improved the correlation with human judgments. On average, it achieved up to an 11.3% relative improvement in Spearman correlation, a key metric for measuring rank agreement, across various score ranges. This demonstrates that contrastive decoding not only reduces the bias but also makes the LLM judge’s evaluations more consistent and robust, regardless of the specific scoring scale used.
Also Read:
- LLMs: Why They Know the Answer But Still Take Shortcuts
- Assessing AI Judges: A New Benchmark for Web Development Quality
Implications and Future Directions
The success of contrastive decoding in mitigating score range bias is a crucial step towards making LLM judges more reliable and trustworthy. It suggests that by carefully designing the decoding process, we can overcome some of the inherent limitations of LLMs when used for direct assessment. This robustness opens up possibilities for exploring and utilizing optimal score ranges beyond the traditional 1-5 scale, which was previously problematic due to the bias.
While the findings are promising, the research acknowledges certain limitations. The experiments were conducted on models up to 14 billion parameters and primarily focused on the English language and summarization tasks. Future work could involve scaling to larger models, expanding to other languages, and applying the technique to a broader range of evaluation tasks. Nevertheless, this study provides a valuable framework for enhancing the alignment of LLM judges with human evaluators, paving the way for more accurate and dependable AI-powered assessment systems.
