TLDR: A new research paper investigates the inconsistent correlation between LLMs’ answer generation and judgment abilities. It finds that under standard methods, these capabilities are only weakly correlated due to LLMs’ sensitivity to the evaluated response and a bias towards predicting ‘Correct.’ The paper proposes a ‘self-reference-guided evaluation’ strategy, where an LLM uses its own generated answer as a reference. This approach significantly strengthens the correlation between generation and judgment, making generation performance a reliable proxy for judgment quality, especially when external references are unavailable.
Large Language Models (LLMs) are increasingly used not just to generate text, but also to evaluate other AI models. This approach, often called ‘LLM-as-Judge,’ has become quite popular. However, a critical question has lingered: how well does an LLM’s ability to create answers (generation) truly align with its ability to assess them (judgment)? Recent research has presented conflicting views, with some studies suggesting a strong link and others finding little correlation.
A new research paper, titled Do Before You Judge: Self-Reference as a Pathway to Better LLM Evaluation, delves deep into this inconsistency. The authors, Wei-Hsiang Lin, Sheng-Lun Wei, Hen-Hsen Huang, and Hsin-Hsi Chen, conducted a comprehensive study across 11 different LLMs and 21 diverse tasks to understand this relationship better. Their findings shed light on why these capabilities often seem disconnected and propose a novel solution to bring them into alignment.
The Initial Disconnect Between Generation and Judgment
Initially, the study observed a positive correlation at a high level: models that were better at generating answers generally also performed better at judging them. This seemed to support the idea that strong generators make strong judges. However, a more detailed analysis revealed a subtle but significant issue. The researchers found that LLMs often exhibit a strong bias towards predicting answers as ‘Correct.’ This means that a judge model might appear to perform well simply because it’s more likely to say an answer is correct, rather than truly understanding and evaluating it based on its own knowledge.
When the team looked at individual instances of judgment, controlling for whether the answer being judged was actually correct, the picture changed dramatically. They found that an LLM’s own ability to generate a correct answer was only weakly correlated with its ability to accurately judge another model’s answer. This suggests that, under standard evaluation methods, being a good generator doesn’t automatically make an LLM a good, unbiased judge.
Introducing Self-Reference-Guided Evaluation
To address this disconnect, the researchers proposed a new strategy: ‘self-reference-guided evaluation.’ Instead of relying on external ‘gold standard’ answers or outputs from even stronger models as references (which can be costly or unavailable), this method leverages the judge model’s *own* generated answer as a reference during the evaluation process. Essentially, the LLM is asked to generate its own best answer to a question first, and then use that self-generated answer as a benchmark when evaluating another model’s response.
A Stronger Link: The Impact of Self-Reference
The results of this self-reference-guided approach were striking. The correlation between an LLM’s answer generation ability and its judgment ability significantly strengthened, with an average increase of 0.35 across all evaluated cases. This means that when an LLM uses its own answer as a reference, its generation performance becomes a much more reliable indicator of its judgment quality.
This method proved particularly effective in helping models identify *incorrect* answers, especially when the judge model itself possessed the correct knowledge. It shifted the evaluation dependency from merely the correctness of the agent’s answer to the judge’s own internal knowledge base, transforming the task into a more robust verification process.
Also Read:
- The Hidden Flaws in AI Evaluation: Why LLM Judge Benchmarks Need a Rethink
- Unlocking Dynamic Problem-Solving in AI with Explanatory Verifiers
Practical Implications and Future Directions
The findings have important practical implications for anyone using LLMs for evaluation. When high-quality external references are scarce, selecting a judge model based on its generation performance, combined with this self-reference-guided strategy, can lead to much more reliable evaluations. It offers a practical way to ensure that as LLMs continue to evolve, their generation and judgment capabilities remain aligned.
The study acknowledges some limitations, including the rapidly evolving nature of LLMs, the focus on pointwise answer judgment tasks (rather than more complex formats like pairwise comparisons), and the absence of multi-turn interaction analysis. There’s also the potential for error propagation if the judge model’s self-generated reference is itself incorrect. However, the research strongly recommends applying this method with judge models that already demonstrate high generation accuracy in the specific domain, mitigating this risk.
In essence, this research provides a valuable pathway to making LLM-as-Judge frameworks more robust and trustworthy, by enabling models to ‘know themselves’ better before they judge others.
