Rethinking Language Model Evaluations: BT-sigma and the Jury Approach

Large language models (LLMs) have become the go-to for assessing natural language generation, but their evaluations often fall short. Here's why: consistency and bias issues plague the current methods. Typically, these evaluations rely on either single judges or an average of multiple ones, assuming all judges perform equally well. The reality is quite different.

The Problem with Current Evaluations

LLM judges aren't created equal. They vary widely in how they perform across different tasks and evaluation aspects. This inconsistency limits their effectiveness, especially direct, probability-based rankings. And to make matters worse, we often lack human-labeled data for calibrating these judges.

Here's what the benchmarks actually show: direct probability-based methods don't hold up under scrutiny. They fail to deliver accurate rankings because they don't account for the variable reliability of different judges. It's like stacking apples and oranges and calling them equal.

Introducing BT-sigma

Enter BT-sigma, a judge-aware extension of the Bradley-Terry model. This new approach introduces a discriminator parameter for each judge, aiming to infer both item rankings and judge reliability from just pairwise comparisons. By focusing on judge reliability, BT-sigma outperforms traditional averaging-based aggregation methods.

The architecture matters more than the parameter count here. BT-sigma's strength lies in its ability to function as an unsupervised calibration mechanism. It improves aggregation by understanding which judges can be trusted more, and which can't.

Why This Matters

The significance of BT-sigma can't be overstated. In experiments with benchmark NLG evaluation datasets, BT-sigma consistently outshone its predecessors. Not only did it deliver more accurate rankings, but it also showed a strong correlation with independent measures of LLM judgment consistency.

Why should you care? Because this is a step toward more reliable and unbiased evaluations of language models. In a field where accuracy is everything, relying on flawed evaluations isn't just inefficient, it's irresponsible. Isn't it time we expected more from our evaluators?

Strip away the marketing and you get a simple truth: better models need better evaluations. BT-sigma is a leap forward in ensuring that the models we rely on are judged fairly and accurately. The numbers tell a different story now, one where judge reliability takes center stage.