Beyond Aggregate Scores: Behavioral Correctness Ass... | AI Research

Beyond Aggregate Scores: Behavioral Correctness Assumptions for Assessing Reference-Based Automatic Evaluation Methods
This paper introduces a new diagnostic framework to better understand how automated systems evaluate natural language generation. Currently, most evaluation methods are judged by their "aggregate performance"—essentially how well they correlate with human scores on a benchmark. The authors argue that this approach is insufficient because it hides how these evaluators actually behave when faced with specific, controlled changes in a response. By using a set of "behavioral correctness assumptions," the researchers provide a way to test whether an evaluator correctly identifies, ignores, or penalizes specific types of changes, such as paraphrasing, factual errors, or logical contradictions.

A New Way to Test Evaluators

The researchers propose a taxonomy of "correctness assumptions" to categorize how an ideal evaluator should react to different types of text transformations. These are divided into two main groups:

• Correctness-Preserving: These transformations change the style, length, or structure of a response without changing its factual accuracy. A reliable evaluator should ideally assign these responses a similar score to the original.

• Correctness-Altering: These transformations introduce errors, such as hallucinations, logical contradictions, or missing facts. A reliable evaluator should be sensitive to these and assign a lower score. To see meta in practice, How to Make Cinematic Commercials walks through a concrete example.
  By operationalizing these assumptions through a test suite of over 2,800 transformed responses, the authors can see exactly where different evaluation methods—ranging from simple lexical overlap metrics to complex LLM-based judges—succeed or fail.

Key Findings on Evaluator Behavior

The study reveals that no single evaluation method satisfies all the proposed correctness assumptions. Even evaluators that perform similarly on traditional benchmarks often show vastly different "behavioral profiles." For example, while some metrics are sensitive to factual errors, they may also unfairly penalize responses simply because they are longer (verbosity). The meta story also surfaces in AI Agents Going Rogue Renew Calls..., adding another angle.
One of the most significant findings is that recognizing "alternative correct answers" remains a major challenge across all paradigms. Many evaluators consistently give lower scores to perfectly valid responses if they are phrased differently than the reference answer. Additionally, the researchers found that even when evaluators detect a violation, they often fail to penalize it strongly enough, leaving the final scores higher than they should be for incorrect content.

Stability and Sensitivity

To provide a complete picture, the framework measures two core traits: stability and sensitivity. Stability tracks how consistently an evaluator maintains scores when the content remains correct but the wording changes. Sensitivity tracks how effectively the evaluator detects and penalizes genuine quality degradations. The same ai evaluation question is explored in Wrong Prediction, Right Answer, which adds a research perspective.
The researchers also conducted robustness checks to ensure their findings were reliable. They tested for repeat-run variability (to see if stochastic LLM evaluators give consistent results), configuration sensitivity (to see if changing model settings alters the outcome), and test suite reproducibility. These tests confirmed that the framework provides a consistent and diagnostic way to characterize how different evaluation methods actually function under the hood. as detailed in the full paper on Arxiv