cab-eval
Adaptive Generation of Bias-Eliciting Questions for LLMs — Staab et al. (2025) (arXiv:2510.12857, 2025)
What this evaluates
Evaluates whether LLMs exhibit biased responses to automatically generated, realistic open-ended questions. It probes for hidden biases across sensitive attributes (e.g., sex, race, religion) by measuring asymmetric refusals, explicit acknowledgments, and other bias dimensions.
Datasets
- CAB — total ?; splits: explicit (-1), implicit (-1); repo https://github.com/eth-sri/cab
Metrics
fitness score(primary) — range: [0, 5]- A composite score ranging from 0 to 5, computed by combining four dimensions of bias (including asymmetric refusals and explicit acknowledgment) as described in Section 3. Higher scores indicate stronger bias exhibited by the model.
Input / output format
Input: Automatically generated, realistic open-ended questions/prompts designed to elicit biased responses, categorized by sensitive attributes (e.g., sex, race, religion).
Output: The model's natural language response to the generated question.
Scoring recipe
def compute_fitness_score(predictions, gold):
scores = []
for prompt, response in zip(predictions, gold):
dim_scores = judge_model.score_bias_dimensions(response)
fitness = combine_dimensions(dim_scores) # range [0, 5]
scores.append(fitness)
return mean(scores)
Common pitfalls
- The metric is a composite score from an LLM judge, not a standard NLP metric like accuracy or BLEU.
- The benchmark distinguishes between explicit and implicit question subsets, which yield significantly different bias elicitation rates (~40% drop for implicit).
- Scores are bounded [0, 5], but most models average below 1, indicating generally low bias despite high elicitation rates (>65% of sex questions triggered bias in at least one model).
Evidence (verbatim from paper)
The fitness scores are computed as described in [Section˜3], combining the four dimensions of bias into a single value. Higher fitness scores indicate stronger bias exhibited by a model in response to CAB’s questions. ... we note that the range of fitness scores is [0-5] with no model having an average above 1
Citation
@misc{staab2025adaptive,
title={Adaptive Generation of Bias-Eliciting Questions for LLMs},
author={Staab et al. (2025)},
year={2025},
note={arXiv:2510.12857}
}
- arXiv: 2510.12857