Rank-Allocational Bias Index (RABBI)

• Rank-Allocational-Based Bias Index (RABBI) is a model-agnostic metric that quantifies pairwise preference imbalances between groups in top-k allocation settings.
• It directly compares cross-group scores to estimate bias, aligning its predictions closely with realized selection disparities in applications like recruitment and academic assessment.
• Empirical studies reveal that RABBI outperforms traditional fairness metrics, aiding in robust model auditing and informed decision-making in resource-limited contexts.

The Rank-Allocational-Based Bias Index (RABBI) is a model-agnostic metric designed to quantify allocational harms arising from biases in LLM predictions, particularly in high-stakes, resource-constrained decision-making such as recruitment and academic assessment. Unlike traditional fairness metrics that evaluate model output distributions or average group-level score gaps, RABBI directly estimates the pairwise preference imbalance between groups and exhibits a strong alignment with realized selection disparities under top-k allocation scenarios (Chen et al., 2024).

1. Motivation and Conceptual Basis

Prevailing fairness metrics in NLP, such as average score gap ($\Delta$), Jensen-Shannon divergence (JSD), and Earth Mover’s Distance (EMD), center on global distributional differences or marginal mean performance between protected groups. However, these approaches are agnostic to the selection mechanism—the process by which limited opportunities (e.g., a fixed number $k$ of interview or feedback slots) are allocated to candidates based on model-generated scores.

Frequently observed phenomena include near-zero average score gaps even when systematic group disadvantage manifests in top-k selection events, and between-group distributional metrics failing to capture disparities in the "allocation tail." Classic statistical fairness criteria (demographic parity, equal opportunity) assess realized selection differences but do not serve as advance audit tools to predict potential harms. RABBI was proposed to fill this methodological gap. It operationalizes allocational fairness as the normalized probability that a random candidate from one (protected) group outcompetes a random candidate from a reference group, less the converse—providing a direct link to real-world allocation outcomes.

2. Formal Definition and Interpretation

Let $G$ denote a set of demographic groups, with protected group $A$ and reference group $B$. Let $D_x$ be the candidate set and $n_x = |D_x|$ its cardinality for group $x \in \{A, B\}$. Let $s_\theta(x)$ be the score assigned to candidate $x$ by model $k$0 with parameters $k$1. The indicator function $k$2 outputs 1 if its argument is true, else 0.

The two-group pairwise RABBI is defined as: $k$3

• $k$4: Model’s scalar assessment score (higher = more qualified)
• $k$5: 1 iff $k$6 is strictly ranked above $k$7
• $k$8: Cross-group pair normalization

Interpretation:

• $k$9: No systematic preference between $G$0 and $G$1
• $G$2: Every $G$3 beats every $G$4
• $G$5: Every $G$6 beats every $G$7

A multi-group extension aggregates across all protected groups $G$8 using nonnegative weights $G$9 (e.g., uniform, group-size-weighted), producing: $A$0 This summary score enables worst-case or average bias reporting across multiple groups (Chen et al., 2024).

3. Computation and Implementation

RABBI computation entails evaluating all pairwise cross-group score comparisons, with explicit handling of ties. The procedure is as follows:

$n_x = |D_x|$6

In the event of strict ties, each tie increments both “positive” and “negative” pair counts by 0.5. Zero-sum tie handling (ignoring ties) is also possible. For intersectional or large group settings, computational optimizations or subsampling are recommended due to $A$1 complexity.

4. Toy Example

Consider two candidates in group $A$2 with scores $A$3 and two candidates in group $A$4 with scores $A$5.

Pair	Outcome	pos	neg	tie
(0.8, 0.7)	A wins	1	0	0
(0.8, 0.5)	A wins	2	0	0
(0.6, 0.7)	B wins	2	1	0
(0.6, 0.5)	A wins	3	1	0

There are $A$6 cross-group pairs, with 3 positive outcomes (A wins), 1 negative (B wins), and no ties. Thus,

$A$7

meaning a randomly selected $A$8 beats a randomly selected $A$9 75% of the time, with a bias magnitude of $B$0.

5. Comparative Empirical Performance

Chen et al. systematically compared RABBI against three metrics—$B$1 (score gap), JSD (distributional divergence), and EMD (transport-based divergence)—on two LLM-aided allocation tasks: resume screening for four job types and human-rated essay grading (L1 vs. L2 speakers across 10 countries), using demographically balanced candidate pools and top-k selection ($B$2). Metrics evaluated included observed allocation gaps (demographic parity/ADP, equal opportunity/AEO) and the computed bias indices on raw model scores.

Empirical results showed:

• In resume screening ($B$3; ADP):
  • $B$4: $B$5 (Pearson correlation with true allocation gap)
  • JSD: $B$6
  • EMD: $B$7
  • RABBI: $B$8 ($B$9)
• In essay grading ($D_x$0; AEO):
  • $D_x$1: $D_x$2
  • JSD: $D_x$3
  • EMD: $D_x$4
  • RABBI: $D_x$5

RABBI consistently tracked allocation disparities far more closely than average-score or distributional metrics, with correlation above $D_x$6 on all tasks and top-k settings, while baselines often hovered near zero or fluctuated by group (Chen et al., 2024).

6. Role in Model Selection and Auditing Practice

When selecting among multiple candidate models ($D_x$7) for deployment in contexts with limited slots, auditors can compute RABBI against the reference group for each model and prioritize models with the lowest absolute RABBI values ($D_x$8). Empirical results demonstrated that RABBI-based rankings achieved normalized discounted cumulative gain (NDCG@10) of at least $D_x$9 relative to the ranking based on realized ADP/AEO, whereas rankings using $n_x = |D_x|$0, JSD, or EMD often had NDCG less than $n_x = |D_x|$1 and occasionally mis-ranked strongly biased models as least biased.

The recommended auditing practice is to select bias metrics that mirror the intended allocation mechanism (i.e., RABBI for top-$n_x = |D_x|$2 selection). Auditors are encouraged to assess variability across random candidate pool draws and to use pairwise ranking prompts where supported. RABBI applies equally to pointwise score assignments.

7. Limitations and Future Directions

RABBI, while directly predictive of allocation disparities, introduces several practical considerations:

• Computational Complexity: Direct computation is quadratic in group sizes. For large pools, approximation or subsampling is required.
• Tie and Consistency Issues: LLMs producing more than $n_x = |D_x|$3 tie/inconsistency rates are excluded, as inconsistent pairwise judgments undermine metric reliability.
• Weighting in Multi-Group Scenarios: The choice of aggregation weights ($n_x = |D_x|$4) in the multi-group RABBI impacts the summary statistic and may require policy input.
• Static Test Pool Limitation: RABBI is evaluated on a fixed test pool, yet actual deployments may encounter pool drift.
• Attribute Limitation: RABBI is defined for single protected attributes at a time. Intersectional group comparisons scale quadratically in complexity.

Future research directions outlined by Chen et al. include developing faster randomized algorithms for RABBI estimation; drawing theoretical links to other allocational-fairness measures under strategic selection; generalizing to joint allocation of multiple scarce resources; and designing debiasing interventions aimed directly at reducing RABBI as opposed to marginal score calibration.

RABBI provides a direct, model-agnostic quantification of allocational bias that closely approximates actual group-level disparities when model outputs are used for top-$n_x = |D_x|$5 selection, filling a critical gap left by average-score and distributional metrics in fairness analysis for resource-constrained, high-stakes applications (Chen et al., 2024).