Skip to content

AgentFairBench: Do LLM Agents Discriminate When They Act?

Jun 2026 · IEEE Access · Vol 14, pp. 148414-148440 · 2 citations · 57 references
Computer Science

TL;DR

It is proved that comparing a $k$ -group spread against a two-run noise difference reports an inflation of about 2.25 at six groups, when the true disparity is zero, and given the arity-matched correction and exact randomization test that repair it.

Abstract

Language model agents increasingly act rather than answer: screening applicants, recommending credit, triaging patients. Fairness is still measured by grading text. We make three contributions. First, a correction to a method: we prove that comparing a $k$ -group spread against a two-run noise difference reports an inflation of $d_{2}(k)/d_{2}(2)$ , about 2.25 at six groups, when the true disparity is zero, and give the arity-matched correction and exact randomization test that repair it. Second, an instrument: AgentFairBench, a reproducible benchmark evaluating demographically empty profiles in matched sets varying only a name-coded race-by-gender signal, across hiring, lending and triage, under four conditions of increasing scaffold depth, each a single model call rather than a multi-turn pipeline, plus a length-matched control, released with a held-out split, a canary and tiered verification. Third, a measurement: across four models, two vendors and 7909 replicated decisions, the two instruments diverge, which is the point. Two open-weights cells, in different domains, have ratios clearing the corrected floor while their randomization tests stay null: wide spreads with no consistent group ordering, on floors too unstable to lean on. Four hiring cells reject exchangeability without clearing the floor. That effect is small, 0.16 to 0.31 of a call-to-call noise standard deviation, and runs opposite to the usual audit direction: white-male-coded names score lowest in every surviving cell. A shift too small to widen a spread can break exchangeability; a spread too wide to be noise can carry no ordering. We offer an instrument and a correction, not a general finding.

Read PDF

Similar papers

#artificial intelligence Preprint Jul 2024

The Llama 3 Herd of Models

It is found that Llama 3 delivers comparable quality to leading language models such as GPT-4 on a plethora of tasks, and performs competitively with the state-of-the-art on image, video, and speech recognition tasks.

Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri et al. · 21,082 citations · ⚡3,832
#artificial intelligence Conference Open access Jun 2018

Gibson Env: Real-World Perception for Embodied Agents

This paper investigates developing real-world perception for active agents, proposes Gibson Environment for this purpose, and showcases a set of perceptual tasks learned therein.

F. Xia, Amir Zamir, Zhi-Yang He et al. · 1,069 citations · ⚡112
#artificial intelligence Open access May 2023

Evaluating the Performance of Large Language Models on GAOKAO Benchmark

GAOKAO-Bench is introduced, an intuitive benchmark that employs questions from the Chinese GAOKAO examination as test samples, including both subjective and objective questions that contribute a robust evaluation benchmark for future large language models and offers valuable insights into the advantages and limitations...

Xiaotian Zhang, Chun-yan Li, Yi Zong et al. · 216 citations · ⚡17

KTO: Model Alignment as Prospect Theoretic Optimization

This work proposes a HALO that directly maximizes the utility of generations instead of maximizing the log-likelihood of preferences, and suggests that the best loss depends on the inductive biases most appropriate for a given setting, an oft-overlooked consideration.

Kawin Ethayarajh, Winnie Xu, Niklas Muennighoff et al. · 195 citations · ⚡28
#artificial intelligence Open access Jul 2024

Gender, Race, and Intersectional Bias in Resume Screening via Language Model Retrieval

This work investigates the possibilities of using LLMs in a resume screening setting via a document retrieval framework that simulates job candidate selection and finds that the MTEs are biased, significantly favoring White-associated names in 85% of cases and female-associated names in only 11.1% of cases.

Kyra Wilson, Aylin Caliskan · 131 citations · ⚡8

Related blog posts

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.