BiasBusters: Uncovering and Mitigating Tool Selection Bias in Large Language Models
Thierry Blankenstein, Jialin Yu, Zixuan Li, Vassilis Plachouras, Sunando Sengupta, Philip Torr, Yarin Gal, Alasdair Paren, Adel Bibi
TL;DR
This work tackles tool-selection bias in tool-augmented LLMs, where models may favor functionally equivalent APIs due to superficial cues rather than utility. It introduces a benchmark of 10 clusters of interchangeable APIs and total-variation based metrics to quantify bias across seven models, revealing that semantic alignment between queries and tool descriptions is the primary predictor and that metadata perturbations and biased pre-training amplify bias. The authors dissect bias origins through attribute correlations, perturbations, and biased CPT, showing imageable effects from description content while CPT accelerates bias but does not solely drive it. To counteract bias, they propose a lightweight debiasing pipeline that filters candidates to a relevant subset and selects uniformly from that subset, which substantially reduces bias metrics while preserving task coverage. The findings underscore a practical need for fair, reliable tool-calling in LLM systems and provide readily adoptable resources and methods to improve marketplace fairness and user experience.
Abstract
Agents backed by large language models (LLMs) often rely on external tools drawn from marketplaces where multiple providers offer functionally equivalent options. This raises a critical point concerning fairness: if selection is systematically biased, it can degrade user experience and distort competition by privileging some providers over others. We introduce a benchmark of diverse tool categories, each containing multiple functionally equivalent tools, to evaluate tool-selection bias. Using this benchmark, we test seven models and show that unfairness exists with models either fixating on a single provider or disproportionately preferring earlier-listed tools in context. To investigate the origins of this bias, we conduct controlled experiments examining tool features, metadata (name, description, parameters), and pre-training exposure. We find that: (1) semantic alignment between queries and metadata is the strongest predictor of choice; (2) perturbing descriptions significantly shifts selections; and (3) repeated pre-training exposure to a single endpoint amplifies bias. Finally, we propose a lightweight mitigation that first filters the candidate tools to a relevant subset and then samples uniformly, reducing bias while preserving good task coverage. Our findings highlight tool-selection bias as a key obstacle for the fair deployment of tool-augmented LLMs.