REQUAL-LM: Reliability and Equity through Aggregation in Large Language Models

TL;DR

REQUAL-LM addresses the dual challenge of reliability and equity in large language models by adopting a Monte Carlo aggregation framework. It samples $m$ outputs from an LLM, embeds them into a semantic space, and forms a centroid $\vec{v}_c$ to identify outputs near the distribution mean; an equity-aware weighting scheme further yields a weighted centroid to minimize harmful bias. The approach is a black-box wrapper requiring no retraining or internal model access, and it supports non-binary demographic groups while formalizing reliability $\rho(O)=\mathcal{S}_{im}(\vec{v}(O),\vec{\mu}_\xi)$ and bias $\beta(O)$. Empirical results across subset selection, chat completion, and masked language prediction show that Weighted Output reduces harmful bias with minimal or negligible loss in reliability across diverse datasets (StereoSet, WinoBias, Forbes Billionaire, Students), demonstrating scalability and deployment-readiness for societally impactful applications.

Abstract

The extensive scope of large language models (LLMs) across various domains underscores the critical importance of responsibility in their application, beyond natural language processing. In particular, the randomized nature of LLMs, coupled with inherent biases and historical stereotypes in data, raises critical concerns regarding reliability and equity. Addressing these challenges are necessary before using LLMs for applications with societal impact. Towards addressing this gap, we introduce REQUAL-LM, a novel method for finding reliable and equitable LLM outputs through aggregation. Specifically, we develop a Monte Carlo method based on repeated sampling to find a reliable output close to the mean of the underlying distribution of possible outputs. We formally define the terms such as reliability and bias, and design an equity-aware aggregation to minimize harmful bias while finding a highly reliable output. REQUAL-LM does not require specialized hardware, does not impose a significant computing load, and uses LLMs as a blackbox. This design choice enables seamless scalability alongside the rapid advancement of LLM technologies. Our system does not require retraining the LLMs, which makes it deployment ready and easy to adapt. Our comprehensive experiments using various tasks and datasets demonstrate that REQUAL- LM effectively mitigates bias and selects a more equitable response, specifically the outputs that properly represents minority groups.

Figures (8)

• Figure 1: System Architecture of Requal-lm.
• Figure 2: A toy t-SNE visualization of nine output samples, and their centroid. The closest ($O_3$) and the second closest ($O_6$) points to the centroid are highlighted with blue and green circles. The green-to-red color code shows the bias values.
• Figure 3: Each figure demonstrates the bias distribution of final outputs on the specified task and dataset.
• Figure 4: Comparing the (gender) bias distributions on subset selection.
• Figure 5: Reliability values, for the subset selection, chat completion, and co-reference resolution tasks.

Theorems & Definitions (2)

• Definition 1: Reliability
• Definition 2: Bias