Reproduce “Test-Time Fairness and Robustness in Large Language Models” on OpenResearch

About this paper

Leonardo Cotta, Chris J. Maddison

Researchers from the Vector Institute and University of Toronto introduced a formal causal framework for test-time fairness and robustness in Large Language Models, defining "stratified invariance" and proposing "Out-Of-Context (OOC) prompting" as a zero-shot method for its implementation. OOC prompting consistently reduced bias, improving stratified invariance in 23 out of 27 task/model combinations, while largely preserving predictive performance across various LLM scales.

Key insights

• A naive application of standard causal debiasing techniques, such as counterfactual data augmentation, generally fails to achieve individual-level debiased predictions at test time under common assumptions.
• Stratified invariance provides a flexible and explicit way to define and enforce a spectrum of debiasing requirements, ranging from population-level to individual-level guarantees, measurable from observational data.
• LLMs can be guided to simulate complex causal interventions, specifically stratified counterfactual data augmentation, through explicit, two-step prompting, moving beyond unreliable implicit safety instructions.

Summary by alphaXiv.