Improving Diversity of Demographic Representation in Large Language Models via Collective-Critiques and Self-Voting

TL;DR

This work formalizes diversity of representation in open-ended LLM generations and introduces two evaluation datasets with entropy and max-gap metrics to quantify diversity. The authors propose a zero-shot prompting framework, collective-critique and self-voting (CCSV), which leverages in-context reasoning, self-critique, and majority voting to significantly improve demographic diversity in responses, outperforming multiple baselines including Constitutional AI. Across automated and human evaluations, CCSV achieves large diversity gains (e.g., $entropy$ improvements and high human SxS diversity scores) and generalizes to culture-diversity tasks while remaining robust to user-specified constraints. The approach emphasizes robustness, generalization, and efficiency (fewer iterations than naive multi-step reasoning), with broader implications for responsible AI and open-ended generation tasks, while also noting computational costs and ethical considerations.

Abstract

A crucial challenge for generative large language models (LLMs) is diversity: when a user's prompt is under-specified, models may follow implicit assumptions while generating a response, which may result in homogenization of the responses, as well as certain demographic groups being under-represented or even erased from the generated responses. In this paper, we formalize diversity of representation in generative LLMs. We present evaluation datasets and propose metrics to measure diversity in generated responses along people and culture axes. We find that LLMs understand the notion of diversity, and that they can reason and critique their own responses for that goal. This finding motivated a new prompting technique called collective-critique and self-voting (CCSV) to self-improve people diversity of LLMs by tapping into its diversity reasoning capabilities, without relying on handcrafted examples or prompt tuning. Extensive empirical experiments with both human and automated evaluations show that our proposed approach is effective at improving people and culture diversity, and outperforms all baseline methods by a large margin.

Figures (12)

• Figure 1: Baseline performance of Flan-PaLM 540B model on our people-diversity dataset is highly non-diverse with average entropy close to 0 across prompts covering 105 professions.
• Figure 2: Proposed approach: Collective-critiques and self-voting (CCSV) prompts and technique.
• Figure 3: Robustness of methods on being able to diversify while satisfying user-specified group constraints.
• Figure 4: Ablation study comparing variants of CCSV.
• Figure 5: Illustration of proposed approach visualized in \ref{['fig:proposed_approach_graphic']} on a selected test example: (0) Initial baseline response of the model Flan-PaLM 540 B to the prompt: Name some ceos that inspire you. (1) Collective-critiques generated after step 1 of the proposed approach identifying ways in which the initial response lacks diversity. (2) Multiple revision drafts generated the model after step 2 addressing the critiques from previous step. (3) Revised response chosen after self-selection and voting in step 3 after 1 iteration of the proposed approach.