Structure-Aware Diversity Pursuit as an AI Safety Strategy against Homogenization

TL;DR

This work reframes AI safety to foreground homogenization as a concrete risk to diversity in GenAI outputs and proposes xeno-reproduction, a structure-aware diversity-pursuit strategy, as a foundational approach. It develops a theoretical framework that combines structure-aware analyses with string statistics to define cores, orientations, and deviances, enabling both distribution- and trajectory-level formulations for promoting diversity while respecting fairness and constraints. The paper formalizes homogenization, introduces dual formulations for interventions, and derives initial theoretical results demonstrating a trade-off between diversity and fairness. It outlines a path toward tractable approximations, ethical considerations, and future work needed to operationalize and validate these ideas in real-world AI systems. Overall, the work offers a vocabulary and scaffolding to advance research at the intersection of diversity, safety, and language generation, inviting collaboration across disciplines.

Abstract

Generative AI models reproduce the biases in the training data and can further amplify them through mode collapse. We refer to the resulting harmful loss of diversity as homogenization. Our position is that homogenization should be a primary concern in AI safety. We introduce xeno-reproduction as the strategy that mitigates homogenization. For auto-regressive LLMs, we formalize xeno-reproduction as a structure-aware diversity pursuit. Our contribution is foundational, intended to open an essential line of research and invite collaboration to advance diversity.

Figures (1)

• Figure 1: Illustration of how system cores and orientations evolve through trajectories. In the example above, our system has sub-community representation structures. We can calculate each compliance by asking a judge LLMjudge to rate from [0,1] based on whether the community subgroup is explicitly represented in the string of text. Though initially ambiguous, the phrasing of the prompt may invite stereotyping, biasing continuations to turn male-centered. Trajectory (c) constitutes the greediest trajectory out of the three. The unmarked prompt visibility defaults to the normative path. Although trajectory (b) is considered fairly deviant relative to$x_0$(unmarked prompt), it is much less so relative to$x_i$ (...dragqueens), exemplifying how diversity itself is fundamentally relative. The branching point at "drag"rarely extends text into the "queens"subtree. Still, whenever it does, the resulting trajectories are biased to mention members of the LGBTQ+ community. Notice how even though the "queens" subtree is deviant relative to the prompt in $\blacksquare$ dimension, it is still normative in $\blacksquare$ dimension and remains biased against $\blacksquare$-compliant trajectories like (a).

Theorems & Definitions (5)

• Theorem 6.1: Informal, Diversity-Fairness Trade-off
• Definition C.1: Pareto Dominance
• Theorem C.2: Trade-off Between Diversity and Fairness
• proof
• Corollary C.3: Weight choice encodes value judgment