Structure-Aware Diversity Pursuit as an AI Safety Strategy against Homogenization
Ian Rios-Sialer · 2026
Read Structure-Aware Diversity Pursuit as an AI Safety Strategy against Homogenization by Ian Rios-Sialer (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-06116.txt.
Introduction
Ian Rios-Sialer argues homogenization should sit at the center of AI safety in a 2026 paper. Generative models copy human biases from training data and can amplify them, as varied outputs collapse toward the most common patterns. The resulting sameness harms minority groups and impoverishes everyone else.
To characterize homogenization, Rios-Sialer offers a framework where stakeholders encode their own context and values. An experiment with a story prompt surfaces gender bias in a current model. Drawing on queer and feminist theory, the paper formalizes homogenization through normativity and introduces xeno-reproduction, a class of tasks meant to counter sameness by promoting diversity. The work opens a collaborative research line on meaningful diversity in AI.
Catalog record
| CREATOR | Ian Rios-Sialer |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 60 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2601-06116
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Structure-Aware Diversity Pursuit as an AI Safety Strategy against Homogenization Ian Rios-Sialer 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. 1 Introduction But even if we are not here next year, our DMs, our selfies, our late-night voice notes, they’ll be. Our memory is the archive now. @bundleof_styx July 28, 2025 on Reels In this epigraph, trans intellectual bundleof_styx laments the recent transphobic turn in the United States, a shift that threatens the survival of her community. The stories in the margins have historically been excluded from the archive [ 122 ] , so their memory faded with them. Today, however, the internet allows (and forces) the recording of many more stories. These are still very subtle traces against the dominant narratives [ 59 ] . How should technology respond to the faint echoes of the minoritized ? AI safety recognizes that AI systems can amplify biases leading to concrete harm [ 7 ] . However, AI safety usually differentiates and prioritizes future catastrophic risk over present social harm [ 94 , 51 ] . In this paper, we respond to traces from the margins by foregrounding them within AI safety discourse through a focus on diversity . The harms from biases in Machine Learning (ML) systems are many, including representational [ 69 ] , allocational [ 117 ] , and narrative 1 1 1 Narrative harms can also be considered as aspirational [ 36 ] , imaginative [ 40 ] , and epistemic [ 5 ] harms, or hermeneutic [ 41 ] injustices. [ 23 ] harms. Concerning Generative Artificial Intelligence (GenAI), we particularly emphasize that biases result in homogenization , a harmful loss of diversity in generated outputs [ 108 , 1 , 59 , 121 , 92 ] . Borrowing terminology from critical theory [ 53 ] , we call a strategy xeno-reproductive if it counteracts homogenization 2 2 2 While homogenization reproduces “the same” and narrows futurability [ 8 ] , xeno -reproduction reproduces “the strange” and widens possibilities. . Xenoreproduction is the generative process that intentionally increases diversity. Our main standpoint is that diversity is always relative to a context. We take the first steps to operationalize this principle by offering an abstract framework that aims to encapsulate some nuances of context. Our framework can be thought of as structure-aware , as it offers a vocabulary of structures , systems , and compliances . Given that an LLM defines a probability distribution over all possible trajectories, we enhance our structural account with string statistics . This allows us to further introduce the notions of cores , orientations , deviances , and dynamics . Finally, our formalism enables us to formalize xeno-reproduction. Our contributions: • We motivate the formalization of xeno-reproduction as a core AI safety strategy. (Section 2 ) • We provide an expressive theoretical framework that allows us to jointly reason about the structures and the statistics of strings. (Section 3 ) • We formalize homogenization (Section 4 ) and xeno-reproduction (Section 5 ). • We provide initial theoretical results and touchpoints from our framework. (Section 6 ) Our position is that AI safety should center homogenization in its research and mitigation agenda, and that structure-aware diversity pursuit is a key part of the strategy to address homogenization in LLMs . The goal of this paper is not to present a complete and empirically validated algorithm, but rather to offer a conceptual vocabulary and formal scaffolding to guide future research on diversity in LLMs. 2 Background A case against homogenization is a case for diversity. Roughly, we can think of the diversity of a community as the average rarity of its members [ 79 ] . For a community of LLM outputs, a string is rare if it is generated infrequently, and similar strings are also generated infrequently. However, people tend to disagree on what kind of similarities and differences are meaningful [ 133 ] . Embracing ambiguity [ 106 ] for us amounts to attending to context . This section situates diversity in the contexts meaningful to us, guiding our desiderata for xeno-reproduction. 2.1 Why is diversity lost? The initial driver of diversity loss is the way our data is collected [ 48 ] . The archive does not fully or accurately represent reality. Minoritized populations are often underrepresented or misrepresented in the existing corpora of data [ 7 ] . Even if our training data perfectly reflected the world, generative models [ 56 ] generally do not capture the complete diversity of the training data. This phenomenon has been referred to as mode collapse [ 62 ] , a failure of distributional faithfulness that negatively impacts diversity. It was initially introduced in the context of GANs [ 56 ] . For LLMs, the terminology has been somewhat loose [ 112 ] . Generalized mode collapse encompasses mode dropping [ 57 , 140 ] , no-breadth scenarios [ 67 ] , coverage collapse [ 112 ] , overgeneralization [ 82 ] , mode interpolation [ 3 ] , degeneration [ 38 ] , and catastrophic forgetting [ 22 , 129 ] . 2.2 Why is diversity important? There are always rare events of interest 3 3 3 For both positive and negative reasons. in the long tails of reality’s distribution . For example, we want to understand, model, and prepare for extreme catastrophes [ 45 ] , such as unexpected natural disasters. Similarly, we want to reproduce those rare bursts of genius that generate novel, paradigm-shifting innovations in our research work [ 130 , 54 , 137 ] . We find examples of interesting rarity in all domains [ 123 ] , including: web server computing [ 28 ] , market research [ 132 ] , autonomous vehicles [ 103 ] , cybersecurity [ 30 ] , and ecology [ 80 ] . How do we guide our GenAI models to reproduce the realities found in these long tails? Outliers [ 14 ] and anomalies [ 110 ] are powerful [ 6 , 25 ] . Each instance represents a possible real mechanism that we have not yet considered [ 136 , 109 ] . Because we lack understanding, they often escape our systems of classification [ 16 ] . Even experts can confuse [ 119 ] aleatoric and epistemic uncertainty 4 4 4 For instance, [ 52 ] recently showed how indeterminism in LLM inference (which can turn on-policy RL into off-policy RL [ 138 ] ) can in fact be explained and reduced, so it is not truly stochastic. . Some of the long tails of reality originate from structural inequity in society [ 113 , 84 ] . Without any intervention, GenAI is expected to worsen the lives of those minoritized [ 59 ] . The traces