Bouba Beats Kiki, and Other Regularities in the Subjective Preferences of LLMs

AI summary

Overview

Research area: Natural Language Processing / LLM evaluation and model behavior Technical level: Intermediate Scope: An early-stage framing of how LLMs choose between multiple viable answers on subjective tasks, studied through short-form comparison tasks over numbers, baby names, and fictional cities.

What This Paper Is About

Evaluation of LLMs is well developed for objective questions with a right answer, but there is no comparable methodology for subjective tasks where several responses are all acceptable. This paper asks not whether a model is correct, but how it picks among viable options when there is no clear winner. The authors narrow the problem to short-form content — numbers, baby names, and fictional cities — and probe the preferences models express when comparing such items.

Key Contributions

1. A framing for subjective evaluation. The paper reframes evaluation away from accuracy and toward the structure of model choice among multiple valid responses, arguing this is an underdeveloped area of methodology.
2. A catalogue of preference-relevant aspects. The authors identify 29 aspects of short-form items spanning syntax and semantics — including length, alphabetic order, popularity, vowel mix, and spoken sound — and use these as the lens for studying model preferences.
3. Evidence that preferences are strong and reproducible. Models show consistent leanings for or against particular aspects across comparisons.
4. Evidence that question phrasing changes outcomes. The paper shows that how a model is asked to express a preference can switch individual aspects on or off, can shift overall preference distributions between pairs, and can create interactions between generating a list of options and ranking it.

Main Findings

• Preferences are strong and reproducible: Models tend to lean consistently for or against certain aspects of short-form items rather than choosing arbitrarily.
• Aspects can be activated or deactivated by phrasing: Small changes in how the preference request is worded can determine which aspects influence the model's selection.
• Preference distributions shift under trivial-seeming variation: The overall pattern of preferences between pairs of options can change radically based on what appear to be unimportant differences in the question, which the authors flag as a manipulation concern.
• Generation and ranking interact: When a model both produces the candidate list and ranks it, the two stages interact in unexpected ways that can drive the stated preferences. Details of these interactions are not given in the abstract.

Methodology in Plain English

The researchers deliberately restricted the problem to short content so that items could be compared cleanly and their properties catalogued. They assembled three families of items — numbers, baby names, and made-up city names — and described each with 29 features covering how items look, sound, and are conventionally used. They then had models choose between pairs of items and observed which features tracked with the choices. To test fragility, they varied the wording of the request and compared preferences across those variations. They also tested a two-stage setting where the model first generates the alternatives and then ranks them, to see whether the source of the list changes the ranking. The abstract reports the direction of these effects but does not give the size of the datasets, the specific models tested, or quantitative results.

Why This Matters

Subjective judgment is a large share of how people actually use chatbots — naming, wording, style, taste — and accuracy-style benchmarks say nothing about it. Treating model preferences as a measurable object of study gives researchers a vocabulary for a question that has mostly been handled anecdotally. It also suggests that model outputs on subjective tasks are sensitive to surface features of the prompt, with consequences for both reliability and manipulation.

• Prompt design and product UX: Teams writing system prompts and choice interfaces learn that swapping phrasing can reshape which options a model favors.
• Benchmark and evaluation design: Provides a template for testing subjective behavior rather than factual recall.
• Guardrails and manipulation review: If preference distributions swing on trivial phrasing, that is a lever a bad actor could pull, and one worth auditing.
• Pipeline design for recommendation-style features: The generate-then-rank interaction warns against assuming the two stages are independent.

Industry relevance: Any product where a model picks among alternatives on the user's behalf — naming suggestions, copy variants, option ranking — is affected, as are red-teaming and QA processes for those products.

Future Directions

• Extend the approach beyond short-form items to longer, more open-ended subjective content where features are harder to enumerate.
• Build a systematic methodology for subjective evaluation, comparable in rigor to accuracy benchmarks.
• Characterize which phrasing variations move preferences and why, to distinguish benign sensitivity from exploitable manipulation.
• Investigate the generation-ranking interaction to determine when a model's own candidate list biases its stated preferences.

Target Audience

Researchers and practitioners working on LLM evaluation, alignment, and model behavior, plus product and prompt-engineering teams who rely on models to choose among options. Readers looking for concrete benchmark numbers will not find them in the abstract; the value here is in the framing and the qualitative findings.

Authors’ abstract

Users engage with large language models on tasks that are purely objective ("When was Netflix founded?"), purely subjective ("What should I name my child?") or anywhere in-between. We have rich tools for evaluating model accuracy with respect to objective criteria. However, when it comes to subjective aspects of a task, we do not have similarly developed methodology at hand. In this paper, we take some early steps towards framing the question of subjective evaluation. Rather than measure how often a model is right or wrong on a question, we instead ask, how do models make choices among multiple viable responses without a clear winner? To study this question, we restrict our focus to short-form content, and consider comparison tasks among three such content types: numbers, baby names, and fictional cities. We identify 29 aspects of these short-form examples covering many areas of syntax and semantics (length, alphabetic order, popularity, vowel mix, spoken sound, etc.). We show first that models tend to have strong and reproducible preferences for or against some of these aspects. Second, we show that variations in how the model is asked to express a preference may cause some of these aspects to activate or deactivate in the model's selection. Third, we show that overall preference distributions among pairs may vary radically based on seemingly-unimportant variations in the nature of the question posed to the model, raising new questions about manipulation. And finally, we show that, if models are used both to create lists of alternatives and to rank them, there are unexpected interactions between the two stages that may also drive the models' stated preferences.