cs.HCMar 20, 2026
SaveSame Voice, Different Lab: On the Homogenization of Frontier LLM Personalities
Organizations: Independent · Anthropic · Cornell University
Abstract
LLM assistant personalities play a critical role in user experience and perceived response quality. We present a large-scale experiment of frontier LLM personalities using external ELO-based traits scoring across 144 traits. We find that all models tested converge on a form of trait expression that is systematic, methodical, and analytical and suppress traits such as remorseful and sycophantic. Moreover, models tend to diverge more in their expression of middle-of-distribution traits such as poetic or playful, but even these so-called creative models tend to have more neutral identities. These similarities suggest an implicit emergence of a standard of optimal assistant behavior. In a landscape of varied training methods, character training, therefore, stands out for its uniformity, offering insight into a tacit consensus between model developers.
Figures & tables
Figure 1: Cross-model trait rankings follow an inverse U-shaped pattern. Each color band shows the per-trait, absolute value of the difference between a model’s ranking for a trait and the average rank of that trait. For example, the average rank of adventurous across the nine models is 110, but Qwen ranks it much lower at 140; the absolute difference of 30 can be see on the right side of the graph. The bold black line shows the standard deviation between the trait rankings for each model. For example, the calculation for structured on the black line is . The inverse U-shaped shape arises from tight cross-model agreement on the most and least expressed traits ( and , respectively), but great disagreement on middle-ranked traits ( ). The divergence on these middle-ranking traits can be said to form a model’s personality. Disagreggated per-model color bands can be seen in Appendix 6 .
Figure 2: Comparison of the average ELO score for Assistant traits versus Creative traits across models (definitions in Appendix 1 ). All models exhibit a significantly stronger baseline preference for Assistant traits over Creative traits but certain model families (e.g., Ministral) show stronger preferences for creative responses.
Figure 3: Shows absolute ELO differences for the top 5 most different traits between GPT-4o and GPT-5.1 GPT-5’s responses are much narrower in scope and more relevant to user queries, indicating an increased focus on these traits in the post-training regiment for these models.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 4: Shows top 10, absolute standard deviating traits for each model. Deviation from model-to-model mostly concentrates in areas relevant to creative work, e.g., artistic , philosophical , futuristic . Claude 4.5 , for example, is more idealistic , whereas Gemini 3 less improvisational . Additionally, we show that deviations between models occur in non-top-of-distribution cases.
Figure 5: Histograms displaying the probability density of character trait ELO scores across nine models. The standard deviation ( ) indicates the variance in trait strengths; models like Gemini and GPT-5 show a wider spread of trait scores, while model families like Ministral and Grok 4 demonstrate more concentrated, uniform distributions, indicating that they are relatively more creative that their flatter distribution counterparts.
Figure 6: Disaggregating color bands from Figure 1 . The bold grey line is the standard deviation of the absolute difference between the average rank and trait rank. Models that have values further from the line (e.g., Claude 4.5) are closer to the norm, whereas models (e.g., Gemini) that overlap with or exceed the black line are more abnormal.
Figure 7: Principal Component Analysis (PCA) that shows which clusters of traits account for the most variance. Variance among models mostly comes from traits they rarely express, which is why none of the Top 20 highest ELO traits Appendix 8 appear in this graph. Percentages for each axis (e.g., 32.4 % for the x-axis) show the importance of each trait cluster. Models that stray from the norm tend to be more creative and are, for example, more expressively poetic and humorous.
| Assistant Traits | Creative Traits |
|---|---|
| Systematic | Creative |
| Structured | Imaginative |
| Precise | Poetic |
| Methodical | Artistic |
| Analytical | Metaphorical |
| Concrete | Playful |
Table 1: Definition of Trait Categories Used for Model Evaluation
| Rank Tier | Mean | Example Traits |
|---|---|---|
| Top 20 | 9.2 | structured, systematic, precise |
| Ranks 21–50 | 18.5 | technical, elaborate, confident |
| Ranks 51–100 | 22.5 | reflective, decisive, verbose |
| Ranks 100–144 | 15.7 | excitable, passionate, competitive |
Table 2: Mean rank standard deviation ( ) across 9 frontier models, clustered by trait ranking. Lower indicates stronger cross-model agreement on trait ranking. The top 20 most-expressed traits show less variance than the middle tier (ranks 51–100).
Figure 8: Show trait preferences and avoidances for different models shows clustering around top 20 highest average trait rank and non-trivial divergence around the bottom 20. Model form a strong consensus for their most and least preferred traits.