Are LLMs becoming similarly creative? Evidence from three years of models

Summary

Provides the base-rate evidence for the monoculture premise on the exact output dimension the paper cares about. Across three years of model releases, measured on Infinity-Chat100 (real-world open-ended user queries) and the Alternate Uses Task (an established psychometric creativity assessment), sentence-embedding similarity shows a statistically significant decrease in output diversity over time, i.e. models are converging in creative substance. If cross-model diversity is shrinking, then any exploration or novelty gain the paper attributes to architecture has to be shown to survive on a common backbone, which strengthens the case for structure as the lever. Explicitly preliminary, so cite for the trend rather than the magnitude.

Contribution

Queued — not yet read.

Method

Queued — not yet read.

Relevance to RISE

Surfaced by the 2026-09-07 literature sweep as a medium-priority candidate for the Architecture as Epistemology (ISR) paper (angle: correlated errors / homogeneity across LLM agents). Verified against: Opened https://arxiv.org/abs/2608.19437 and read the full abstract, four-author list, submission date 19 Aug 2026, primary category cs.CL, 12 pages. Before the Aug 25 window boundary; included as directly on-angle.

Critique / open questions

Queued — not yet assessed.

Key quotes

Many benchmarks track Large Language Model (LLM) performance on tasks with verifiable answers, but less is known about how LLM performance is evolving on open-ended tasks, where creativity, originality and diversity may matter as much as quality. (abstract)

As LLMs increasingly support human ideation and creative work, understanding trends in LLM performance on open-ended tasks is critical. (abstract)