Transformational Creativity via Multi‑Agent LLM Debate

Determine whether collaborative multi‑agent large language model systems that engage in debate and critique among agents with diverse perspectives can move beyond combinatorial or exploratory creativity to achieve transformational creativity that reshapes the underlying conceptual space of scientific ideas.

Background

The survey argues that single‑agent techniques—prompting, retrieval, and inference‑time scaling—are often insufficient for generating radically novel scientific ideas. It proposes that multi‑agent interaction, especially structured debate and critique among specialized agents, may yield emergent behaviors analogous to those observed in complex systems, potentially enabling higher‑order creativity.

Within this context, the authors explicitly conjecture that such interactions could challenge entrenched assumptions and expand the space of ideas in ways characteristic of transformational creativity, as defined by Boden. Verifying this conjecture would clarify the value of multi‑agent collaboration for scientific ideation beyond incremental recombination.

References

Extending this perspective to collaborative multi-agent LLMs, we can conjecture that debates among agents with diverse perspectives may not only expand the idea space but also challenge entrenched domain assumptions. In doing so, such systems could approach what Boden defines as transformational creativity— that alters the very conceptual space in which ideas are generated — opening pathways to more radical and paradigm-shifting hypotheses.

— Large Language Models for Scientific Idea Generation: A Creativity-Centered Survey  (2511.07448 - Shahhosseini et al., 5 Nov 2025) in Section 6: Collaborative Multi-Agent Systems