LLM Collectives Self-Organize to Solve Problems Following Hallmarks of Biological Collective Intelligence

Just as Large Language Models (LLMs) are now commonly used to generate solutions to problems, biological organisms since the dawn of life have been generating solutions for survival as they continuously face novel challenges in dynamic environments. Collectives of cells must coordinate to solve problems they have never encountered before, generating adaptive responses not explicitly specified in their genome. Understanding how this kind of collective intelligence emerges from local interactions among agents with heterogeneous capabilities remains a central challenge in systems biology. Meanwhile, LLMs continue to struggle with creative problem-solving beyond their training data, especially in solving complex problems, such as mathematical discoveries. These challenges are complementary. Insights from biological collectives can guide the design of more capable LLM systems, while controlled study of LLMs may reveal mechanisms difficult to isolate in living systems. This study introduces LLM-simulated expert conferences as a controllable in silico model system for studying collective problem-solving dynamics. The LLM was prompted to simulate conferences among synthetic agents, each assigned a distinct expertise profile, to solve a mathematical problem (Yu Tsumura's 554th problem) that otherwise could not be solved via direct prompting. Analysis of problem-solving dynamics revealed three hallmarks known for biological collective intelligence. First, division of labor emerged without pre-assignment, with errors detected by agents whose expertise matched the error type (p < 0.05). Second, functional repair chains arose spontaneously following a Detect, Confirm, Repair, Validate sequence analogous to sequential task handoffs in biological systems at multiple scales, such as error correction in DNA or social insect behavior. Third, discourse dynamics exhibited a phase transition from stochastic verification to ordered consensus ultimately providing the solution to the problem. Transition entropy dropped from 2.27 bits in the verification phase to 0.25 bits at consensus, representing a 9-fold collapse. This entropy collapse provided an intrinsic termination signal that characterizes consensus formation in biological collectives. Thus, the result supports the view that the mechanisms and information processing underlying collective intelligence is substrate-independent (either biological or silicon-based) and can be further studied using the new synthetic collective model mainframe. Furthermore, LLM-simulated expert conferences offer a disruptive innovation in LLMs’ problem-solving capabilities (beyond their training data) and may be applied to any complex problem in mathematics or other scientific disciplines that need creative or novel solutions.