Team research project · June 2026
Evolutionary optimisation of communication graphs in GPTSwarm
Reproducing adversarial multi-agent LLM experiments and comparing REINFORCE with an evolutionary edge optimiser.
Conceptual illustration; not an experimental result.
Individual contribution
My role
I contributed to the implementation and analysis, with a particular focus on aggregating experimental results and producing visualisations. I also worked on robustness fixes, statistical interpretation and the report. The contribution statement describes a balanced overall workload across the three team members.
Contribution statement in the reportResearch question
Can a discrete evolutionary algorithm optimise communication between language-model agents as effectively as the REINFORCE edge optimiser used in GPTSwarm? The study examines both task accuracy and computational cost.
Methodology
The team reproduced adversarial MMLU experiments with Llama 3.1 8B via Ollama. Five conditions compare a single agent, fully connected and random graphs, REINFORCE, and an evolutionary algorithm over binary adjacency matrices. Experiments vary the number of truthful and adversarial agents.
Findings
Both optimisers improved on random connections in the reported setting. With three truthful and three adversarial agents, REINFORCE reached 50% accuracy and the evolutionary method 51%. The evolutionary method used substantially more tokens: approximately 3.7 times as many in that configuration, increasing to around 10 times in the largest configuration.
Limitations
The small accuracy difference does not establish a general advantage for the evolutionary method. The report documents a constrained reproduction study with one model family and benchmark setting; token cost is a central trade-off.
Related research
SentenceBot