Language Game framework lets LLMs negotiate with gene networks in virtual worlds

Michael Levin's lab has published a framework called Language Game that enables large language models to hold two-way conversations with non-neural biological systems. The approach, detailed in an arXiv preprint released this week, demonstrates real-time dialogue with gene regulatory networks — collections of genes that switch each other on and off without any neurons involved.

The system works by placing a frozen biological network inside a simple virtual game environment, then training lightweight adapters that translate between three domains: human text, the physics of the game world, and the network's internal state. An LLM reads a user's request, the adapter converts it into environmental conditions (nutrient levels, temperature, spatial layout), the gene network responds to those conditions, and the adapter translates the network's output back into text the LLM can parse. The biological system itself never changes — only the translation layers are trained.

What stands out

1. 01Top-down control without editing genes. Instead of rewriting DNA or injecting chemicals, the framework treats cells as agents you can negotiate with. You describe a goal in plain English; the system figures out which environmental cues will persuade the network to behave that way.
2. 02Frozen networks, trained interfaces. The gene regulatory network remains untouched during training. Only the adapter modules learn, which means the approach scales to any biological system you can simulate — no need to retrain the biology itself.
3. 03Game environments as shared substrate. The virtual worlds are intentionally simple: 2D grids with resources, obstacles, and basic physics. They're abstract enough that the same adapter architecture works across different network types, but concrete enough that the LLM can reason about cause and effect.
4. 04