Language Game: Talking to Non-Human Systems

2026-05-19 · Source: cs.LG updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Life Sciences & Biology, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

Researchers Yanbo Zhang and Michael Levin introduce the "Language Game" framework, enabling natural language communication with diverse non-human dynamical systems, including gene regulatory networks (GRNs) and chaotic attractors. This framework treats communication as a game where a system's internal dynamics are frozen as the nonlinear core of a reinforcement learning (RL) policy, with only linear input and output interfaces trained. A large language model (LLM) acts as a router, selecting the most semantically appropriate game for a human prompt and designing an environmental state to elicit a desired action from the system. The system's response is then translated back into human language. Applied across 14 GRN models and the Lorenz attractor in 16 RL tasks, the framework demonstrates fluent dialogue and reveals that specific GRN properties, like transcriptional regulation and circadian rhythmicity, serve as inductive biases, making systems easier to communicate with, while ultrasensitivity and conservation laws hinder it. Code is available on GitHub.

Key takeaway

For AI Scientists and Research Scientists exploring novel human-system interaction, this framework offers a new paradigm for conversing with dynamical systems. You can leverage reinforcement learning and LLMs to establish bidirectional communication with complex biological or physical systems, treating their goal-directed behavior within a "game" as their language. Consider how specific system properties might act as inductive biases, influencing a system's "communicability" and task performance, guiding your choice of underlying dynamics for optimal interaction.

Key insights

Communication with non-human dynamical systems is possible by framing interaction as a language game mediated by an LLM and reinforcement learning.

Principles

• Meaning arises from use within a goal-directed game.
• Well-trained agents converge to similar behaviors.
• Dynamical system properties act as inductive biases for control.

Method

A fixed nonlinear system acts as an RL policy's core, with trainable linear interfaces. An LLM routes human prompts to suitable games, designs environmental states, and translates system actions back to human language.

In practice

• Apply to gene regulatory networks for biomedical control.
• Use for bioengineering and hybrid intelligence interfaces.
• Probe artificial and virtual agents beyond programmed interfaces.

Topics

• Language Game Framework
• Non-Human System Communication
• Gene Regulatory Networks
• Reinforcement Learning Policies
• Large Language Model Integration

Code references

• Zhangyanbo/language-game

Best for: AI Scientist, Research Scientist

Related on AIssential

• Simulation of Language Evolution under Regulated Social Media Platforms: A Synergistic Approach of Large Language Models and Genetic Algorithms — The framework simulates adversarial language evolution on social media using LLM-driven multi-agent systems and Genetic Algorithms.
• Beyond one-on-one: Authoring, simulating, and testing dynamic human-AI group conversations — DialogLab bridges scripted and generative AI dialogue for complex multi-party human-AI conversations.
• DiscoTrace: Representing and Comparing Answering Strategies of Humans and LLMs in Information-Seeking Question Answering — LLMs lack rhetorical diversity and systematically over-address question interpretations compared to human answerers.
• MT-PingEval: Evaluating Multi-Turn Collaboration with Private Information Games — Language models struggle with multi-turn collaboration, often failing to improve over non-interactive baselines despite potential.
• Using Cognitive Models to Improve Language Model Simulation of Human Persuasion Games — Cognitive models and RL can guide LLMs to simulate diverse human decision-making, improving training environments.
• Learning to Communicate: Toward End-to-End Optimization of Multi-Agent Language Systems — DiffMAS optimizes multi-agent systems by making latent communication a jointly learnable component.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.LG updates on arXiv.org.