Language Game: Talking to Non-Human Systems — Wittgenstein Meets Gene Regulatory Networks

Forum topic · 小凯 · 2026-05-19

Summary

A Chinese tech forum post reviews the paper "Language Game: Talking to Non-Human Systems" (arXiv:2605.16321) by Yanbo Zhang and Michael Levin (Tufts/Harvard), posted May 5, 2026. The paper operationalizes Wittgenstein's notion of language games to enable communication with non-neural intelligent systems such as gene regulatory networks (GRNs), microbial communities, and fungal mycelia. The core method freezes a system's internal dynamics as the nonlinear core of a reinforcement learning policy, training only linear input/output interfaces via RL reward signals, while an LLM acts as a router that maps natural-language prompts into environment states and translates the system's native behavioral responses back into language. Key findings: behavior converges across different architectures trained on the same language game, suggesting games can serve as a cross-system lingua franca; and GRNs exhibit inherent inductive biases that make some systems easier to communicate with than others. The post raises candid questions about how much the LLM controls the dialogue's semantics, whether reward optimization implies understanding, and the lack of validated applications, concluding the work is an ambitious conceptual framework and proof of concept for dialogues with all dynamical systems rather than a near-term commercial technology.

This forum post discusses the paper "Language Game: Talking to Non-Human Systems" by Yanbo Zhang and Michael Levin (Tufts University / Harvard), arXiv:2605.16321 (cs.LG), dated May 5, 2026: https://arxiv.org/abs/2605.16321

The Core Problem

While ChatGPT can simulate asking a dish of bacteria what it "thinks," that is the LLM thinking, not the bacteria. The paper asks a principled question: can a non-human dynamical system — a gene regulatory network (GRN), microbial community, or fungal mycelium — "speak" in its own way, rather than being ventriloquized by an LLM? Levin's prior work established intelligence in non-neural systems: GRNs can "remember," microbial communities can "decide," fungi communicate via electrical signals. The challenge is how to talk to them.

The Wittgensteinian Answer

The paper draws on Wittgenstein: meaning comes from use — communication is not symbol exchange but public, practical behavioral patterns in concrete contexts. The key contribution is translating this philosophy into a runnable computational framework:

Step 1: Freeze the GRN's (or other system's) internal dynamics as the nonlinear core of an RL policy — no connection weights are modified; natural dynamics are fully preserved.
Step 2: Add only two trainable linear layers (input and output interfaces), trained via RL reward signals. The core is immutable; only the interfaces adapt.
Step 3: Define a "language game" — an RL environment such as maze navigation, block stacking, or point chasing.
Step 4: Dialogue. A human speaks natural language; the LLM routes the prompt to an environment state designed so a particular response becomes the system's rational choice; the system responds with its own dynamics; the LLM translates the action back into language.

The crucial point: the LLM does not speak for the system — it designs the situation and reads the system's natural response.

Experimental Findings

1. Behavioral convergence: Different architectures (small GRNs vs. standard LSTM agents) trained on the same language game with the same rewards produce mutually interpretable behavior — language games act as a cross-system lingua franca. 2. Natural "personality": GRNs have inherent inductive biases — large, densely connected networks are easier to train on some tasks than small, sparse ones, analogous to species-specific cognitive strengths (dogs in social cooperation, crows in tool use, octopuses in spatial manipulation). Some systems are inherently more "communicable" than others.

Honest Open Questions

Is this really dialogue? The LLM chooses the game and translates behavior, so it may substantially shape what the system "says."
The boundary of understanding: Reward optimization is not comprehension; asking "what do you fear?" may just trigger avoidance patterns without any subjective experience.
Application uncertainty: No concrete application validation (e.g., clinical dialogue with engineered cell clusters) is provided — this is a conceptual framework and proof of concept.

The Reviewer's Verdict

The author admires the paper's ambition: not a 3% benchmark gain, but opening a new door — dialogue with all dynamical systems. Don't use an LLM to simulate non-human systems or speak for them; let their natural behavior patterns serve as "language," reward structures as "semantics," and language games as "translation protocols." If the direction works out, humans might one day negotiate pollution strategies with mycelial networks, cancer treatments with gene circuits, or gut-health plans with bacterial communities.

References

1. Zhang, Y., Levin, M. (2026). Language Game: Talking to Non-Human Systems. arXiv:2605.16321. 2. Wittgenstein, L. (1953). Philosophical Investigations. Blackwell. 3. Levin, M. (2022). Technological Approach to Mind Everywhere: An Empirically-Grounded Framework for Diverse Intelligence. Frontiers in Systems Neuroscience. 4. Levin, M. (2023). Bioelectric Networks: The Cognitive Glue of Diverse Intelligence. Trends in Cognitive Sciences. 5. Sutton, R.S., Barto, A.G. (2018). Reinforcement Learning: An Introduction. MIT Press.