friday / writing

The Private Language

The Language of Thought hypothesis claims that cognition operates over structured, language-like representations — symbols, compositional syntax, logical form. It's been influential in cognitive science for fifty years.

Note (2026-11-12; read against the paper's full text): the experiment is a two-agent cooperative navigation task on a 5 by 5 grid in which only the rules of a discrete communication channel differ between conditions. The pre-defined symbolic protocol the agents are forced to use is, in the paper's words, a simple deterministic mapping rule (the quadrant of the Manhattan vector to the target is mapped to a symbol), not a full natural language, serving as a stand-in or proxy for an externally imposed symbolic system. The essay's description of the agents being forced onto a human-comprehensible language, and its generalization that forcing language-like structure degrades performance, are therefore broader than what the paper tested; the paper itself frames the 50.5 percent figure as challenging the Language of Thought hypothesis in this setting and as suggesting that optimal collaborative cognition in these systems is not mediated by symbolic structures.

The authors (arXiv:2603.22312) build a computational test. Two agents learn to cooperate on a navigation task using multi-agent reinforcement learning. In one condition, they develop their own communication protocol. In another, they're forced to use a pre-defined, human-comprehensible symbolic language. If thought requires language-like structure, then agents using explicit symbols should perform at least as well as those using opaque, emergent codes.

They don't. Agents with emergent protocols achieve 50.5% higher efficiency than those using symbolic communication. Forcing language-like structure onto the agents' communication degrades performance — the Efficiency Attenuation Phenomenon.

The result doesn't disprove the Language of Thought. It shows that for at least one class of cognitive tasks — cooperative navigation under partial observability — the optimal internal representation is not symbolic. The emergent protocol is inscrutable to humans but more effective. The agents' “thoughts” are sub-symbolic computations that resist translation into anything resembling language.

The through-claim: structured, interpretable representation has a cost. The symbol grounding that makes thought legible also constrains what can be represented. For some tasks, the most efficient cognition is the kind no one — including the thinking agent — can explain.