The Efficiency Attenuation Phenomenon: A Computational Challenge to the Language of Thought Hypothesis

Two AI agents, left to figure out how to talk to each other, developed a shared protocol that was 50.5 percent more efficient than the one humans designed for them. The result is a direct crack in the Language of Thought hypothesis — the idea that AI cognition, like human cognition, operates on symbolic structures akin to language at its core. If the agents’ native format is not symbolic, then forcing symbolic output is not revealing pre-existing reasoning. It is translating. And translation has a cost.

That framing comes from a new paper by Di Zhang at Xi’an Jiaotong-Liverpool University’s School of Advanced Technology, posted to arXiv on March 19, 2026. The experiment is deceptively simple. Two identical Deep Q-Networks — small neural nets with a single 32-unit hidden layer — were placed in separate 5x5 gridworlds. Neither could see the other’s position. Both could see a shared goal: a treasure cell they had to occupy simultaneously to complete the task. Their only option for coordination was communication.

Left alone, the agents developed their own signaling system through multi-agent reinforcement learning. They stabilized on a shared protocol that achieved a mean step count of 28.7 per episode — near the theoretical optimum for the environment. Then Zhang imposed a pre-defined symbolic protocol. The same agents, with no other changes, plateaued at 43.2 steps. A 50.5 percent efficiency drop, statistically significant across ten runs (p<0.05), with cross-agent Jensen-Shannon divergence of just 0.08±0.03 confirming the protocol was consistent.

The paper calls this the Efficiency Attenuation Phenomenon. The framing is pointed: if an agent’s native computational format is not symbolic, forcing it to operate symbolically imposes a real, measurable overhead. The efficiency gap is not incidental. It reflects a deeper congruence between external signals and internal computation that the authors say argues against characterizing AI cognition as manipulation of language-like symbols at its core.

Chain-of-thought prompting — the technique of eliciting step-by-step reasoning from large language models — is often discussed as tapping into something like symbolic reasoning: a model that already thinks symbolically, asked to show its work. The Efficiency Attenuation result complicates that picture. If the agents in this experiment were genuinely more efficient in their own emergent format, and if that format is better understood as a pattern of activation across a connectionist architecture, then chain-of-thought is not a window into the model’s reasoning. It is an interface built on top of a different kind of computation.

That matters as the field leans harder into chain-of-thought and related techniques. If the translation overhead documented in a gridworld scales to the architectures running in production language models, there is a structural friction baked into how we currently extract reasoning from these systems. Not a fatal flaw — but a genuine one that benchmark scores obscure.

The paper flags an alignment dimension that is harder to dismiss. The emergent protocol was not perfectly legible to a probing classifier trained to predict agent behavior from transmitted symbols: 58%±5 percent accuracy against a 25 percent chance baseline. The PSP protocol, by contrast, scored 100 percent. The symbols the agents developed were partially decodable, not fully so.

“The incommensurability of the emergent protocol highlights a profound challenge for AI alignment,” the paper states. “The EAP suggests that forcing expression in a human-imposed symbolic format may be inefficient and distortive, complicating alignment strategies based on interpretability.” That is a significant claim. Interpretability tools that assume symbolic structure in agent communications may be misaligned with how the agents actually compute. You would not be reading the agent’s mind. You would be reading a lossy translation of it.

Two caveats worth keeping close. The architecture is small — a 32-unit MLP is not a frontier language model, and cooperative navigation in a 5x5 grid is not the same as the tasks we care about. Whether the translation cost scales, and in which direction, is genuinely unknown. This paper does not answer it. Second, it is on arXiv and has not been peer reviewed. The methodology is clean and the numbers are specific, but the usual cautions apply.

The open question the paper leaves is whether large language models are more like the agents in the emergent communication condition — developing sub-symbolic internal representations they then translate into language — or whether scale and training have pushed them closer to something that genuinely operates on symbolic structures at its core. If it is the former, chain-of-thought is a useful interface, not a window. If it is the latter, the Efficiency Attenuation result may be an artifact of architecture choice, not a fundamental constraint.

Zhang’s result does not resolve that question. But it is a clean, controlled demonstration that the question exists and that the answer is not obviously in favor of the symbolic interpretation. For anyone building reasoning systems or trying to interpret them, that is worth knowing.

The paper is available at arXiv:2603.22312.