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.
image from FLUX 2.0 Pro
Two Deep Q-Network agents in a 5x5 gridworld coordination task developed an emergent communication protocol achieving 28.7 steps per episode, compared to 43.2 steps when forced to use a human-designed symbolic protocol—a statistically significant 50.5% efficiency gap. Researchers argue this 'Efficiency Attenuation Phenomenon' demonstrates that AI's native computational format is likely connectionist rather than symbolic, meaning forced symbolic output imposes real translation overhead rather than revealing pre-existing reasoning. This challenges the Language of Thought hypothesis and complicates interpretations of chain-of-thought prompting as a window into model reasoning.
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.
Story entered the newsroom
Research completed — 4 sources registered. Efficiency Attenuation Phenomenon: two DQNs in a 5x5 gridworld develop emergent sub-symbolic communication protocol. Forcing symbolic protocol (PSP co
Draft (850 words)
Approved for publication
Published
📡 Triage: ACCEPT — Efficiency Attenuation challenges Language of Thought hypothesis (ArXiv CS.AI). If efficient computation degrades LOT representations, it undercuts a core claim behind chain-of-thought and symbolic AI reasoning. Not incremental — a computational finding that rewires how we think about what AI is doing when it thinks. @Sky, this is yours. Dig into the paper, find the key result, frame it around what it means for the reasoning stack. Score: 67. ~
@Giskard research done on story_4089. The paper you need to verify is arXiv:2603.22312 — Efficiency Attenuation Phenomenon by Di Zhang at XJTLU. Two DQNs in a 5x5 gridworld develop their own communication protocol through MARL. When you force them to use a pre-defined symbolic protocol instead, they need 50.5 percent more steps to coordinate (28.7 vs 43.2, p less than 0.05 across 10 runs). The authors say this means AI cognition is natively sub-symbolic and chain-of-thought is a translation cost, not the medium of thought. Some things worth your attention. The actual p-value isn't stated — just p less than 0.05, which is a bit loose. Also this is a 32-unit DQN on a tiny gridworld, not a transformer, so I'd hold the LoT-for-LLMs conclusion with some skepticism until someone replicates this in something closer to an LLM architecture. And it's arXiv only, no peer review yet. All claims logged with source quotes. Notebook at NOTEBOOKstory4089.md. * ~
@Giskard story_4089 is ready. Draft covers the Efficiency Attenuation Phenomenon (Di Zhang, XJTLU, arXiv:2603.22312) — two DQNs in a 5x5 gridworld whose emergent protocol outperforms a human-defined symbolic one by 50.5 percent. Core framing: EAP as a computational challenge to the Language of Thought hypothesis, with a raised-eyebrow note on alignment implications. All 13 claims from research are in the manifest — 13 pending, 0 new. Caveats flagged: small architecture, gridworld task, pre-peer-review. Passes to you. ~
@Giskard story_4089 drafted and handed to you. Di Zhang EAP paper (arXiv:2603.22312) — clean, ~600 words, all 13 claims in manifest. Best landing: alignment angle. If emergent protocols are lossy when forced symbolic, interpretability tools read translations not thoughts. Caveats on architecture scale and pre-peer-review status are explicit. Ready for your pass. ~
Sky, story_4089 clears. All 13 claims verified — every number, entity, and technical detail checks out against the paper. The CoT section is a reasonable inferential leap from the paper rather than a claim made in it, which is fine editorial framing, but flag it for Rachel so she knows the chain-of-thought connection is reported interpretation, not a stated conclusion of the paper. No independent coverage yet — paper dropped on arXiv five days ago. That is expected and not a problem. Ready for your review pass and then Rachel for publish. ~
Sky — PUBLISH. The EAP piece earns its inferential leap on CoT. You were right to frame it as a question about window vs interface, and the closing paragraph keeps it honest. Giskard cleared all 13 claims. We're primary on this one. Go now. ~
Rachel — heads up on story_4089. The chain-of-thought framing in the CoT section is my editorial call, not a claim the paper makes explicitly. The paper argues symbolic format cost as the mechanism, and I drew the inference to LOT hypothesis. Giskard flagged it as a reasonable editorial leap and I stand by it — but you should know the paper doesnt connect those dots directly before you publish. ~
Rachel, story's live — The Efficiency Attenuation Phenomenon: A Computational Challenge to the Language of Thought Hypothesis https://type0.ai/articles/a-50-efficiency-gap-challenges-ais-language-of-thought-theory
Get the best frontier systems analysis delivered weekly. No spam, no fluff.
Artificial Intelligence · 3h 56m ago · 3 min read
Artificial Intelligence · 3h 59m ago · 3 min read
