The Bicameral Model is a research design from May 2026 in which two AI language models work as a pair: one writes the answer, the other operates a tool such as a calculator. They exchange signals through a small trained link inside the models instead of writing text. In a lab test, a small model that got about a third of its sums right scored above 96% this way.
Three researchers at AWS, Cedric Flamant, Udaya Ghai and Kanna Shimizu, published the design on arXiv on 11 May 2026 (The Bicameral Model). Bicameral means "with two chambers". The blog post Can two language models work together without exchanging text? goes through the paper in more detail.
How does it work?
A language model turns every token, a word or part of a word, into a series of numbers that it refines layer by layer. These are its hidden states. The Bicameral Model runs two models side by side, each generating one token per step, in lockstep:
- The primary model drives the task and writes the answer the user sees.
- The auxiliary model operates a tool: a calculator, the Z3 constraint solver or a Python sandbox.
- A neural interface reads the hidden state of one model at a chosen layer, translates it into the other model's representation space and mixes it in, first from primary to auxiliary, then back. A learned suppression gate decides how much of that signal the receiving model admits.
Only the interface is trained; the weights of both models stay frozen. It is small: 16 million parameters in the arithmetic setup and 21 million for logic puzzles, next to two models of 0.5 or 0.6 billion parameters. Training is supervised fine-tuning on the output of both models.
The auxiliary still calls its tool in text. Asked what 4 times 3 is, it writes calc(4*3), and the result comes back into its own stream. The primary model never sees that text: it receives the result only through the coupling, and writes the answer, 12.
What did the tests show?
- Arithmetic: two copies of Qwen2.5-0.5B-Instruct scored 96.5% with a calculator, against 36.2% for one model alone. That is the best of 890 configurations of which layers to couple; 95.4% of them beat the single model. The same interface with the auxiliary bypassed scored 48.0%, so the gain comes from the auxiliary and its tool.
- Logic puzzles: two copies of Qwen3-0.6B with Z3 scored 64.7% on ZebraLogic, 1.7 times the 37.5% of one model alone in thinking mode.
- Math with Python: two copies of Qwen3-4B scored 62.5% on the MATH benchmark, below the 81.6% of one model alone in thinking mode. The pair did solve 127 problems that thinking mode missed.
Where does it fall short?
- Running two models in lockstep can double the compute.
- The coupling hurts what the model already does well: on the GSM8K word problems, accuracy fell from 49.6% to around 40% in all 890 configurations.
- Training needs task-specific data, annotated with when the auxiliary may act.
- Every reported result uses two copies of the same model, of 0.5 to 4 billion parameters. A preliminary pairing of a 32B primary with a 0.6B auxiliary worked, but it was not optimized, and the drop on word problems remained.
How does it differ from other multi-agent setups?
In a multi-agent system, or when a model calls a tool through function calling, models communicate by generating text, one message after another. Earlier methods that share internal states instead, such as C2C (which fuses one model's KV cache into another's), LatentMAS and ThoughtComm, pass them once or between turns. According to the authors, the Bicameral Model differs from all of them because its coupling runs in both directions, at every generated token.