Can two language models work together without exchanging text?

Yes, on narrow tasks. On 11 May 2026, three researchers at AWS published the Bicameral Model on arXiv: two language models that run side by side and exchange signals through their hidden states, the internal representations a model computes for every token. Only a small interface between them is trained; the models' own weights stay unchanged. With a calculator, two coupled 0.5B models scored 96.5% on arithmetic, against 36.2% for one model alone. The price: compute can double, and the coupling hurt tasks the model already did well.

Why would two models skip text?

When two language models cooperate, in a multi-agent system or when a model calls a tool, they communicate by generating text. Every exchange passes through discrete tokens, one message after the other. The paper asks whether two pretrained models can coordinate through a continuous channel instead, at the same time, and what that channel learns to carry when it is trained on nothing but the task.

How does the Bicameral Model work?

The system has three parts:

• a primary model that drives the task and writes the answer the user sees
• an auxiliary model that runs alongside it and operates a tool: a calculator, the Z3 constraint solver or a Python sandbox
• a neural interface between the two, the only part that is trained

Both models generate one token per step, in lockstep. At chosen layers of the transformer, the interface reads the hidden state of one model, 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 lets in. The interface is small: 16 million parameters in the arithmetic setup and 21 million for the logic puzzles, next to models of 0.5 and 0.6 billion parameters each.

The auxiliary still calls its tool in text. Asked what 4 times 3 is, it writes calc(4*3), and the calculator's result is fed back into its own stream. The primary model never sees that result as text. It only receives it through the coupling, and writes the answer, 12, without ever seeing the calculator's syntax.

Training uses supervised fine-tuning on the output of both models, with the weights of both models frozen. The training examples need extra annotation: a constraint solver places each tool call at a point where the auxiliary can already know what it needs from the primary. A second phase of reinforcement learning with GRPO, on the interface alone, added 4.7 points on a harder set of word problems in a preliminary test.

What did the experiments show?

The authors picked tasks where the tool creates a large, clear gap between what the models can do with and without it.

• Arithmetic: two copies of Qwen2.5-0.5B-Instruct, with a calculator for the auxiliary, reached 96.5%, against 36.2% for the primary model alone. That is the best of 890 configurations of which layers to couple, and 95.4% of those beat the model alone. With the same interface but the auxiliary bypassed, the score was 48.0%, so the gain comes from the auxiliary's tool use rather than from the extra parameters.
• Logic puzzles: two copies of Qwen3-0.6B with the Z3 solver scored 64.7% on ZebraLogic, a benchmark of 1,000 logic grid puzzles, averaged over five runs. Qwen3-0.6B alone, with thinking, scored 37.5%, so the coupled system reached 1.7 times that. The same small model, prompted to write solver code as text, scored 24.6% and never got a successful query through to the solver.
• Math with Python: two copies of Qwen3-4B scored 62.5% on the MATH benchmark, below the 81.6% of Qwen3-4B alone with thinking. They did solve 127 problems that thinking alone missed, and the auxiliary wrote code for the specific problem without ever seeing the problem text.

For the puzzles, the paper adds the ZebraLogic leaderboard for context: Claude 3.5 Sonnet at 36.2% and GPT-4o at 31.7%, both without a solver. Trained without ZebraLogic-style clues, the coupled system dropped to 10.3% on ZebraLogic while scoring 81.1% on puzzles the authors generated themselves. The auxiliary leans heavily on clue patterns it has memorized.

What does the channel between the models carry?

The gate learned when to open from the task alone, without a prescribed format. In a word problem, the forward channel opened on words like "sum" and "total", and stayed shut on the dollar amounts, which the auxiliary had already received earlier. The reverse channel stayed near zero until the calculator returned its result. During training, the coupling developed in a fixed order: first the forward channel, then correct calculator calls, then a sudden jump in accuracy.

What are the limits?

• Running two models in lockstep can double the compute.
• The coupling hurts tasks the base model already handles well: on the GSM8K word problems, accuracy fell from 49.6% to around 40% across all 890 configurations.
• Training needs task-specific data, annotated with when the auxiliary may act.
• All reported results use two copies of the same model, of 0.5 to 4 billion parameters. A preliminary test with a 32B primary and a 0.6B auxiliary functioned, but it was not optimized, and the drop on word problems remained.

The authors want to pair other complementary models next, such as retrieval models or models for other modalities, and to train the interface further with reinforcement learning.

How does it compare with other ways to share hidden states?

The paper lists earlier approaches that let models exchange internal states instead of text. C2C fuses one model's KV cache into another's, once, before generation starts. LatentMAS passes full KV caches between agents in a sequential pipeline, without training. ThoughtComm injects latent thoughts between the rounds of a debate between agents. According to the authors, the Bicameral Model differs from all of them because its coupling runs in both directions, at every generated token.

Other multi-agent work from the same week still coordinated through text. EVOCHAMBER (11 May 2026) lets a pool of agents reflect together after a failed task and pass the lessons from strong to weak agents. DOLORES (12 May 2026) builds a reasoning scaffold for each task at inference time and spreads the work over separate reasoning threads.