RecursiveMAS: AI Agents Pass Hidden States Instead of Text
RecursiveMAS links small open models into a loop that exchanges hidden states instead of text. In the authors' tests it was more accurate than a text loop and used up to 75.6 percent fewer tokens.
TL;DR
- RecursiveMAS chains small open models into a loop that exchanges hidden states instead of text, training only a small connector module [^1].
- Against the same agents looping in text, it was more accurate, up to 2.4 times faster and used up to 75.6 percent fewer tokens [^1].
Background
A multi-agent system splits a task among several language models with different roles, such as a planner, a critic and a solver, which usually pass text to each other [^1]. Separately, recursive or looped language models deepen their reasoning by running the same layers several times over hidden states, the numeric vectors a model computes before it picks a word [^1]. The new paper asks whether collaboration between agents can be scaled the same way [^1].
What happened
Researchers from UIUC, Stanford University, NVIDIA and MIT posted RecursiveMAS on arXiv on April 28 [^1]. Each agent is an off-the-shelf open model whose weights stay frozen. Instead of writing text, an agent generates a sequence of latent thoughts, hidden states it feeds back into itself, and a small module called RecursiveLink passes them to the next agent, translating between models of different sizes [^1]. After the last agent, the result loops back to the first, and only in the final round does the last agent write a text answer [^1]. Training touches only the links, 13.12 million parameters or 0.31 percent of the system in one setup: first for each agent, then for the whole loop, with gradients flowing back through all rounds [^1].
The team tested four collaboration patterns on nine benchmarks in math, science, medicine, search and code: a planner, critic and solver in sequence; math, code and science specialists with a summarizer; a large expert paired with a smaller learner; and a reflector paired with a tool-calling agent [^1]. The agents are small open models, from Llama3.2-1B to Qwen3.5-9B, drawn from the Qwen, Llama, Gemma and Mistral families [^1]. Against Recursive-TextMAS, the same agents looping in text, RecursiveMAS was on average 8.1 percent more accurate after one round and 20.2 percent after three. It was 1.2 times faster after one round and 2.4 times after three, and used 34.6 to 75.6 percent fewer tokens, because the text version decodes intermediate text in every round [^1].
In a wider comparison after three rounds, RecursiveMAS beat single agents fine-tuned on the same data, TextGrad, Mixture-of-Agents and the looped model LoopLM on all six benchmarks, by 8.3 percent on average over the strongest rival on each [^1]. On AIME2025, scored as the share of problems solved in any of 10 attempts, it reached 86.7 against 73.3 for the best alternatives [^1]. Training the links cost an estimated $4.27, against $6.64 for LoRA and $9.67 for full fine-tuning, and gave the highest average accuracy of the three [^1].
Why it matters
The paper's case is that text is an expensive channel between models. Every intermediate message has to be decoded word by word, and the authors argue theoretically that training through text makes gradients vanish across rounds, while their links keep them stable [^1]. Their numbers show what that means for adding rounds. In the text baseline, more rounds cost more time and tokens and sometimes lowered accuracy: on AIME2025 with the light agents, it fell from 24.0 to 18.0 percent over three rounds, while RecursiveMAS rose from 30.7 to 34.0 [^1].
The trade-offs follow from the design. Because agents exchange hidden states, the method needs models whose internals you can reach and train links for, like the open models in the paper; a model that is only available as a text API doesn't fit. The loop also has to be trained on a task mix, here public sets for math, medicine, science, code and tool use [^1]. Since the agents talk only in vectors until the final round, there is no readable transcript of their exchange to inspect. And the results come from the authors alone, on small models, as means over five runs [^1].
Practical example
Say Lena runs a startup that explains math homework, with three open models on its own GPUs: a planner breaks each problem down, a critic checks the plan and a solver writes the answer. Every handoff is written out as text, so running the chain for three rounds multiplies tokens and waiting time.
With RecursiveMAS, Lena keeps the three models frozen and trains only the RecursiveLink modules on worked math problems, as the paper did with the s1K set [^1]. Now the planner's latent thoughts flow straight into the critic, then the solver, and back to the planner, for three rounds; only the solver's final answer is turned into text [^1]. The paper's light planner, critic and solver setup shows what to expect on MATH500: after three rounds, RecursiveMAS took 1,360 seconds against 2,952 for the text loop, at 77.8 versus 69.1 percent accuracy [^1].
The catch shows up when a parent reports a wrong explanation. Lena's team can read the final answer, but not what the planner and critic told each other, because that exchange only existed as hidden states [^1]. So the team logs every disputed problem and reruns it through the old text chain, where each step can still be read.
Related gear
We recommend this book because it explains hidden representations and why gradients vanish through long chains of computation, the two ideas the RecursiveMAS design rests on.