RecursiveMAS: Scaling Multi-Agent Collaboration Through Latent-Space Recursion

Researchers from the University of Illinois Urbana-Champaign, Stanford University, NVIDIA, and the Massachusetts Institute of Technology have come together to build RecursiveMAS. It's a recursive multi-agent framework that treats the entire system as a single latent-space recursive computation. The work, currently available as a preprint on arXiv, extends the idea of recursive language models from single models to multi-agent systems, asking whether agent collaboration itself can be scaled through recursion.

Why Latent-Space Recursion?

Here's the core insight: most computation in multi-agent systems happens in text space. Each agent has to decode its thoughts into tokens before passing them to the next agent. That's expensive, and it introduces a semantic bottleneck. RecursiveMAS keeps all intermediate reasoning entirely in continuous latent space, only decoding to text in the final round.

The framework connects heterogeneous agents through a lightweight RecursiveLink module, a two-layer residual network that handles two types of transitions. Inner links map an agent's last-layer hidden state back into its input embedding space for continued latent generation. Outer links transfer latent thoughts between agents, even when those agents have different hidden sizes. The residual design preserves original latent semantics, so the module only needs to learn the distributional shift.

Architecture and Training

RecursiveMAS chains all agents into a single recursive loop. Each agent uses the inner link to generate latent thoughts, then passes them to the next agent via the outer link. After the last agent finishes, its latent outputs feed back to the first agent, closing the loop. Training uses a two-stage inner-outer loop algorithm. The inner loop warm-starts each agent's inner link with a regression objective that aligns latent thoughts with the ground-truth answer's input embedding distribution. The outer loop unrolls the full system for multiple recursion rounds and optimizes all RecursiveLinks via backpropagation through the entire recursive trace. Only about 13 million parameters, roughly 0.31% of the full system, are trainable, with base LLM parameters frozen.

Theoretical Guarantees

The paper delivers two key theoretical results. First, a runtime complexity analysis shows that text-based recursive multi-agent systems pay a per-step decoding cost proportional to the vocabulary size. RecursiveMAS replaces this with a much cheaper latent-space transformation, removing the vocabulary projection bottleneck. Second, a gradient stability theorem shows that under realistic assumptions with confident token predictions, text-based recursive supervised fine-tuning suffers from vanishing gradients, while RecursiveMAS maintains stable, near-constant gradients across looped backpropagation, enabling stable whole-system co-optimization.

Performance and Efficiency Gains

Evaluated across nine benchmarks covering mathematics, science, medicine, code, and search, RecursiveMAS consistently outperforms strong baselines. On MATH500, AIME 2025, AIME 2026, GPQA-D, LiveCodeBench, and MedQA, the framework delivers an average accuracy improvement of 8.3% over the best baseline under matched training budgets. Notably, on the logic-dense AIME 2025 competition math tasks, RecursiveMAS achieves an 18.1% improvement, and on the held-out code generation benchmark MBPP+, it shows a 13.0% gain.

The efficiency improvements become even more striking at deeper recursion depths. At recursion round 3, RecursiveMAS achieves up to a 2.4× end-to-end inference speedup compared to text-based recursive multi-agent systems, while reducing token usage by 75.6%. These advantages compound as recursion deepens, because most rounds happen entirely in latent space, avoiding repeated text decoding.

Generalization Across Collaboration Patterns

RecursiveMAS is structure-agnostic. It's been instantiated under four representative collaboration patterns: sequential (planner-critic-solver), mixture (domain experts with a summarizer), distillation (expert-learner), and deliberation (reflector and tool-caller). Across all patterns, the framework improves over the strongest standalone agent, by 6.2% in mixture style, 4.8% in deliberation style, and 8.0% in distillation style, where it lifts a smaller learner while preserving a 1.5× speed advantage over a larger expert.

Implications for Multi-Agent Systems

The work points to a new scaling axis for multi-agent systems: instead of simply adding more agents or larger models, recursion depth in latent space can yield significant performance and efficiency gains. The approach also opens the door to more flexible agent coordination without the overhead of decoding to text at every step, potentially making complex multi-agent architectures more practical for resource-constrained deployments.