RecursiveMAS cuts multi-agent token usage 34.6% to 75.6%

A team from UIUC, Stanford, NVIDIA, and MIT has published RecursiveMAS, a multi-agent framework that treats a heterogeneous agent collective as a single unified latent-space recursive computation. It delivers an 8.3% average accuracy gain across nine benchmarks, cuts token usage by 34.6–75.6%, and accelerates end-to-end inference 1.2–2.4× over state-of-the-art text-based multi-agent baselines.

Recursive language models improve reasoning by iteratively refining the same computation over latent states rather than generating fresh token streams each pass. RecursiveMAS extends that loop across an entire pipeline: multiple heterogeneous agents iterate together, with each agent's latent output feeding directly into the next agent's input space rather than being decoded to text first.

A lightweight module called RecursiveLink solves two problems. First, it keeps each agent's generated latent thoughts in-distribution—critical because agents may differ in architecture or fine-tune history. Second, it manages cross-agent latent state transfer, passing compressed representations across the collaboration loop without the vocabulary-level serialization that dominates agentic pipelines. The system is optimized end-to-end by an inner-outer loop learning algorithm that performs whole-system co-optimization using shared gradient-based credit assignment across recursion rounds.

The efficiency gains matter most for practitioners. The 34.6–75.6% reduction in token usage compresses inference cost directly—a key lever for enterprises running high-throughput agentic workflows where per-token pricing dominates operating expense. The 1.2–2.4× speedup eliminates repeated tokenization and detokenization round-trips that impose serial latency in text-based pipelines. The paper's theoretical analysis establishes that RecursiveMAS maintains stable gradients during recursive training, addressing a practical concern that deeper recursion would complicate fine-tuning on proprietary data.

The benchmarks span mathematics, science, medicine, search, and code generation across nine datasets under four representative agent collaboration patterns. The consistent 8.3% average accuracy improvement across that spread shows the recursion mechanism is task-agnostic.

RecursiveMAS does not establish how the RecursiveLink module performs when agents are drawn from entirely different model providers—a common constraint when proprietary, open-weight, and vendor-hosted models coexist in one pipeline. The inner-outer loop training requires gradient access to all agents in the loop, which rules out black-box API-only deployments without additional engineering. The team released code and data at recursivemas.github.io.