Editorial illustration for RecursiveMAS cuts multi-agent inference time 2.4×, slashes token use 75%
RecursiveMAS cuts multi-agent inference time 2.4×,...
Everyone's building AI agents now, and they're getting expensive fast. A new paper offers a direct, boring fix: cut the talking.
Researchers introduced RecursiveMAS, a framework that makes multiple language models collaborate without generating text at every step. They tested it on nine benchmarks in math, science, medicine, code, and search Q&A. The system used open-weight models like Qwen, Llama-3, Gemma3, and Mistral, assigning each a specific role.
Because it skips the text generation, inference runs 1.2 to 2.4 times faster. Token use plummets. The first round cuts usage by 34.6%, and by the third round, the reduction hits 75.6% compared to a text-based baseline.
Training is lighter too. Only the RecursiveLink modules are updated—about 13 million parameters, a mere 0.31% of the frozen models' trainable weights. This design lowers peak GPU memory and chops training costs by more than half versus a full fine-tuning.
RecursiveMAS was compared to baselines under identical training budgets, including standalone models enhanced with LoRA or full supervised fine-tuning, alternative multi-agent frameworks like Mixture-of-Agents and TextGrad, and recursive baselines like LoopLM. It was also compared to Recursive-TextMAS, which uses the same recursive loop structure as RecursiveMAS but forces the agents to explicitly communicate via text.
RecursiveMAS achieved an average accuracy improvement of 8.3% compared to the strongest baselines across the benchmarks. It excelled particularly on reasoning-heavy tasks, outperforming text-based optimization methods like TextGrad by 18.1% on AIME2025 and 13% on AIME2026.
The promise is real speed and cost savings. A 2.4x inference speedup with a 75% drop in token use would lower compute bills for many teams. The system consistently beat other methods under the same training budget.
But the paper doesn't report real-world deployment metrics. It's unclear if the token savings mean proportional energy or latency benefits on production-scale workloads. The evaluation also doesn't cover scaling beyond a handful of agents, or using proprietary models with different licensing. For now, it's a promising research direction, not a proven commercial tool.
Further Reading
Related Reading
Research & Benchmarks Google's FACTS benchmark shows 70% factuality ceiling across four tests
Research & Benchmarks Databricks finds multi-step agents beat single-turn RAG by 21% to 38% on STaRK
Research & Benchmarks Nvidia's DLSS 4.5 beta adds 6x Multi Frame Generation for RTX 50 GPUs
From Archives ArXiv to ban authors of papers with unchecked LLM‑generated content
From Archives BenchJack proposes secure-by-design AI benchmark audit with eight flaw taxonomy
Common Questions Answered
How does RecursiveMAS reduce inference time and token usage compared to traditional multi-agent systems?
RecursiveMAS achieves a 2.4x inference speedup and 75% reduction in token usage by eliminating unnecessary text generation between agent steps. Instead of having multiple language models generate text at every collaboration point, the framework enables direct communication between agents, significantly reducing computational overhead and token consumption.
What types of tasks was RecursiveMAS tested on, and which models were used?
RecursiveMAS was evaluated on nine benchmarks spanning math, science, medicine, code, and search Q&A tasks. The researchers used open-weight models including Qwen, Llama-3, Gemma3, and Mistral, with each model assigned a specific role within the multi-agent framework.
What are the practical cost implications of RecursiveMAS's token reduction for AI teams?
The 75% drop in token usage with RecursiveMAS would substantially lower compute bills for teams running multi-agent systems. This efficiency gain makes deploying multiple language models more economically viable, as the reduced token consumption directly translates to lower infrastructure and API costs.
What limitations does the RecursiveMAS paper acknowledge regarding real-world deployment?
The paper does not report real-world deployment metrics, making it unclear whether the token savings translate to proportional energy or latency benefits in production-scale workloads. Additionally, the evaluation does not cover scaling beyond a handful of agents or performance with proprietary models, which limits conclusions about broader applicability.
Further Reading
- Recursive Multi-Agent Systems - arXiv — arXiv
- Paper page - Recursive Multi-Agent Systems — Hugging Face
- Recursive Multi-Agent Systems - arXiv — arXiv
- Scaling Agent Collaboration via Recursion — StartupHub.ai
- Offical Implementation for "Recursive Multi-Agent Systems" — GitHub
