HELENA: New multi-agent AI framework crushes benchmarks
Researchers propose HELENA, a hierarchical sparse coordination framework for LLM-based multi-agent systems that outperforms baselines by up to 10.34% on MMLU-Pro.
Researchers from three Chinese institutions have unveiled HELENA (Hierarchical Sparse Coordination over a Union of Complementary Topologies for MAS), a groundbreaking framework designed to supercharge multi-agent AI systems. Unlike traditional approaches that lock multi-agent systems (MAS) into a single reasoning topology—limiting analytical depth—HELENA dynamically constructs a union graph from complementary candidate topologies using Monte Carlo Tree Search and Determinantal Point Processes.
The system then employs a hierarchical sparse coordination module that activates only task-relevant agent subgraphs at each step, exchanging compressed latent summaries to minimize noise. A final Local Self-Refinement stage identifies and corrects decision discrepancies when evidence confirms failure and improvement. Across eight benchmark tests, HELENA delivered state-of-the-art results, averaging 3.47% gains over the strongest baseline and reaching 10.34% on MMLU-Pro, with greater improvements on harder tasks at manageable additional compute cost.
- HELENA uses Monte Carlo Tree Search and Determinantal Point Processes to build optimal multi-agent topologies
- Achieved average 3.47% benchmark improvement and up to 10.34% on MMLU-Pro
- Hierarchical sparse coordination reduces noise by activating only relevant agent subgraphs
Why It Matters
HELENA represents a leap in multi-agent AI reasoning, enabling more accurate, comprehensive analysis for complex real-world problems while maintaining computational efficiency.