The Big Picture
Build a union of diverse agent team layouts, but only let small, trusted pieces talk at a time and locally fix risky parts — that cuts distracting noise and improves final answers (avg +3.47%, up to +10.34%).
ON THIS PAGE
The Evidence
Selecting a small set of complementary collaboration patterns and merging them into a union keeps diverse reasoning paths available. complementary collaboration patterns Limiting active connections at each step and exchanging compact summaries prevents irrelevant signals from spreading. A local self-check step then spots and repairs unreliable agent outputs, producing more accurate results while keeping cost practical.
Data Highlights
1Average performance gain of 3.47% over the strongest baseline across eight benchmarks
2Maximum single-task improvement of 10.34% on the hardest benchmarks
3Evaluated on 8 benchmarks covering knowledge, math, and code (e.g., MMLU, GSM8K, HumanEval)
What This Means
Engineers building multi-agent orchestration and evaluation pipelines who want better accuracy without exploding inter-agent chatter. Technical leaders and researchers focused on multi-agent trust or agent reliability can use these ideas to make agent teams both more diverse in reasoning and less prone to noisy feedback.
Not sure where to start?Get personalized recommendations
Key Figures
Fig 1: Figure 1: Problem Illustration: I) A single topology restricts reasoning to a narrow trajectory. II) Composite graph propagates irrelevant noise across redundant connections.
Fig 2: Figure 2: The designing workflow of our proposed HELENA .
Fig 3: Figure 3: Cost–accuracy comparison and difficulty-dependent scaling of HELENA . Left: MATH and MATH-Lv5. Right: MMLU and MMLU-Pro.
Fig 4: Figure 4: Sensitivity analysis of HELENA . Left: performance and relative token cost as the number of fused topologies K K varies. Right: Right: performance and relative token cost as the lower-confidence quantile α \alpha varies.
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreConsiderations
The approach depends on a search budget to generate candidate topologies and uses frozen priors during inference, so quality depends on the offline search and scoring stages. The union-graph step reduces but does not eliminate the risk of including useless connections if complementary selection fails. Real-world tasks with different cost or latency constraints may require tuning how many topologies to fuse and how aggressive the sparsity is. scoring stages
Methodology & More
Helena (Hierarchical Sparse Coordination over a Union of Complementary Topologies) tackles a key trade-off: one fixed agent layout narrows reasoning, while naively merging many layouts spreads irrelevant noise. The method first generates many candidate team layouts by making small edits (role changes, prompt tweaks, or edge edits) guided by a search procedure that estimates which edits look promising for the query. From that pool, it selects a complementary subset — chosen to cover diverse reasoning perspectives — and merges them into a single union graph so multiple viewpoints remain available. search procedure To stop the union from turning into a noisy tangle, Helena activates only a sparse subgraph at each communication step and has agents exchange compressed latent briefs instead of full verbose messages. After agents produce candidate solutions, a local self-refinement stage selects reliable proposals, flags high-risk units, and revises them based on clear evidence and alternatives. Across eight benchmarks in knowledge, math, and code domains, this combination yields an average improvement of 3.47% and up to 10.34% on the hardest tasks, while remaining cost-competitive. The pattern is practical: keep diversity in structure, but limit who actually talks and add local checks to catch mistakes. Blackboard Pattern
Avoid common pitfallsLearn what failures to watch for
Credibility Assessment:
No author h-indexes or affiliations provided and only an arXiv preprint with zero citations—insufficient signals of credibility.