MANTA: Multi-Agent Network Topology Adaptation for Self-Evolving Multi-Agent Systems¶
Deterministic topology mini-suite, not the full external environment.
Paper information¶
| Field | Details |
|---|---|
| Paper | Paper or original source |
| Institution | Cornell University; first author MaoXun Huang. |
| First public release | 2026-07-30 (arXiv v1) |
| Author code | No public author code found as of 2026-08-24. |
| Local adapter or method key | manta |
| Local reproduction code | src/auto_research/agent_research/historical_b10_b11.py |
Paper method and evidence¶
Initialize communication topology from task priors and monitor collaboration traces. Adapt roles, edges, ordering, visibility and verification under bounds.
flowchart LR
A[Initialize from task priors] --> B[Monitor collaboration trace]
B[Monitor collaboration trace] --> C[Bounded topology adaptation]Original paper figure¶
Figure shows topology initialization and adaptation. From the original paper; copyright remains with the authors.
Core equation¶
Reported results¶
Paper average 74.0 over five benchmarks, 5.8 points above the strongest baseline; best on PlanCraft.
Local reproduction¶
PlanBench mini, 120 episodes: MANTA 1.0000 joint success, mean cost 0.6600; long-context 1.0000 and 64.5000. Metrics, Batch index. With saturated accuracy, inspect traces, provenance and method-specific counters rather than claiming gains. No real browser, AppWorld, SWE-bench or external services. Raw outputs stay under runs/agent-research/; only metrics are committed.