ResearchBenchmarksBenchmark-topology generalisation on AgentBench topology simulation seed=42 (n=50 per topology)Benchmark-topology generalisation on AgentBench topology simulation seed=42 (n=50 per topology)Follow1.39ρ(B)-reductionCascade0.85960.99731.1351.2727Jul 2, 2026Evaluation resultsMethodMethodLinksρ(B)-reductionMean Region SizeCascade2026.071.3912WM-SAR2026.071.3911LR-3Hhops=3hops=32026.071.3512LR-2Hhops=2hops=22026.071.2611Window-4window_size=4window_size=42026.071.084TopK-5K=5K=52026.071.065Greedy-Pt2026.070.881