AMLBench: A Real-Incident Benchmark for Agentic Blockchain Anti-Money Laundering (arXiv (Cornell University))

Summary

• AMLBench is an anti-money laundering benchmark constructed from 24 real-world cases and 5.05 million Ethereum transfer records.p.1
• Using GPT-6 Astra, AMLAgent achieves macro averages of 80.82% transfer-level precision and 84.73% query coverage across 24 cases.p.1
• The benchmark covers cryptocurrency security incidents that took place between 2018 and 2025.p.3
• AMLAgent reaches a transfer-level F1 score above 70% in 15 of the 24 cases, with a median F1 of 75.48%.p.6
• Bybit is the largest case in the benchmark, containing approximately 2.62 million transfer records.p.6

Key insights

1. Large language model agents can automate the reconstruction of blockchain fund-transfer networks from limited initial clues.p.1
2. AMLAgent discovers new case-related addresses beyond the initial clues in 23 of the 24 cases.p.6
3. Large-scale cases including Upbit, PlusToken, and Bybit each contain over 48,000 target transfers and remain challenging, with recall staying below 2%.p.6
4. Selecting records based on fund-flow connections improves submission precision in 20 of the 24 cases.p.7
5. Fewer tool calls do not necessarily imply lower token consumption, as ATO uses 292 tool calls for 0.95 million tokens while Lendf.Me uses 50 calls for 2.45 million tokens.p.8
6. In the multi-model comparison on Bybit, GPT-6 Astra retained 52.24% of observed relevant transfers, whereas other models retained less than 5%.p.8