ABF-Bench: A Verifiable Benchmark for Agentic Blockchain Forensics
The authors’ agent harness recovered 75% of fund movements, but still finished below the best fixed-order tracer.
ABF-Bench requires every claimed movement to cite ledger records the agent retrieved. Its tasks draw on real laundering activity, with tracing halted wherever the records leave the next step unresolved. The harness recognized undisclosed custodial services in 89% of cases, compared with 0% for fixed-order tracers. Agents tended to stop early on larger tasks, leaving parts of the money trail uncovered.
HF Daily Papers' note
ABF-Bench requires every claimed movement to cite ledger records the agent retrieved. Its tasks draw on real laundering activity, with tracing halted wherever the records leave the next step unresolved. The harness recognized undisclosed custodial services in 89% of cases, compared with 0% for fixed-order tracers. Agents tended to stop early on larger tasks, leaving parts of the money trail uncovered.
HF Daily Papers' note
score 4