ABF-Bench makes blockchain fund tracing a verifiable LLM-agent task
Original title: ABF-Bench: A Verifiable Benchmark for Agentic Blockchain Forensics
Authors report 75% edge recall for their harness, versus 47% for Codex and 10% for Claude Code.
Researchers introduce ABF-Bench, an interactive benchmark that turns evidence-backed fund tracing into a verifiable agent task. Every claimed movement must cite ledger records the agent retrieved, and reference graphs come from conservative rules that stop when records leave a continuation unresolved. Tasks drawn from real laundering include multi-hop transfers, DeFi exchanges, cross-chain round trips, custodial and mixer deposits, and address poisoning. Their harness reports 75% edge recall and 98% fidelity, yet agents stop early and trail the best tracer on larger tasks.
Unconfirmed: Reported scores come from the authors' evaluation; independent replication is not stated.
- The harness hits 75% edge recall, Codex 47%, and Claude Code 10%.
- It recognizes undisclosed custodial services at 89%, against 0% for fixed tracers.
- Claude Code cites ledger records it left unretrieved, cutting recall through overconfidence.
- The harness acquires evidence faster than the best tracer early in the budget.
- Agents tend to stop early and lose recall on larger tasks.
From the original article
1 min read at huggingface.co
Blockchain forensics is adversarial and open-ended: money launderers split, swap, and bridge funds, cut their trail at mixers or custodial services, and change tactics from one incident to the next. Investigators must keep adapting their tracing in response. Yet the immutability of public ledgers preserves every recorded movement in public view, making much of the investigation independently verifiable.
By Xiaohang Yu, Hejia Geng, William Knottenbelt · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.