DeXposure-Claw brings forecast-grounded LLM supervision to DeFi credit risk
Researchers Aijie Shu, Bowei Chen, and Wenbin Wu introduce DeXposure-Claw, an agentic supervision system that routes LLM decisions through structured forecasts and confidence gates to reduce false alarms in DeFi risk oversight.
Score breakdown
The system replaces unconstrained LLM escalation with a structured, forecast-grounded pipeline and introduces a regulator-aligned evaluation metric for false interventions — two gaps the authors identify as absent from existing DeFi supervision approaches.
- 01DeXposure-Claw is an agentic supervision system for DeFi credit risk, introduced by Aijie Shu, Bowei Chen, and Wenbin Wu.
- 02General-purpose LLM agents over-read weak evidence and recommend high-stakes interventions in DeFi settings, motivating the work.
- 03DeXposure-FM, a graph time-series foundation model, forecasts future exposure networks as the first stage of the pipeline.
Aijie Shu, Bowei Chen, and Wenbin Wu identify a core problem with applying general-purpose LLM agents to DeFi risk supervision: these agents over-read weak evidence and recommend high-stakes interventions, while existing evaluation frameworks offer no regulator-aligned way to measure the resulting false alarms. Their proposed solution, DeXposure-Claw, structures LLM decision-making through a three-stage pipeline. First, DeXposure-FM — a graph time-series foundation model — forecasts future exposure networks. Second, deterministic monitors and stress scenarios convert those forecasts into typed alerts, attribution signals, and scenario evidence. Third, data-health and confidence gates constrain escalation before the system emits auditable supervisory tickets with rationales.
Alongside the system, the authors introduce DeXposure-Bench, a six-axis evaluation harness.
Alongside the system, the authors introduce DeXposure-Bench, a six-axis evaluation harness. Its decision axis scores supervisory tickets against a regulator-aligned absolute-loss ground truth and includes an explicit false-intervention rate metric — filling a gap the authors identify in existing evaluation approaches. The full system was validated on five years of weekly real DeFi data, and the code is publicly available at https://github.com/EVIEHub/DeXposure-Claw.
Key facts
- 01DeXposure-Claw is an agentic supervision system for DeFi credit risk, introduced by Aijie Shu, Bowei Chen, and Wenbin Wu.
- 02General-purpose LLM agents over-read weak evidence and recommend high-stakes interventions in DeFi settings, motivating the work.
- 03DeXposure-FM, a graph time-series foundation model, forecasts future exposure networks as the first stage of the pipeline.
- 04Deterministic monitors and stress scenarios convert forecasts into typed alerts, attribution signals, and scenario evidence.
- 05Data-health and confidence gates constrain escalation before auditable supervisory tickets with rationales are emitted.
- 06DeXposure-Bench is a six-axis evaluation harness with a decision axis scored against a regulator-aligned absolute-loss ground truth and an explicit false-intervention rate.
- 07Experiments were conducted on five years of weekly real DeFi data; code is publicly available on GitHub.
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 19, 2026 · 10:25 UTC. How this works →