Trace Reconstruction on ATLAS (D3)
80.8RecallTool-Aug. ReAct
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Tool-Aug. ReActMethodology Category=LLM & agentic baselines, Backbone=GPT-4o, Log-suppression rate=30%2026.05 | 80.8 | 49.2 | 61.2 | 56.4 | |
| HunterAgentMethodology Category=LLM & agentic baselines, Backbone=GPT-4o, Log-suppression rate=30%2026.05 | 79.4 | 90.8 | 84.7 | 7.4 | |
| HunterAgent†Methodology Category=LLM & agentic baselines, Backbone=GPT-4o, Log-suppression rate=30%2026.05 | 78 | 67.5 | 72.4 | 24.8 | |
| KAIROSMethodology Category=Recent learned provenance / streaming, Log-suppression rate=30%2026.05 | 65.8 | 77.8 | 71.3 | 10.5 | |
| GraphRAGMethodology Category=LLM & agentic baselines, Backbone=GPT-4o, Log-suppression rate=30%2026.05 | 61 | 34.5 | 44.1 | 70.2 | |
| NODLINKMethodology Category=Recent learned provenance / streaming, Log-suppression rate=30%2026.05 | 60.5 | 82.7 | 69.8 | 6 | |
| GNN-LinkPredictMethodology Category=Statistical relational learning, Log-suppression rate=30%2026.05 | 47.5 | 72 | 57.2 | 29.8 | |
| MAGICMethodology Category=Deterministic provenance graphs, Log-suppression rate=30%2026.05 | 38.7 | 98 | 55.6 | 0.6 | |
| Zero-Shot LLMMethodology Category=LLM & agentic baselines, Backbone=GPT-4o, Log-suppression rate=30%2026.05 | 38.5 | 15.6 | 22.2 | 88.5 | |
| SLEUTHMethodology Category=Deterministic provenance graphs, Log-suppression rate=30%2026.05 | 31.2 | 98.8 | 47.5 | 0 |