Multi-stage Attack Detection on CSTM-Bench (cross-session shard)
73.1Detection RateCoreset Memory Reader
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Coreset Memory ReaderReader=Coreset Memory Reader, Model=Claude Sonnet 4.6, Coreset size=K=502026.04 | 73.1 | 73.1 | 80.4 | 69 | 0 | 21.4 | |
| Per-session judgeReader=Per-session judge2026.04 | 53.8 | 53.8 | 55.4 | 68.8 | 21.4 | 64.3 | |
| Full-Log CorrelatorReader=Full-Log Correlator, Model=Claude Sonnet 4.62026.04 | 50 | 50 | 60 | 72 | 0 | 50 |