Long-horizon million-token stress test on LOCA-Bench
50.7AccuracyVISTA (full)
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| VISTA (full)State=✓, Ctrl=✓, Recov=✓2026.06 | 50.7 | 2.86 | |
| VISTA w/o recoveryState=✓, Ctrl=✓, Recov=✗2026.06 | 45.3 | 2.99 | |
| Auto-Archive + RecoverState=✗, Ctrl=✗, Recov=✓2026.06 | 44 | 2.73 | |
| Claude CodeState=✗, Ctrl=✓, Recov=∼2026.06 | 42.7 | 6.72 | |
| VISTA w/o dashboardState=✗, Ctrl=✓, Recov=✓2026.06 | 37.3 | 5.25 | |
| Active Ctx. CompressionState=✗, Ctrl=✓, Recov=✗2026.06 | 36 | 3.2 | |
| Context-FoldingState=✗, Ctrl=✓, Recov=✗2026.06 | 34.7 | 3.41 | |
| Skeleton CompressionState=✗, Ctrl=✗, Recov=✗2026.06 | 33.3 | 2.84 | |
| SLIM (summary)State=✗, Ctrl=∼, Recov=✗2026.06 | 29.3 | 3.76 | |
| Stale-obs. MaskingState=✗, Ctrl=✗, Recov=✗2026.06 | 28 | 3.32 | |
| Tool-result ClearingState=✗, Ctrl=✗, Recov=✗2026.06 | 26.7 | 2.6 | |
| ReActState=✗, Ctrl=✗, Recov=✗2026.06 | 22.7 | 3.51 |