Multi-domain knowledge reasoning on HLE 500-question ablation
57.3Success Rate (Last)MemRL
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| MemRLBackbone=Gemini-3-pro2026.03 | 57.3 | 61.3 | |
| MemPBackbone=Gemini-3-pro2026.03 | 52.8 | 58.2 | |
| Mem0Backbone=Gemini-3-pro2026.03 | 51.2 | 56 | |
| RAGBackbone=Gemini-3-pro2026.03 | 50 | 54.8 | |
| Self-RAGBackbone=Gemini-3-pro2026.03 | 48.8 | 54.8 | |
| APEX-EM: Entity graph (A3, E10)Backbone=Claude Opus 4.52026.03 | 48 | 53.3 | |
| APEX-EM: Full memory (A2, E10)Backbone=Claude Opus 4.52026.03 | 46.8 | 52.3 | |
| APEX-EM: Semantic only (A1, E10)Backbone=Claude Opus 4.52026.03 | 45.6 | 52.3 | |
| APEX-EM: Judge + iteration (A5, E10)Backbone=Claude Opus 4.52026.03 | 40.4 | 52.9 | |
| No MemoryBackbone=Gemini-3-pro2026.03 | 35.7 | — | |
| APEX-EM: No Memory (A0)Backbone=Claude Opus 4.52026.03 | 25.2 | — | |
| APEX-EM: Memory, no judge (A4, E10)Backbone=Claude Opus 4.52026.03 | 19.4 | 47.9 | |
| Pass@10Backbone=Gemini-3-pro2026.03 | — | 52.4 | |
| ReflexionBackbone=Gemini-3-pro2026.03 | — | 53 |