Knowledge Graph Question Answering on KGQAGen-10k 1,079 unseen questions (test)
84.9LASM AccuracyGPT-4o w/ SP
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| GPT-4o w/ SPType=Oracle subgraph, Base Model=GPT-4o, Memory=Frozen2026.03 | 84.9 | — | |
| APEX-EM (EG1: Entity graph)Type=Rich feedback, iter., EG, Memory=Frozen2026.03 | 73.7 | 31.7 | |
| APEX-EM (A3: Judge + iteration)Type=Memory, binary, iteration, Memory=Frozen2026.03 | 73.5 | 31.5 | |
| APEX-EM (A5: Opus judge)Type=All + Opus judge, Memory=Frozen2026.03 | 72.7 | 30.7 | |
| APEX-EM (R1: Semantic only)Type=Rich feedback, iter., semantic, Memory=Frozen2026.03 | 71.6 | 29.7 | |
| APEX-EM (EG2: Full memory)Type=Rich feedback, iter., all, Memory=Frozen2026.03 | 71.2 | 29.2 | |
| APEX-EM (A5: GPT 4o as base)Type=All, Base Model=GPT-4o, Memory=Frozen2026.03 | 66.2 | 12 | |
| PoG (GPT-4o)Type=KG-RAG, plan-on-graph, Base Model=GPT-4o, Memory=Frozen2026.03 | 58.1 | — | |
| ToG (GPT-4o)Type=KG-RAG, think-on-graph, Base Model=GPT-4o, Memory=Frozen2026.03 | 56.3 | — | |
| GPT-4oType=Single-shot LLM, Memory=Frozen2026.03 | 54.2 | — | |
| GCR (GPT-4o)Type=KG-RAG, constrained gen., Base Model=GPT-4o, Memory=Frozen2026.03 | 54 | — | |
| APEX-EM (A1: Memory, no judge)Type=Memory, binary signal, Memory=Frozen2026.03 | 43.1 | 1.1 | |
| APEX-EM (A0: No Memory)Type=Sonnet 4.5, single-shot, Memory=Frozen2026.03 | 42 | — | |
| APEX-EM (A2: Memory + judge)Type=Memory, rich feedback, Memory=Frozen2026.03 | 41.8 | 0.2 | |
| RoG (LLaMA2-7B)Type=KG-RAG, graph reasoning, Base Model=LLaMA2-7B, Memory=Frozen2026.03 | 31.4 | — |