Single-Hop Fact-based Reasoning on MAB FC-SH 262K v3 (test)
93AccuracyHeadline pipeline
Evaluation Results
| Method | Links | |
|---|---|---|
| Headline pipelineBackbone=gpt-4o2026.05 | 93 | |
| Headline pipelineBackbone=gpt-4o-mini2026.05 | 82 | |
| Ablation ABackbone=gpt-4o-mini, Pipeline architecture=chunk-40962026.05 | 73 | |
| LLM-judgment baselineBackbone=gpt-4o-mini2026.05 | 61 | |
| GPT-4o (long-context)Backbone=gpt-4o, Pipeline architecture=long-context2026.05 | 60 | |
| HippoRAG-v2 (best published)Backbone=gpt-4o-mini2026.05 | 54 | |
| BM25Backbone=gpt-4o-mini, Pipeline architecture=MAB chunk-512 default2026.05 | 48 | |
| Zep / GraphitiBackbone=gpt-4o-mini, Pipeline architecture=temporal KG2026.05 | 7 |