Fact Consolidation on MAB FC-SH (262K context) v3 (full)
93Accuracy (SubEM)SH-conflict (fact + Python max)
Evaluation results
| Method | Links | ||
|---|---|---|---|
| SH-conflict (fact + Python max)architecture=SH fact + Python max, backbone=gpt-4o2026.05 | 93 | — | |
| SH-conflict (fact + Python max)architecture=SH fact + Python max, backbone=gpt-4o-mini2026.05 | 82 | — | |
| SH-conflict (chunk4096 + Python max)architecture=SH chunk4096 + Python max, backbone=gpt-4o-mini2026.05 | 73 | — | |
| GPT-4omode=long-context2026.05 | 60 | -33 | |
| HippoRAG-v2retrieval=hippocampal retrieval2026.05 | 54 | -39 | |
| BM25retrieval=simple lexical RAG, chunking=MAB chunk-512 default2026.05 | 48 | -45 | |
| GPT-4o-minimode=long-context FIFO2026.05 | 45 | -48 | |
| Claude-3.7-Sonnetmode=long-context2026.05 | 43 | -50 | |
| GPT-4.1-minimode=long-context2026.05 | 36 | -57 |