Public KG record
Zyphra Hierarchical Attention (ZHA)
Recorded overview
Sparse attention method using a tree-structured indexer to search for relevant k
An evidence-backed entity from the project knowledge graph; detailed tool metadata is still being added.
Recorded details
- Benchmarks: LongBench v2 (deep tree vs dense: 63.4% vs 58.0%)
- Benchmarks: RULER-13 (deep tree vs dense: 93.3% vs 96.0%)
- Benchmarks: LongBench v2（thinking）63.4%（deep tree）
- Benchmarks: RULER-13（256K）deep tree 93.3%，wide tree 96.2%
- Benchmarks: 在 Nemotron-3.5-Lightning-30B-A3B 上 deep-tree RULER-13 比 dense 低 25.3 分
- Benchmarks: 28 种树几何下，读取 10% 至 30% 键时 RULER 准确率保持稳定
- Benchmarks: LongBench v2
- Benchmarks: RULER-13
Evidence coverage
The KG currently records 10 known information dimensions and has 9 undocumented dimensions, linked to 1 public source articles. These counts describe record coverage, not model performance or verified fit.
Related public records
- Zyphra — releases
- MiniMax Sparse Attention (MSA) — competes with
- Qwen3.6-35B-A3B — applied to
- MiniMax Sparse Attention (MSA) — competes with
- LongBench v2 — evaluated on