| layout | home |
|---|---|
| title | HongLin He |
English
中文简体
中文繁體
日本語
한국어
Français
Deutsch
Español
🌙
Independent researcher in hyperbolic geometry, quantum-inspired IR, and LLM context extension
Density matrix representations and measurement collapse for context selection
Adaptive compression with closed-loop quality feedback for long-context processing
I designed UCEF, a model-agnostic framework enabling any LLM to process documents far exceeding its native context window through hyperbolic retrieval, quantum-inspired selection, adaptive compression, and quality feedback.
Results on LongBench (8 tasks × 30 samples/task × 2 models):
| Model | Method | ROUGE-L | TokenF1 | vs RAG |
|---|---|---|---|---|
| GLM-4-flash | RAG top-k | 0.1340 | 0.1498 | — |
| GLM-4-flash | UCEF | 0.1479 | 0.1671 | +10.3% |
| DeepSeek-v3 | RAG top-k | 0.1800 | 0.1919 | — |
| DeepSeek-v3 | UCEF | 0.2146 | 0.2306 | +19.3% |
Statistical significance (DeepSeek-v3): Wilcoxon p=0.011, paired t-test p=0.007 ✅
📄 GitHub · 📊 Documentation · 📝 Paper
- May 2026 — Real benchmark experiments: UCEF on LongBench with GLM-4-flash and DeepSeek-v3
- Apr 2026 — FeMeter firmware Phase 1-4 complete, FreeRTOS 11.x migration done
- Mar 2026 — UCEF v2026.5.2 released: all phases complete
- Jan 2026 — Competitive analysis of hyperbolic RAG landscape