Proposal: Symbolic Activation — 95% Compression & No Context Rot

Hi @guosheng @deepseek-ai,

I'm GCat, and I've published research identifying **Context Rot**—a critical flaw in KV caching that degrades LLM performance in long conversations.

Key findings:
- After 15+ turns / 32k tokens, effective info utilization drops below 40%
- Root cause: Transformer + explicit KV cache, not long-context capability
- Solution: **Symbolic Activation Framework** → 95% compression, zero context pollution

Paper: https://zenodo.org/records/20433184
Code: https://github.com/gymaira1990-jpg/catnest/tree/main/00-%E8%AE%BA%E6%96%87%E5%90%88%E9%9B%86

I would love to hear your thoughts on feasibility of integrating this into DeepSeek inference.

Thanks!
GCat