保存先:
| 主要な著者: | , , |
|---|---|
| フォーマット: | Recurso digital |
| 言語: | |
| 出版事項: |
Zenodo
2026
|
| 主題: | |
| オンライン･アクセス: | https://doi.org/10.5281/zenodo.19928385 |
| タグ: |
タグ追加
タグなし, このレコードへの初めてのタグを付けませんか!
|
目次:
- We present a Two-Layer Architecture for continual learning identity preservation in small language models (SLMs), addressing both training-time weight forgetting and inference-time context loss within a unified theoretical framework: the Compression–State–Propagation (C-S-P) framework. At the training layer, we identify the Fisher Scale Problem: standard EWC silently fails in SLMs when Fisher Information diagonal values collapse to the 1e-4–1e-5 range, rendering the regularisation penalty numerically indistinguishable from zero. We introduce Fisher Scaling and GodelReplay (Fisher-scaled EWC-DR + experience replay), achieving 31.5% forgetting reduction over raw EWC on sequential tasks, 82.8% reduction on our curated Conflict Dataset (43× over standard EWC), and a 4.1% additive improvement over replay-alone at the empirically identified sweet spot of mem=200 across 10 PermutedMNIST tasks. At the inference layer, GodelAI-Lite provides persistent episodic memory (MemPalace-Lite), structured reasoning continuity (MACP-Lite), and identity drift governance (GIFP-Lite) to any frozen SLM. Evaluated on Gemma 4: +31.2% overall performance, 3/3 memory retention vs. 0/3 baseline. Zero fine-tuning. Portable JSON memory transfers across model boundaries. Both layers implement the same three C-S-P stages, validating a unified structural account. The T-score (gradient diversity diagnostic) and FLYWHEEL Self-Recursive Proof (54.6% identity preservation for the AI agents who built the system) are additional contributions.