Live learning for language models: a self-judging, quarantined write path protected by live importance. Technical record of design and first results
A dated record of the design and first results of a live-learning system for language models, made before full papers are written. A claim from a live stream passes a surprise check, is judged by the model itself, waits in quarantine when uncertain, is kept as a note when nothing can confirm it, counts only independent sources, and is written into the weights with live importance protection (Press-fit: per-weight standing earned from the model's own right and wrong answers while writing, with a steep cube brake), after a simulated write checks refusals, maths, language and neighbouring facts; checked night folding merges what was learned. First results on small networks and on Qwen2.5 7B, 21B and 72B are reported, including the open problem of a lie repeated by many ordinary-looking outlets.
Authors
- Parag Garg
Institutions
- University of Cambridge (GB)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-18
- DOI
- https://doi.org/10.5281/zenodo.24355332
- Primary Topic
- Topic Modeling
- Type
- article
- Field-Weighted Citation Impact
- 0.00