Sparad:
| Huvudupphovsman: | |
|---|---|
| Materialtyp: | Recurso digital |
| Språk: | engelska |
| Publicerad: |
Zenodo
2026
|
| Ämnen: | |
| Länkar: | https://doi.org/10.5281/zenodo.20113783 |
| Taggar: |
Lägg till en tagg
Inga taggar, Lägg till första taggen!
|
Innehållsförteckning:
- <p>This paper presents an architecture where a lightweight outer learning layer uses a frozen large language model (LLM) as an environment for autonomous knowledge acquisition. The outer layer observes the LLM's outputs, extracts concepts and features, builds a knowledge graph, and generates questions driven by the graph's own structural gaps. Crucially, the outer layer forms its own abstractions and then queries the LLM about these self-generated abstractions, creating a cycle where learned structure drives new inquiry.</p> <p>A four-way ablation study over 200 steps shows that (1) feedback-driven questioning produces 4.9x denser knowledge graphs than structure-driven questioning without feedback, (2) every question strategy improves by 34-97% when feedback is enabled, and (3) structure alone performs worse than random exploration. An extended 2000-step run demonstrates that information gain accelerates rather than plateaus as the knowledge graph densifies. All code and data are publicly available.</p>