-д хадгалсан:
| Үндсэн зохиолч: | |
|---|---|
| Формат: | Recurso digital |
| Хэл сонгох: | англи |
| Хэвлэсэн: |
Zenodo
2025
|
| Нөхцлүүд: | |
| Онлайн хандалт: | https://doi.org/10.5281/zenodo.17105709 |
| Шошгууд: |
Шошго нэмэх
Шошго байхгүй, Энэхүү баримтыг шошголох эхний хүн болох!
|
Агуулга:
- <p>Scaling has produced surprising “emergent” behaviors in modern ML systems, <em>yet the mechanisms behind robust emergence remain unclear.</em> <strong>This paper argues that durable emergence is not a mystery of scale but a consequence of invariant-preserving feedback loops.</strong> When self-modifying agents update in ways that maintain internal stability while expanding representational reach, new behaviors crystallize as robust attractors; when loops erode invariants, apparent gains collapse into drift and brittleness.</p> <p>The paper formalizes a stability functional S(M) that gates self-improvement (ΔS(M) > 0), outlines practical proxies for invariant preservation (entailment, paraphrase stability, tool pre/post-conditions), and proposes falsifiable protocols for testing the framework. Empirical footholds from ARC-AGI, AlphaGeometry, and large proof libraries (Coq, Lean, Isabelle) suggest that <em>systems enforcing invariants already outperform pure stochastic scaling on reasoning-heavy tasks.</em></p> <p><strong>The arguement is that invariants unify capability and safety: </strong>the same substrate that yields robust emergence also prevents drift. The implication is a reframing of the bottleneck: not FLOPs, but invariants (the universal substrate of adaptive stability) quantified via an Invariant Data-Processing Inequality and a No-Free-Stability bound on verification work.</p>