Research Archive

Mechanisms before claims.

The archive tracks the public paper record. KALAVAI opens as a local research note because the site carries its figures, slides, PDF, and artifacts; the other entries currently resolve to arXiv.

2603.22755

KALAVAI

Predicting when independent specialist fusion works: a quantitative model for post-hoc cooperative LLM training.

Open note
2603.06728

Orion

Characterizing and programming Apple's Neural Engine for LLM training and inference, bypassing CoreML with a native runtime.

Read paper
2602.21231

ACAR

Adaptive Complexity Routing for multi-model ensembles with auditable decision traces and self-consistency variance routing.

Read paper
2605.11153

Evolutionary Mixture-of-LoRA

Decomposes the routing lever, lifecycle penalty, and substrate-conditional boundary in evolutionary adapter architectures.

Read paper