Capacity as a decision
Adapter growth, reuse, and consolidation are modeled as explicit structural decisions.
Paper
Read the technical paper.
A technical paper view focused on the compiled PDFs, with bilingual paper asset switching and no LaTeX source reader.
Abstract preview
NH-LoRA combines a frozen Vision Transformer backbone with shared low-rank memory, expandable task-specific slots, instance-level routing, and a horizon planner that regulates structural capacity over time.
Shared and task-specific memory
The method separates reusable cross-task knowledge from fast task-specific plasticity.
Sparse instance routing
Inference activates a compact subset of relevant slots for each sample.
Future-aware planning
Task-state signals guide layer-wise novelty, conflict, rank, and retention behavior.
Viewer
Bilingual PDF reader.
Switching language changes only the paper asset. The website interface remains in English.
PDF viewer
Preparing the client PDF viewer...
Supplementary links
Continue through the research artifacts.
Internal
MethodOpen a related NH-LoRA research resource.Internal
ExperimentsOpen a related NH-LoRA research resource.External
RepositoryOpen a related NH-LoRA research resource.Paper PDF slots currently use duplicated placeholder assets until final bilingual PDFs are supplied.