Slow plasticity
Shared Core LoRA
Reusable cross-task low-rank memory for stable knowledge that should survive across tasks.
KCVANGUARD 2026 Final Project
NH-LoRA
A future-aware continual learning method that treats LoRA capacity as a structural decision, not a fixed adapter budget.
Rehearsal-Free CIL
Frozen ViT
Dynamic Rank
Sparse Routing
Why NH-LoRA
Adapter structure becomes part of the learning problem.
NH-LoRA is framed around the stability-plasticity pressure in rehearsal-free continual learning.
Continual learning without replay
Class-incremental systems must absorb new classes without revisiting old task data, making stability and plasticity hard to balance.
Adapter capacity is a moving target
A fixed low-rank budget can be too rigid when task novelty, conflict, and retention pressure change over time.
Structure becomes part of learning
NH-LoRA models when to reuse shared memory, expand task slots, open new capacity, and consolidate stable updates.
Architecture
Four moving parts, one structural control loop.
The method separates reusable memory, task-specific plasticity, per-instance routing, and future-aware planning.
Fast plasticity
Expandable Task Slot Bank
Task-specific slots that can grow when the planner sees novelty or conflict evidence.
Conditional routing
Instance Router
Sparse per-sample slot activation to reduce inter-slot interference at inference time.
Structural control
Horizon Planner
A future-aware controller that predicts novelty, conflict, rank budget, and consolidation signals.
Benchmark teaser
Reported logs are shaped into an explorer.
Numbers below come from the migrated project logs and remain presented as reported artifacts, not final paper claims.
Custom Dataset
2
Ready for final benchmark values.
Deployment reported top average
71.20%
Paired with reported forgetting metric.
Tracked benchmark groups
3
CIFAR-100, ImageNet-A, and Custom dataset for Deployment.
Experiment Explorer
Compare benchmark groups, config changes, and task-level metrics.
Interactive demo shell
The public demo is linked externally while this site explains the method, artifacts, and known limitations around it.
Bilingual Paper
Read the compiled paper PDFs with language-specific paper asset switching.
Resources
Direct paths into the project.
Jump into the demo, paper, experiment logs, or source repository.