I
Sparse boundary characterization.
Which parameters can move?
Research
மர்மம்
MARMAM
Heritable Sparse Mutability Masks. Parameter-efficient adaptation research, grounded in a substrate of published papers.
Why this exists
Most parameter-efficient fine-tuning treats the base model as a black box and bolts adaptation on top.
LoRA, adapters, and prefix tuning are useful, but MARMAM goes the other way: what if adaptation paths are heritable properties of the model's parameter topology itself?
Heritable Sparse Mutability Masks identify which parameters can change without destabilizing the model, treating adaptation as a sparse, structured search rather than a dense uniform perturbation.
Approach
II
Heritable mask propagation.
How do masks transfer across base models?
III
Conditional adaptation.
Adaptation under masks versus without.
Status
- Phase 1 boundary characterization active.
- Correction logged: Phase 1 used uniform full-LoRA perturbation; the HSMM mechanism has not been tested yet.
- Phase 2 with CH-LoRA conditionally authorized.
- Foundation of the broader UYIR research line.
Related publications
- UYIR - Evolutionary Mixture-of-LoRA Decomposition. Foundation of the MARMAM direction.
- KALAVAI - Cooperative Specialist Fusion via Frozen-Layer MoE Routing. Powers PINNAL's adapter fusion.
- ORION - Direct Paths into Local Neural Hardware. First open end-to-end system programming Apple's ANE.
- ACAR - Adaptive Complexity Routing with Auditable Traces. Foundation for the routing program.