Dimension D5

Scope

How much of the model structure is modified?

10 categories. Other dimensions: D1 D2 D3 D4 D6

IV. No Parametric Footprint

Category	Definition	Techniques
Activ.-Space Activation-Space	Direct modulation of the model's intermediate forward-pass representations with zero modification to any underlying parameter weight tensors. While the model's internal activation state is transiently altered during the forward pass, no persistent change to any stored weight tensor occurs; the model's checkpoint is identical before and after the intervention.	Activation Steering 1 technique(s)
I/O Space Input/Output-Space	No modification is made to any model parameter or internal activation state; behavioral conditioning occurs exclusively through external input-space manipulation, output candidate selection, or inference-time search over the model's generative distribution. The model's computational graph and all stored weight tensors remain completely unaltered. The adaptation footprint is confined to the interface between the model and its deployment environment, not to any internal model state.	PE, APO, RAG, ICL, CE, Test-Time Compute Scaling 6 technique(s)

II. Targeted Component Validation

Category	Definition	Techniques
Compress. Artif. Compressed Artifact	A structurally reduced, pruned, or lower-precision derivative of the original base model. that occupies a strictly smaller representational capacity than the original while targeting preservation of the original model's functional behavior within an acceptable accuracy tolerance. The compressed artifact is evaluated as a modified derivative of the original model rather than an independently trained entity.	Model Compression 1 technique(s)
Part. Partial	Only a designated component or layer subset of the model is modified while the remainder of the parameter space is frozen. The bounded modification surface restricts behavioral change to the updated component subset, enabling targeted validation and localized regression testing of only the modified layers; unmodified layers require no re-evaluation.	FT (partial), PEFT (LoRA, adapters), DA, FSL, CL, TIL, DIL, SFT, RLHF (LLM Alignment), DPO, RLAIF/CAI, RLVR/GRPO, LEACE, Meta-Learning, SSL / CPT, DP-FT, Long-Context Ext., MoE, Prompt Learning, TTA, Cross-Modal Alignment, Multimodal Instr. Tuning, Machine Unlearning, Knowledge Editing 24 technique(s)
Surrog. Model Surrogate Model	A completely new, independently deployable model artifact trained to replicate a teacher model's outputs or to provide a learned scoring signal as an intermediate training artifact. Because it is a structurally new model produced as a byproduct of the adaptation procedure, it constitutes a distinct artifact from the adapted model itself and carries an independent representational provenance.	RLHF (LLM Alignment), KD 2 technique(s)

V. Distributed Aggregation

Category	Definition	Techniques
Dist. Distributed	The global model state is constituted by the aggregation of locally computed parameter updates from multiple decentralized clients, such that the adaptation signal is distributed across a federated system. The persistent artifact is the aggregated global model whose parameter state incorporates learning signals from all contributing local data shards without any raw data leaving the originating device. Validation focuses on the consistency of the aggregated global state rather than on any single local update.	FL 1 technique(s)

I. Full-System Validation

Category	Definition	Techniques
Fused Fused Composition	A derived model artifact produced by the algebraic combination of two or more parent model checkpoints, resulting in a merged weight state in which the individual contributing models are no longer trivially separable. The fused artifact's behavior is not directly attributable to any single parent model, making its representational lineage structurally opaque relative to the source checkpoints.	Task Arith. & Model Merging, Machine Unlearning 2 technique(s)
Whole Whole-Model	All layers and components of the model are modified simultaneously, producing the most comprehensive structural footprint in the taxonomy. Because the entire parameter space participates in the modification, no modular isolation boundary exists; any behavioral change introduced propagates across all task dimensions simultaneously, and the model's complete representational state is overwritten or jointly updated. Rollback requires full-checkpoint reversion rather than targeted component substitution.	Training, FT (full), Retraining, DIL, SFT, RLHF (LLM Alignment), DPO, RLAIF/CAI, RLVR/GRPO, Adversarial Training, MTL, Self-Play, SSL / CPT, DP-FT, TTA, Multimodal Instr. Tuning 16 technique(s)

III. Modular Isolation

Category	Definition	Techniques
Modular Modular	The adaptation is encapsulated as a decoupled, independently stored module that can be attached to or detached from a frozen base model without altering the base model's parameter state in any way. The base model's existing representational state is completely preserved, and the modular component constitutes an independent artifact whose versioning is entirely distinct from the base checkpoint. Rollback is achieved by removing or swapping the module without any base-model reversion.	PEFT (LoRA, adapters), ReFT, TIL, LP, MoE, Calibration, Modality-Spec. Adapters 7 technique(s)

VI. Inherited Burden

Category	Definition	Techniques
Pipe.-Dep. Pipeline-Dependent	The structural scope of the adaptation is not intrinsic to the technique itself but is entirely delegated to and inherited from the downstream training event that the technique feeds. Because the pipeline technique modifies the data distribution rather than the model's computational graph, it imposes no independent structural footprint; the component-level regression testing burden is entirely subsumed by the downstream training method that consumes the modified data.	Curriculum Learning, Active Learning, Data Augmentation, Semi-SL 4 technique(s)

Category definitions verbatim from the paper's appendix