Dimension D1

Mechanism

What changes?

9 categories. Other dimensions: D2 D3 D4 D5 D6

III. Inference-Time / Zero-Footprint

Category	Definition	Techniques
Act.-Space Manip. Activation-Space Manipulation	A mechanism that adds learned steering vectors or direction-specifying perturbations to the intermediate activation representations of a model during the forward pass—typically targeting specific residual stream positions or attention head outputs—thereby redirecting the model's behavioral trajectory along a target direction without modifying any stored weight tensors. Because the intervention operates in activation space rather than parameter space, the model's checkpoint is entirely unaffected; the effect is confined to the forward pass and leaves no trace in the persistent parameter state.	Activation Steering 1 technique(s)
Context Inj. Context Injection	An inference-time conditioning mechanism that prepends, appends, or structurally assembles task-relevant information—including natural-language instructions, retrieved documents, few-shot demonstrations, or structured tool outputs—into the model's input context window without modifying any stored parameter weights. Behavioral changes are achieved entirely through the model's existing attention and generation mechanisms acting on the enriched input; no gradient computation occurs, and the model's parameter state is identical before and after the inference session.	PE, APO, RAG, ICL, CE 5 technique(s)
Infer.-Time Search Inference-Time Search	An inference-time compute-allocation mechanism that scales response quality by executing multiple internal generation processes without modifying model parameters, injecting external retrieval context, or steering activations. The mechanism trades additional inference compute for improved output quality by exploring a broader portion of the model's generative distribution and selecting the highest-quality candidate according to a process-internal scoring criterion. Compute is consumed at inference time rather than training time.	Test-Time Compute Scaling 1 technique(s)

II. Gradient-Free / Weight Manipulation

Category	Definition	Techniques
Arch. Architectural Modification	A topological alteration of the model's computational graph that introduces new nodes, routing pathways, or expert sub-networks—such as dynamically initializing MoE routing modules, inserting gating components, or appending new expert layers—without modifying the pre-trained weights of existing base components. The mechanism expands the model's capacity or specialization range through structural graph surgery rather than parameter optimization on existing tensors.	Long-Context Ext., MoE 2 technique(s)
Param. Comp. Parameter Composition	A gradient-free algebraic operation applied directly to the parameter vectors of one or more pre-trained model checkpoints—such as task-vector addition, negation, or weight averaging—that derives a new behavioral configuration by combining or transforming weight-space representations without executing any training loop on task data. No forward passes over labeled examples and no loss function are required; the mechanism operates entirely in parameter space.	LEACE, Task Arith. & Model Merging, Machine Unlearning, Knowledge Editing 4 technique(s)
Param. Compres. Parametric Compression	A structural or numerical intervention that reduces a single existing model's parameter count, memory footprint, or arithmetic precision—through techniques such as magnitude pruning, structured sparsification, or weight quantization—producing a deployment-optimized derivative artifact. Unlike cross-model transfer, no independent teacher model is required; the compression operation modifies the model's representational capacity directly by eliminating or reducing its own parameter tensors.	Model Compression 1 technique(s)

I. Gradient-Based / Compute-Intensive

Category	Definition	Techniques
Cross-Model Transf. Cross-Model Transfer	A training procedure in which a new, independently deployable student model is optimized to replicate the output distributions, intermediate representations, or decision boundaries of a pre-existing teacher model. Rather than updating the teacher's parameters, the mechanism generates a derived artifact—the student—by minimizing a surrogate objective (e.g., KL divergence over soft logits, feature-map matching, or contrastive alignment) that transfers the teacher's generalization surface without requiring access to the original training data. The student and teacher are structurally distinct deployable artifacts.	KD 1 technique(s)
Param. Upd. Parametric Update	A gradient-based optimization procedure that computes loss-driven parameter gradients—using a task-specific or dataset-specific objective—and applies them to alter the stored weight tensors of one or more model components. The mechanism can operate across the full parameter space simultaneously (full FT), or be constrained to a designated subset of layers, adapter matrices, or task-specific heads (partial and parameter-efficient variants); in all cases it consumes forward and backward passes—and therefore training-time floating-point operations—to modify the model's learned representations. No architectural constraints are assumed: the defining property is that GD acts directly on stored parameters.	Training, FT (full), FT (partial), PEFT (LoRA, adapters), ReFT, DA, FSL, Retraining, CL, TIL, DIL, SFT, RLHF (LLM Alignment), DPO, RLAIF/CAI, RLVR/GRPO, LP, Adversarial Training, Meta-Learning, MTL, Self-Play, SSL / CPT, DP-FT, FL, Long-Context Ext., MoE, Prompt Learning, TTA, Calibration, Cross-Modal Alignment, Modality-Spec. Adapters, Multimodal Instr. Tuning, Machine Unlearning, Knowledge Editing 34 technique(s)

IV. Pipeline-Mediated

Category	Definition	Techniques
Pipe.-Med. Pipeline-Mediated	An indirect adaptation pathway in which the technique modifies the composition, ordering, labeling, or synthesis of training examples supplied to a downstream learning algorithm without directly modifying any model parameter. Because the technique operates on the data pipeline rather than on the model artifact itself, the persistence and structural scope of the resulting adaptation are entirely determined by and inherited from the downstream training method it feeds.	Curriculum Learning, Active Learning, Data Augmentation, Semi-SL 4 technique(s)

Category definitions verbatim from the paper's appendix