Dimension D4

Persistence

How long does the change last?

8 categories. Other dimensions: D1 D2 D3 D5 D6

II. Event-Driven Unplanned Modification

Category	Definition	Techniques
Ad-hoc Perm. Ad-hoc Permanent	An unplanned, event-driven parameter update triggered by sudden data drift, an emergent vulnerability, a targeted knowledge correction, or an arbitrary developer decision rather than a pre-specified schedule. Like scheduled permanent updates, the change produces a new model checkpoint that permanently supersedes the prior parameter state; unlike them, the timing, trigger, and scope of the intervention cannot be declared in advance and are determined reactively.	Training, FT (full), FT (partial), PEFT (LoRA, adapters), ReFT, DA, FSL, Retraining, SFT, RLHF (LLM Alignment), DPO, RLAIF/CAI, RLVR/GRPO, LP, LEACE, Adversarial Training, Meta-Learning, MTL, Self-Play, DP-FT, KD, Model Compression, Task Arith. & Model Merging, Long-Context Ext., MoE, Prompt Learning, Calibration, Cross-Modal Alignment, Modality-Spec. Adapters, Multimodal Instr. Tuning, Machine Unlearning, Knowledge Editing 32 technique(s)

I. Pre-Specifiable Baselines

Category	Definition	Techniques
Bound. Cumul. Bounded Cumulative	Incremental parameter accumulation across a finite, pre-enumerable set of new tasks or domains, in which each update extends the model's knowledge without overwriting prior representations. The finite, pre-specified character of the task sequence means the total knowledge accumulation horizon can, in principle, be declared in advance, making this the only cumulative persistence mode whose modification trajectory admits complete pre-specification.	CL, TIL 2 technique(s)
Sched. Perm. Scheduled Permanent	Pre-planned, periodic parameter updates executed on a fixed temporal or data-volume cadence, resulting in a permanently updated model checkpoint that fully supersedes the prior parameter state. The modification is permanent at the model-parameter level—the updated weights replace the previous checkpoint and persist across all subsequent inference sessions—and its triggering conditions and approximate timing can be declared in advance.	Training, FT (full), FT (partial), PEFT (LoRA, adapters), ReFT, DA, FSL, Retraining, SFT, RLHF (LLM Alignment), DPO, RLAIF/CAI, RLVR/GRPO, LP, Adversarial Training, Meta-Learning, MTL, Self-Play, SSL / CPT, DP-FT, FL, KD, Model Compression, Task Arith. & Model Merging, Long-Context Ext., MoE, Prompt Learning, Calibration, Cross-Modal Alignment, Modality-Spec. Adapters, Multimodal Instr. Tuning 31 technique(s)
Ver.-Persist. Version-Persistent	A context-only modification at the model-parameter level—such as a maintained retrieval corpus, a persistent steering-vector store, or a versioned production system prompt—that does not alter model weights but is maintained as a version-controlled, persistent system artifact whose update constitutes a system-level change event. While the model's internal parameter state is unchanged, the system's behavioral profile is durably modified by the maintained artifact across inference sessions.	PE, APO, RAG, CE 4 technique(s)

V. Inherited Burden

Category	Definition	Techniques
Pipe.-Dep. Pipeline-Dependent	A persistence profile in which the temporal durability of the adaptation is not intrinsic to the technique itself but is entirely delegated to and inherited from the downstream learning method that the technique augments. Because the technique operates on the training data pipeline rather than directly on model parameters, its effective change-control classification is entirely contingent on the downstream training method it feeds; no independent persistence characterization applies to the pipeline technique in isolation.	Curriculum Learning, Active Learning, Data Augmentation, Semi-SL 4 technique(s)

IV. Session-Scoped

Category	Definition	Techniques
Sess.-Eph. Session-Ephemeral	A behavioral modification achieved entirely through input-context conditioning or internal generation processes at inference time, with the model's parameter state remaining identical before and after each session. The adaptation is local to the active inference request or session and clears completely upon session termination, leaving no trace in the model's stored parameter state. While ephemeral at the parameter level, the system-level behavioral effect may be functionally persistent if the conditioning mechanism is applied uniformly across all deployed inference sessions.	PE, ICL, Test-Time Compute Scaling 3 technique(s)
Transient Transient	A modification strictly confined to a temporary inference session that does not persist once that session concludes. The transient category encompasses two mechanistically distinct implementations: parameter-modifying variants (e.g., test-time adaptation), in which gradient updates are applied during inference and then explicitly rolled back upon session completion via weight restoration; and activation-modifying variants (e.g., activation steering), in which no weight write occurs and the effect naturally clears at session end.	TTA, Activation Steering 2 technique(s)

III. Continuous Surveillance Required

Category	Definition	Techniques
Unbound. Cumul. Unbounded Cumulative	Open-ended, continuous parameter accumulation over an indefinite or unknown horizon of streaming data, in which the model's behavioral state is never stable but continuously incorporates new representational commitments. The absence of a pre-specifiable termination boundary is the defining property: unlike bounded cumulative persistence, the total scope of adaptation cannot be declared in advance, and the model's behavioral state at any future checkpoint is structurally indeterminate at deployment time.	CL, DIL, SSL / CPT 3 technique(s)

Category definitions verbatim from the paper's appendix