{
"meta": {
"version": "1.0.0",
"released": "2026-08-06",
"n_techniques": 48,
"n_reference_rows": 1,
"centerpiece_note": "Abbreviations are listed in the manuscript before the bibliography and in Supplementary Section S1. Highly flexible techniques such as CE (e.g., prompting for step-by-step rationales) and RLHF (e.g., penalizing unexplainable refusals) can dynamically adopt Explainability as a primary goal when engineered specifically to satisfy regulatory transparency mandates.",
"crosscheck": {
"cells_compared": 294,
"cells_identical": 294,
"outstanding": 0
}
},
"techniques": [
{
"slug": "training",
"tech_key": "training",
"name": "Training",
"family": "Reference Baseline",
"is_reference_row": true,
"footnote_markers": [],
"footnotes": [],
"notes": [
"definition not found in Appendix C"
],
"d1": [
"parametric-update"
],
"d2": [
"build-capability"
],
"d3": [
"large-labeled"
],
"d4": [
"ad-hoc-permanent",
"scheduled-permanent"
],
"d5": [
"whole-model"
],
"d6": [
"dl",
"fm",
"llm",
"ml",
"mllm"
],
"definition_verbatim": "",
"source_ref": "",
"classification_tension": "",
"summary_editorial": "The original build of a model — learning its parameters from scratch on a large dataset — included here as a reference point rather than as a post-training technique."
},
{
"slug": "fullft",
"tech_key": "fullFT",
"name": "FT (full)",
"family": "Knowledge Transfer and Task Specialization",
"is_reference_row": false,
"footnote_markers": [],
"footnotes": [],
"notes": [],
"d1": [
"parametric-update"
],
"d2": [
"task-specialization"
],
"d3": [
"large-labeled"
],
"d4": [
"ad-hoc-permanent",
"scheduled-permanent"
],
"d5": [
"whole-model"
],
"d6": [
"dl",
"fm",
"llm",
"mllm"
],
"definition_verbatim": "Full FT is the maximally expressive parametric realization of TL: GD (Parametric Update) applies to every model layer without exception, producing a Whole-Model footprint that carries the highest structural validation burden in the taxonomy. The unrestricted update surface structurally necessitates the Large Labeled data regime—with all parameters free to drift, the optimization landscape has sufficient degrees of freedom to overfit or catastrophically corrupt prior knowledge unless constrained by large-scale supervision [ref]. The Task Specialization goal draws the sharpest coordinate boundary with Retraining: both are mechanistically identical across D1 (Mechanism), D4 (Persistence), and D5 (Scope), yet retraining targets Drift Remediation on a fixed behavioral envelope rather than reshaping that envelope for a new task. Compared to Partial FT, Full FT sacrifices the reduced testing surface (D5 (Scope) contracts from Whole-Model to Partial) in exchange for broader update expressiveness; compared to PEFT, it eliminates the Modular isolation boundary that permits component-level rollback, replacing it with whole-model entanglement. Because Full FT modifies the entire parameter space with no modular isolation boundary, rollback requires complete checkpoint restoration rather than targeted component substitution, and the validation obligation extends to full-system behavioral regression testing across all previously certified behavioral boundaries before any re-deployment.",
"source_ref": "Knowledge Transfer and Task Specialization",
"classification_tension": "",
"summary_editorial": "Continues training every weight in the model on new task data — maximum flexibility, maximum cost, and no part of the model left unchanged."
},
{
"slug": "partft",
"tech_key": "partFT",
"name": "FT (partial)",
"family": "Knowledge Transfer and Task Specialization",
"is_reference_row": false,
"footnote_markers": [],
"footnotes": [],
"notes": [],
"d1": [
"parametric-update"
],
"d2": [
"task-specialization"
],
"d3": [
"small-labeled"
],
"d4": [
"ad-hoc-permanent",
"scheduled-permanent"
],
"d5": [
"partial"
],
"d6": [
"dl",
"fm",
"llm",
"mllm"
],
"definition_verbatim": "Partial FT targets Task Specialization via the same Parametric Update mechanism as Full FT but constrains updates to later, task-specific layers while freezing the early layers that encode broadly transferable representations [ref]. This architectural discipline collapses the update surface from Whole-Model to Partial—a coordinate divergence from Full FT that carries both a data consequence and a validation consequence: the data requirement contracts to Small Labeled (the frozen layers perform feature extraction, leaving only the task-specific layers requiring supervision), and the regression testing footprint is structurally isolated to the modified layer subset, enabling targeted component-level rollback. Compared to PEFT, which introduces a fully decoupled Modular adapter that can be removed without touching any base-model layer, Partial FT embeds updates directly into the base model's later layers—rollback requires layer-level checkpoint restoration, not a clean module swap. The practical distinction matters for change-control documentation: a PEFT event generates a standalone adapter artifact; a Partial FT event generates a modified base-model checkpoint whose changed layers must be explicitly versioned. Partial FT necessitates versioned checkpointing of the modified layer subset and localized parametric regression testing bounded strictly to those layers; because the frozen early layers remain mathematically unchanged, they require no re-validation.",
"source_ref": "Knowledge Transfer and Task Specialization",
"classification_tension": "",
"summary_editorial": "Retrains only the later layers and freezes the early ones, so the general features learned during pre-training stay intact and less needs re-testing."
},
{
"slug": "peft",
"tech_key": "peft",
"name": "PEFT (LoRA, adapters)",
"family": "Knowledge Transfer and Task Specialization",
"is_reference_row": false,
"footnote_markers": [],
"footnotes": [],
"notes": [],
"d1": [
"parametric-update"
],
"d2": [
"computational-efficiency",
"task-specialization"
],
"d3": [
"small-labeled"
],
"d4": [
"ad-hoc-permanent",
"scheduled-permanent"
],
"d5": [
"modular",
"partial"
],
"d6": [
"dl",
"fm",
"llm",
"mllm"
],
"definition_verbatim": "PEFT is taxonomically the most distinctive member of the FT family, differentiated by its Modular, Partial scope, which no full or partial FT variant achieves. The defining operational property is that the base model's weights remain strictly frozen throughout training; all adaptation occurs within a small set of newly introduced parameters appended to or inserted into the frozen architecture [ref]. This produces the Modular coordinate: the adapter is a structurally independent artifact that can be loaded, swapped, or removed without touching the base checkpoint. This Modular property creates a structural bridge to Task Arithmetic and Model Merging, where adapter task-vectors can be algebraically combined. The data requirement contracts to Small Labeled relative to Full FT because the adapter's parameter count is orders of magnitude smaller than the base. At the LLM and MLLM tiers, PEFT simultaneously absorbs the distributional gap goal of DA when applied to domain-specific corpora, and its modular architecture makes it the preferred complement to alignment techniques (RLHF, DPO) in production stacks. Because a change-control event in a PEFT deployment affects only the adapter artifact—not the frozen base—regression testing scope is bounded to the adapter module alone; the base model's existing validation record remains unmodified, enabling clean rollback by removing the adapter checkpoint without any base-model reversion. While categorized as a single structural decision unit in the taxonomy, the PEFT paradigm encompasses a wide array of architectural variants. These include quantization-aware methods (e.g., QLoRA [ref]), continuous prompt interventions (prefix-tuning, prompt-tuning, p-tuning), sparse bias updates (BitFit [ref]), and attention-scaling (IA3 [ref]). Furthermore, dynamic adapter composition strategies, such as adapter fusion and LoRA routing (e.g., X-LoRA [ref], LoraHub [ref], S-LoRA [ref]), functionally blur the boundary between modular PEFT and MoE by dynamically routing inputs to specialized adapter modules.",
"source_ref": "Knowledge Transfer and Task Specialization",
"classification_tension": "",
"summary_editorial": "Freezes the base model and trains a small add-on instead, so the original weights are never touched and the adapter can be swapped out or removed cleanly."
},
{
"slug": "reft",
"tech_key": "reft",
"name": "ReFT",
"family": "Knowledge Transfer and Task Specialization",
"is_reference_row": false,
"footnote_markers": [],
"footnotes": [],
"notes": [],
"d1": [
"parametric-update"
],
"d2": [
"task-specialization"
],
"d3": [
"small-labeled"
],
"d4": [
"ad-hoc-permanent",
"scheduled-permanent"
],
"d5": [
"modular"
],
"d6": [
"llm",
"mllm"
],
"definition_verbatim": "ReFT presents an alternative to standard weight-based PEFT by intervening on the internal activation streams of the model [ref]. While it utilizes a Parametric Update, it does not alter the base weights; instead, it trains lightweight, Modular projection matrices (interventions) that alter the model's hidden representations during the forward pass. It targets Task Specialization using Small Labeled datasets. Because it introduces permanent, swappable components, it maintains Ad-hoc Permanent, Scheduled Permanent persistence. It is distinct from Activation Steering (which calculates transient vectors) because ReFT permanently trains its intervention matrices via GD.",
"source_ref": "Knowledge Transfer and Task Specialization",
"classification_tension": "",
"summary_editorial": "Trains small permanent modules that adjust the model's internal activations as it runs, leaving the base weights themselves unchanged."
},
{
"slug": "da",
"tech_key": "da",
"name": "DA",
"family": "Knowledge Transfer and Task Specialization",
"is_reference_row": false,
"footnote_markers": [
"\\S"
],
"footnotes": [
{
"marker": "\\S",
"text": "Classification tensions involving these techniques are discussed in Supplementary Section S4."
}
],
"notes": [],
"d1": [
"parametric-update"
],
"d2": [
"distributional-gap-bridging"
],
"d3": [
"small-labeled",
"unlabeled"
],
"d4": [
"ad-hoc-permanent",
"scheduled-permanent"
],
"d5": [
"partial"
],
"d6": [
"dl",
"fm"
],
"definition_verbatim": "DA addresses Distributional Gap Bridging—reducing the feature-space mismatch between source and target domains when labeled target data is unavailable or scarce. Its Small Labeled, Unlabeled profile reflects the defining operational condition: the target domain lacks sufficient labeled examples to support task-specific FT, The Parametric Update mechanism operates by training a feature extractor alongside an explicit domain discriminator or statistical divergence minimizer, yielding a Partial footprint (the alignment layers and feature extractor are updated; task heads may remain frozen) [ref]. The defining taxonomic boundary is DL, FM: at the LLM tier, the DA goal persists but the mechanism is superseded—CPT achieves distributional alignment by continued self-supervised training on unlabeled target-domain corpora, and domain-specific PEFT injects target-domain signal via lightweight adapters. System documentation for LLM-level DA must therefore specify the concrete CPT or PEFT pathway rather than invoking classical DA terminology. At the DL, FM tier, change-control documentation should specify which alignment layers were updated and which domain-discriminator architecture was used, as these define the precise structural boundary of the modification event; regression testing should verify that source-domain performance is preserved alongside target-domain improvement—necessitating a dual-distribution evaluation protocol as part of the validation record.",
"source_ref": "Knowledge Transfer and Task Specialization",
"classification_tension": "DA. Stops at DL, FM: the goal of bridging domain gaps persists at the LLM level, but the mechanism is superseded by CPT and domain-specific PEFT. PE vs.\\ ICL ( vs. ). While structurally identical on their underlying mechanism (D1), scope (D5), and model tier (D6), and partially overlapping in their goals (D2) and persistence (D4), the shift in data requirements (Zero-Shot vs. Few Demonstrations) causes their profiles to diverge. ICL uniquely introduces Task Specialization, while PE distinctly extends to Behavior Control and Version-Persistent implementations.",
"summary_editorial": "Trains a model to close the gap when the data seen at deployment looks different from the training data, without needing many labels for the new domain."
},
{
"slug": "fsl",
"tech_key": "fsl",
"name": "FSL",
"family": "Knowledge Transfer and Task Specialization",
"is_reference_row": false,
"footnote_markers": [
"\\P"
],
"footnotes": [
{
"marker": "\\P",
"text": "For in-context FSL at the LLM level, see ICL."
}
],
"notes": [],
"d1": [
"parametric-update"
],
"d2": [
"task-specialization"
],
"d3": [
"few-demonstrations",
"small-labeled"
],
"d4": [
"ad-hoc-permanent",
"scheduled-permanent"
],
"d5": [
"partial"
],
"d6": [
"dl",
"fm"
],
"definition_verbatim": "FSL is defined within this taxonomy exclusively as a Parametric Update intervention under extreme data scarcity. Depending on the implementation, it relies on either 1–5 examples for metric-based support sets (Few Demonstrations) or tens of examples for gradient-based updates (Small Labeled), producing a persistent weight modification (Ad-hoc Permanent, Scheduled Permanent) [ref]. This strict boundary resolves the taxonomy's most consequential terminological ambiguity: colloquial ``few-shot'' usage conflates two mechanistically and operationally distinct operations. The parametric form produces a versioned model artifact that has undergone a parameter-modifying training event and therefore demands formal change-control tracking and baseline regression testing. The non-parametric, prompting-based form shares Few Demonstrations but operates via Context Injection, producing only a Session-Ephemeral behavioral shift that leaves the underlying model artifact unmodified; this is reassigned to ICL. The D4 (Persistence) divergence is the taxonomic anchor: FSL creates a new versioned artifact requiring software change-control; ICL creates no artifact whatsoever. FSL's DL, FM ceiling mirrors DA's: at the LLM tier, ICL achieves rapid few-shot task specialization through context conditioning alone, making formal FSL training pipelines architecturally redundant for pure few-shot tasks at that scale. FSL necessitates version-control tracking of the resulting weight checkpoint; the extremely small Few Demonstrations dataset constitutes the sole training provenance record and should be fully documented.",
"source_ref": "Knowledge Transfer and Task Specialization",
"classification_tension": "FSL. In this framework, the FSL categorization is strictly reserved for parametric approaches—such as gradient-based meta-learning or metric-based networks—where the model's weights undergo a Ad-hoc Permanent, Scheduled Permanent update based on limited labeled data. While the term FSL is colloquially used to describe providing examples in an LLM prompt, this framework classifies that non-parametric, Session-Ephemeral mechanism distinctly as ICL. ICL functions primarily to induce a Knowledge Update, Task Specialization mapping during inference, bypassing parameter updates entirely.",
"summary_editorial": "Learns a new task from a handful of labelled examples by actually updating weights — as opposed to just showing examples in the prompt, which is ICL."
},
{
"slug": "retraining",
"tech_key": "retraining",
"name": "Retraining",
"family": "Temporal Adaptation and Maintenance",
"is_reference_row": false,
"footnote_markers": [],
"footnotes": [],
"notes": [],
"d1": [
"parametric-update"
],
"d2": [
"drift-remediation"
],
"d3": [
"sequential-incremental"
],
"d4": [
"ad-hoc-permanent",
"scheduled-permanent"
],
"d5": [
"whole-model"
],
"d6": [
"dl",
"fm",
"llm",
"ml",
"mllm"
],
"definition_verbatim": "Retraining updates (Parametric Update) a deployed model on newer or additional data to restore performance degraded by distribution shift, representing the foundational lifecycle maintenance operation across all model types. Its universality across DL, FM, LLM, ML, MLLM reflects reflects the fact that distributional drift is an architecture-agnostic failure mode: any model whose training distribution diverges from its deployment distribution will degrade, regardless of architectural family [ref]. The critical D2 (Goal) divergence from (Full/Partial) FT is precise: both operate identically on D1 (Mechanism) and D4 (Persistence), yet retraining preserves the model's original task objective (Drift Remediation) rather than reshaping its capability boundaries for a new task (Task Specialization). This D2 (Goal) distinction carries a direct validation consequence—a retraining event that strictly adheres to the original task specification does not require an intended-use review; an FT event does. The Sequential / Incremental coordinate reflects that retraining data arrives in temporally ordered batches whose distributional properties evolve across the model's operational lifecycle—this is structurally distinct from the static, curated datasets of (Full/Partial) FT (Large Labeled, Small Labeled) or FSL (; Few Demonstrations, Small Labeled). The Ad-hoc Permanent, Scheduled Permanent coordinate marks the boundary with CL: unconstrained retraining overwrites prior weight states with a new fixed checkpoint, whereas CL's Bounded Cumulative, Unbounded Cumulative encodes a commitment to accumulation rather than replacement. The Whole-Model scope means that every weight is overwritten at each retraining event; there is no modular isolation boundary, so any behavioral drift introduced by the new training data propagates across all task dimensions simultaneously—necessitating automated full-system regression testing against a frozen behavioral baseline before re-deployment, with the historical checkpoint retained as the rollback artifact.",
"source_ref": "Temporal Adaptation and Maintenance",
"classification_tension": "",
"summary_editorial": "Trains the existing model again on newer data to recover accuracy lost as the world drifted away from the original training distribution."
},
{
"slug": "cl",
"tech_key": "cl",
"name": "CL",
"family": "Temporal Adaptation and Maintenance",
"is_reference_row": false,
"footnote_markers": [],
"footnotes": [],
"notes": [],
"d1": [
"parametric-update"
],
"d2": [
"continual-adaptation"
],
"d3": [
"sequential-incremental"
],
"d4": [
"bounded-cumulative",
"unbounded-cumulative"
],
"d5": [
"partial"
],
"d6": [
"dl",
"fm",
"llm",
"mllm"
],
"definition_verbatim": "CL addresses the stability-plasticity trade-off [ref]: how to accumulate knowledge across sequential tasks or shifting distributions without overwriting previously learned representations. Its defining coordinate is D4 (Persistence): Bounded Cumulative, Unbounded Cumulative—a deliberate architectural commitment to knowledge accumulation rather than replacement—which is the main axis that separates CL from Retraining (, Ad-hoc Permanent, Scheduled Permanent) despite their identical D1 (Mechanism) and D3 (Data Requirements) profiles. The bounded variant corresponds to sequential acquisition over a finite, pre-enumerable task set (TIL, ); the unbounded variant corresponds to open-ended streaming adaptation over an indefinite sequence of distributions (DIL, ). This cumulative persistence coordinate is operationally consequential: the open-ended modification trajectory of Unbounded Cumulative fundamentally breaks the assumptions of discrete, periodic regression testing, because the model's behavioral state is never stable—it is continuously accumulating new representational commitments. At the LLM tier, CL encompasses three sequential stages—continued pre-training, domain-adaptive pre-training, and continual FT [ref]—each carrying distinct D3 (Data Requirements) and D5 (Scope) profiles. The D5 (Scope) coordinate (Partial) reflects that production CL deployments at scale almost always apply constrained, parameter-isolated updates (adapters, selective layer freezing) rather than whole-model retraining, precisely to limit the forgetting surface. CL's cumulative persistence necessitates continuous behavioral drift monitoring rather than periodic snapshot regression testing; each task or distribution boundary must be logged as a discrete lifecycle event with its own targeted stability evaluation verifying that prior-task performance has not regressed beyond a predefined tolerance.",
"source_ref": "Temporal Adaptation and Maintenance",
"classification_tension": "",
"summary_editorial": "Learns a sequence of tasks over time while trying not to forget earlier ones — the central difficulty being catastrophic forgetting."
},
{
"slug": "til",
"tech_key": "til",
"name": "TIL",
"family": "Temporal Adaptation and Maintenance",
"is_reference_row": false,
"footnote_markers": [],
"footnotes": [],
"notes": [],
"d1": [
"parametric-update"
],
"d2": [
"continual-adaptation"
],
"d3": [
"sequential-incremental"
],
"d4": [
"bounded-cumulative"
],
"d5": [
"modular",
"partial"
],
"d6": [
"dl",
"fm"
],
"definition_verbatim": "TIL is a specific operational variant of CL in which a model acquires a finite, pre-enumerable sequence of distinct tasks—each presented without access to prior task data—with task identity explicitly provided at inference time [ref]. The Bounded Cumulative coordinate is the main definitional anchor: the finite, pre-specified task sequence means the total knowledge accumulation horizon can, in principle, be declared in advance, making TIL the only CL variant whose change trajectory is pre-specifiable in a technical plan. Its Modular, Partial scope reflects that the dominant practical implementations at the DL and FM tiers use parameter-isolated architectures—dedicated output heads, task-specific adapters, or masked parameter subsets—that physically separate each task's learned representation from prior ones, directly mitigating the forgetting problem through structural disentanglement rather than regularization. The DL, FM ceiling is imposed by architectural supersession: at the LLM tier, domain-adaptive pre-training and continual PEFT subsume TIL's goals under the broader CL framework, rendering dedicated TIL protocols redundant. Compared to DIL, TIL's inference-time task identifier is a system-level dependency that must be tracked in deployment documentation—if the task-identity signal is unavailable or corrupted, the deployed system's behavior is undefined, constituting a unique failure mode that requires explicit specification in the operational validation plan. TIL's bounded, pre-specifiable task sequence supports formal pre-deployment documentation of the complete capability accumulation trajectory; the modular scope permits per-task regression testing at each task boundary, isolating validation to the newly added task module without re-validating prior task components.",
"source_ref": "Temporal Adaptation and Maintenance",
"classification_tension": "",
"summary_editorial": "Continual learning where you know which task an input belongs to, so the model can route to task-specific parameters."
},
{
"slug": "dil",
"tech_key": "dil",
"name": "DIL",
"family": "Temporal Adaptation and Maintenance",
"is_reference_row": false,
"footnote_markers": [],
"footnotes": [],
"notes": [],
"d1": [
"parametric-update"
],
"d2": [
"continual-adaptation",
"drift-remediation"
],
"d3": [
"sequential-incremental"
],
"d4": [
"unbounded-cumulative"
],
"d5": [
"partial",
"whole-model"
],
"d6": [
"dl",
"fm"
],
"definition_verbatim": "DIL is the open-ended (Unbounded Cumulative) variant of CL: the model encounters a potentially indefinite sequence of domains or class distributions without access to prior training data and—critically—without a task identifier at inference time [ref]. This absence of inference-time task identity is the defining coordinate divergence from TIL. The Partial, Whole-Model scope reflects that some DIL approaches retrain the entire network while others selectively update domain-adaptive components. The compound D2 (Goal) profile—Continual Adaptation, Drift Remediation—captures DIL's dual operational goal: it simultaneously accumulates new distributional knowledge (Continual Adaptation) and preserves performance on prior distributions that may re-emerge (Drift Remediation). The DL, FM ceiling is imposed by the same supersession mechanism as TIL: at the LLM tier, continued pre-training and domain-adaptive PEFT subsume DIL's goals. DIL's unbounded modification trajectory is the most operationally demanding persistence profile in the taxonomy: the system's behavioral state cannot be fully characterized at any fixed checkpoint, necessitating continuous runtime monitoring with per-distribution performance tracking and explicit drift-impact assessments at each domain boundary transition.",
"source_ref": "Temporal Adaptation and Maintenance",
"classification_tension": "",
"summary_editorial": "Continual learning where the task stays the same but the input distribution keeps shifting, so the model must track a moving target."
},
{
"slug": "sft",
"tech_key": "sft",
"name": "SFT",
"family": "Alignment, Reasoning, and Trustworthiness",
"is_reference_row": false,
"footnote_markers": [],
"footnotes": [],
"notes": [],
"d1": [
"parametric-update"
],
"d2": [
"alignment",
"safety",
"task-specialization"
],
"d3": [
"instruction-response-corpus"
],
"d4": [
"ad-hoc-permanent",
"scheduled-permanent"
],
"d5": [
"partial",
"whole-model"
],
"d6": [
"llm",
"mllm"
],
"definition_verbatim": "SFT, also termed instruction tuning, establishes the foundational behavioral capability for an LLM to interact via natural language instructions [ref], and its taxonomic position demands precise separation from classical task-specific FT. Both employ Parametric Update via supervised GD, but the D3 (Data Requirements) divergence is categorical: task FT operates on Small Labeled datasets of narrowly scoped, single-function examples yielding a bounded operational envelope, whereas SFT employs a massively heterogeneous Instruction-Response Corpus—spanning question answering, summarization, code generation, reasoning chains, and multilingual instructions—to instill a general instruction-following capability that cuts across all task boundaries. This D3 (Data Requirements) divergence propagates into the D2 (Goal) profile: SFT achieves a multi-domain form of Task Specialization alongside baseline Alignment, Safety, rather than the single-function specialization of task FT. While traditional SFT relies on human-curated datasets, contemporary variants frequently utilize synthetic, model-generated data [ref] (e.g., Instruction Backtranslation [ref], RAFT [ref], and Self-Rewarding models [ref]). This introduces a critical D3 (Data Requirements) nuance: it shifts the compliance and bias-auditing burden away from human annotator governance and onto the upstream teacher model generating the synthetic corpus. Safety SFT is a specific instantiation that explicitly enforces Safety by restricting the corpus to refusal demonstrations and harm-category filtered examples. The D5 (Scope) coordinate—Partial, Whole-Model—is implementation-dependent: SFT applied to all parameters yields Whole-Model with a full-system regression burden. Because SFT reshapes the model's behavioral envelope across all tasks simultaneously—not just a single bounded function—it demands comprehensive cross-domain behavioral regression testing before re-deployment; a task-specific regression suite is insufficient; the validation protocol must cover the full instruction-following surface, including all task types represented in the instruction-response corpus.",
"source_ref": "Alignment, Reasoning, and Trustworthiness",
"classification_tension": "",
"summary_editorial": "Trains a model on curated instruction-and-response pairs so it answers in the format and manner people expect."
},
{
"slug": "rlhf",
"tech_key": "rlhf",
"name": "RLHF (LLM Alignment)",
"family": "Alignment, Reasoning, and Trustworthiness",
"is_reference_row": false,
"footnote_markers": [
"*"
],
"footnotes": [
{
"marker": "*",
"text": "While preference-based RL predates foundation models, this row explicitly scopes its modern instantiation for generative alignment at the lm tiers. flushleft"
},
{
"marker": "Note",
"text": "Highly flexible techniques such as CE (e.g., prompting for step-by-step rationales) and RLHF (e.g., penalizing unexplainable refusals) can dynamically adopt Explainability as a primary goal when engineered specifically to satisfy regulatory transparency mandates."
}
],
"notes": [],
"d1": [
"parametric-update"
],
"d2": [
"alignment",
"safety"
],
"d3": [
"preference-pairs"
],
"d4": [
"ad-hoc-permanent",
"scheduled-permanent"
],
"d5": [
"partial",
"surrogate-model",
"whole-model"
],
"d6": [
"llm",
"mllm"
],
"definition_verbatim": "RLHF. RLHF trains a reward model on human preference pairs and optimizes an LLM policy against it, constituting the canonical alignment technique for generative models [ref]. The Preference Pairs coordinate is the mechanism's defining input requirement and its most consequential data-governance obligation: unlike Small Labeled datasets annotated for a fixed objective, preference pairs encode comparative behavioral judgments—which of two model outputs a human annotator prefers—making annotator demographics, selection criteria, and inter-rater reliability first-class artifacts in the training provenance record. The Alignment, Safety goal and the strict LLM, MLLM bound reflect that preference-based alignment is architecturally contingent on the model's capacity for open-ended natural-language generation; the preference signal is structurally meaningless for discriminative or shallow architectures. The D5 (Scope) coordinate is implementation-dependent in a way that carries direct validation consequences: applying RL policy optimization across all network parameters yields Whole-Model, demanding full-system regression testing; applying it via PEFT adapters restricts the scope to Partial, isolating the change-control footprint to the adapter; this structural distinction is operationally critical, as it determines whether an alignment modification event demands full-system regression testing or a targeted, component-level validation assessment. Training the reward model introduces a third Surrogate Model component—a fully independent model artifact that triggers its own conformity validation chain. This last point is the taxonomically critical divergence from DPO: both share D1 (Mechanism), D2 (Goal), D3 (Data Requirements), and D4 (Persistence), but RLHF's reward model produces a Surrogate Model footprint that DPO eliminates entirely by recasting alignment as a direct classification objective. RLHF deployments require independent validation of the reward model artifact as a separate system component; any update to the reward model constitutes a distinct change-control event from the policy update, as the reward model defines the optimization target for all subsequent policy training.",
"source_ref": "Alignment, Reasoning, and Trustworthiness",
"classification_tension": "",
"summary_editorial": "Uses human preference comparisons to train a reward model, then tunes the model against it — the standard route to making outputs helpful and safe."
},
{
"slug": "dpo",
"tech_key": "dpo",
"name": "DPO",
"family": "Alignment, Reasoning, and Trustworthiness",
"is_reference_row": false,
"footnote_markers": [],
"footnotes": [],
"notes": [],
"d1": [
"parametric-update"
],
"d2": [
"alignment",
"safety"
],
"d3": [
"preference-pairs"
],
"d4": [
"ad-hoc-permanent",
"scheduled-permanent"
],
"d5": [
"partial",
"whole-model"
],
"d6": [
"llm",
"mllm"
],
"definition_verbatim": "DPO eliminates the explicit reward model required by RLHF, recasting alignment as a supervised classification objective directly on preference pairs [ref]. DPO shares D1 (Mechanism), D2 (Goal), D3 (Data Requirements), and D4 (Persistence) with RLHF—making the D5 (Scope) divergence the sole but operationally decisive difference between them. DPO's elimination of the reward model collapses D5 (Scope) to Partial, Whole-Model alone, removing the Surrogate Model component entirely. This D5 (Scope) reduction is the most system-engineering-consequential property of DPO: it eliminates the need for a separate reward model validation pipeline, and the three-stage orchestration infrastructure. Compared to RLAIF/CAI, DPO shares the same D5 (Scope) simplification but retains the Preference Pairs human-annotation requirement; RLAIF substitutes AI-generated constitutional preferences (Constitution-Derived), shifting the validation burden from annotator oversight to constitutional text auditing. DPO's structural simplicity narrows the validation surface to a single policy artifact; however, the preference pair dataset remains the sole provenance record determining the alignment boundary, and its annotator-selection protocol, demographic coverage, and inter-rater reliability metrics must be fully documented as the system's behavioral specification. Recent advancements have expanded this family into broader preference-optimization variants, including IPO [ref], KTO [ref], and SimPO [ref]. Notably, ORPO [ref] collapses the traditional multi-stage pipeline by integrating baseline instruction tuning and alignment into a single gradient update.",
"source_ref": "Alignment, Reasoning, and Trustworthiness",
"classification_tension": "",
"summary_editorial": "Reaches the same goal as RLHF directly from preference pairs, skipping the separate reward model and the reinforcement-learning loop."
},
{
"slug": "rlaif",
"tech_key": "rlaif",
"name": "RLAIF/CAI",
"family": "Alignment, Reasoning, and Trustworthiness",
"is_reference_row": false,
"footnote_markers": [],
"footnotes": [],
"notes": [],
"d1": [
"parametric-update"
],
"d2": [
"alignment",
"safety"
],
"d3": [
"constitution-derived"
],
"d4": [
"ad-hoc-permanent",
"scheduled-permanent"
],
"d5": [
"partial",
"whole-model"
],
"d6": [
"llm",
"mllm"
],
"definition_verbatim": "RLAIF/CAI share the optimization infrastructure of RLHF but fundamentally substitutes the D3 (Data Requirements) profile [ref]: human-annotated preference pairs are replaced by synthetic preferences generated by an AI model operating under a set of constitutional principles (Constitution-Derived) [ref]. This D3 (Data Requirements) substitution is the technique's defining taxonomic property and its consequential system-engineering decision: it transfers the entire source of the alignment signal from human annotators—whose demographics, biases, and consistency are tracked through inter-rater reliability metrics—to a text-based constitutional document and an AI feedback model whose behavioral properties, latent biases, and edge-case failure modes must instead be audited. Compared to RLHF and DPO, RLAIF/CAI shares D1 (Mechanism), D2 (Goal), D4 (Persistence), D5 (Scope), and D6 (Model Type) but diverges exclusively on D3 (Data Requirements)—making the data-source substitution the main structural change. The scalability advantage is architectural: AI annotation parallelizes without human bottlenecks, enabling preference dataset generation at orders-of-magnitude larger scale than human annotation pipelines permit. Because the constitutional text and the AI feedback model jointly constitute the alignment specification, both must be treated as version-controlled engineering artifacts; any modification to either triggers a new change-control event, and the validation protocol must include semantic-stability testing of the constitutional principles across edge-case behavioral scenarios and regression testing of the AI critic pipeline's preference consistency.",
"source_ref": "Alignment, Reasoning, and Trustworthiness",
"classification_tension": "",
"summary_editorial": "Replaces human preference labels with AI-generated ones, guided by a written set of principles rather than case-by-case human judgement."
},
{
"slug": "rlvr",
"tech_key": "rlvr",
"name": "RLVR/GRPO",
"family": "Alignment, Reasoning, and Trustworthiness",
"is_reference_row": false,
"footnote_markers": [
"\\S"
],
"footnotes": [
{
"marker": "\\S",
"text": "Classification tensions involving these techniques are discussed in Supplementary Section S4."
}
],
"notes": [],
"d1": [
"parametric-update"
],
"d2": [
"reasoning"
],
"d3": [
"verifiable-task"
],
"d4": [
"ad-hoc-permanent",
"scheduled-permanent"
],
"d5": [
"partial",
"whole-model"
],
"d6": [
"llm",
"mllm"
],
"definition_verbatim": "RLVR/GRPO targets Reasoning—multi-step problem-solving, planning, and chain-of-thought verification—a goal that diverges fundamentally from both the alignment goal of RLHF/DPO/RLAIF (Alignment, Safety) and the Task Specialization goal of FT [ref]. The D3 (Data Requirements) coordinate is also a differentiator: Verifiable Task relies on programmatic reward signals (e.g., mathematical correctness verified by symbolic solvers, code execution verified by test suite pass/fail, logical entailment verified by formal proof checkers) rather than subjective human preference annotation. This substitution does not merely simplify the data pipeline; it categorically changes the nature of the validation obligation. The preference-pair regime (RLHF/DPO) requires mitigating annotator bias and tracking demographic representation; the verifiable-task regime requires instead certifying the correctness and coverage of the programmatic oracle—a shift from human-factors verification to formal functional-specification testing. The D3 (Data Requirements)-based differentiation between RLVR and RLHF is the clearest demonstration of the multi-dimensional taxonomy's analytical value: a mechanism-only single-axis framework collapses both into a generic ``RL-based tuning'' category, completely obscuring that they target distinct D2 (Goal) goals and require radically different D3 (Data Requirements) provenance documentation.",
"source_ref": "Alignment, Reasoning, and Trustworthiness",
"classification_tension": "",
"summary_editorial": "Trains on tasks where answers can be checked automatically — maths, code, proofs — so the reward signal comes from a verifier rather than an opinion."
},
{
"slug": "lp",
"tech_key": "lp",
"name": "LP",
"family": "Alignment, Reasoning, and Trustworthiness",
"is_reference_row": false,
"footnote_markers": [
"\\S"
],
"footnotes": [
{
"marker": "\\S",
"text": "Classification tensions involving these techniques are discussed in Supplementary Section S4."
}
],
"notes": [],
"d1": [
"parametric-update"
],
"d2": [
"explainability"
],
"d3": [
"small-labeled"
],
"d4": [
"ad-hoc-permanent",
"scheduled-permanent"
],
"d5": [
"modular"
],
"d6": [
"dl",
"fm",
"llm",
"mllm"
],
"definition_verbatim": "LP is framed as an interpretability and explainability intervention—determining its D2 (Goal) assignment as Explainability—rather than merely as a downstream evaluation metric, task performance, or behavioral alignment [ref]. By freezing all base model parameters and training only a lightweight linear classifier appended to a designated intermediate layer, LP surfaces latent representational structure as interpretable task-specific outputs without modifying the base model's core feature extraction function [ref]. The Parametric Update classification applies exclusively to the probe head's weight matrix; the base remains mathematically unchanged. Its Modular scope indicates that LP probes constitute structurally decoupled artifacts whose installation and removal leave the base model's parametric state and validated behavior unaffected. Compared to PEFT, which also introduces a modular artifact onto a frozen base, LP diverges critically on D2 (Goal): PEFT targets task performance (Computational Efficiency, Task Specialization), while LP targets representational transparency. Compared to LEACE, which operates via Parameter Composition to algebraically rewrite the base model's weight matrices, LP never touches the base weights at all. For production deployments requiring verifiable transparency, the probe's classification performance across sensitive concept dimensions can be documented as engineering evidence of representational alignment or bias. Because LP does not modify the base model's parameters, it does not trigger a model-level change-control event; the decoupled probe artifact (Modular) requires its own version-controlled documentation specifying the target layer, concept labels, and labeled probe dataset—but the base model's existing validation record remains entirely unaffected.",
"source_ref": "Alignment, Reasoning, and Trustworthiness",
"classification_tension": "LP. A strictly frozen linear probe does not modify the base model's internal representations or predictive behavior; it merely surfaces intermediate states. It is included in the adaptation framework strictly because appending a probe fundamentally alters the deployed system's explainability footprint and output characteristics (Explainability), triggering regulatory transparency mandates (e.g., EU AI Act Article 13) even when the core parameters remain untouched.",
"summary_editorial": "Trains a small classifier on the model's internal representations to see what information they contain, without changing the model itself."
},
{
"slug": "leace",
"tech_key": "leace",
"name": "LEACE",
"family": "Alignment, Reasoning, and Trustworthiness",
"is_reference_row": false,
"footnote_markers": [],
"footnotes": [],
"notes": [],
"d1": [
"parameter-composition"
],
"d2": [
"fairness"
],
"d3": [
"small-labeled"
],
"d4": [
"ad-hoc-permanent"
],
"d5": [
"partial"
],
"d6": [
"dl",
"fm",
"llm",
"mllm"
],
"definition_verbatim": "LEACE targets Fairness through a mechanistically novel D1 (Mechanism) pathway: rather than employing iterative GD, it operates via Parameter Composition—computing a covariance-based orthogonal projection from a Small Labeled and algebraically projecting the model's weight matrices into the null space of a target concept's linear representation [ref], a parameter composition operation analogous to Task Arithmetic but operating on concept subspaces rather than task capability directions. This closed-form linear algebra operation bypasses the computational footprint of traditional parameter-updating methods entirely, producing a Ad-hoc Permanent, Partial modification at negligible training cost. Compared to LP, which trains a probe head on top of a frozen base to expose linear concept encodings, LEACE inverts the operation: it takes the same linear concept subspace identified by probing and algebraically projects it out of the base model's weight matrices—making LP and LEACE complementary, with LP as the diagnostic step and LEACE as the corrective intervention targeting the same identified representation. The Ad-hoc Permanent Permanent coordinate reflects that LEACE produces an irreversible, event-driven modification at negligible compute cost—unlike gradient-based fairness interventions that require training runs, LEACE can be applied as a post-hoc weight surgery. The Partial scope is bounded by the targeted layers: only the weight matrices encoding the concept's linear subspace are projected; all other parameters remain unchanged. Its operational and validation significance lies in the deterministic auditability of the intervention: because the modification is derived analytically from a strict concept specification, it is fully reproducible and mathematically characterizable; the concept subspace, projection matrix, and labeled probe dataset constitute a complete audit trail for the fairness intervention—validation requires verifying that the modified representation is linearly indecipherable for the target concept while baseline task performance is preserved, both of which are statistically testable properties.",
"source_ref": "Alignment, Reasoning, and Trustworthiness",
"classification_tension": "",
"summary_editorial": "Surgically removes a concept's linear influence from the model's representations — used for fairness work — with a closed-form edit rather than training."
},
{
"slug": "advrstrn",
"tech_key": "AdvrsTrn",
"name": "Adversarial Training",
"family": "Alignment, Reasoning, and Trustworthiness",
"is_reference_row": false,
"footnote_markers": [],
"footnotes": [],
"notes": [],
"d1": [
"parametric-update"
],
"d2": [
"robustness"
],
"d3": [
"large-labeled"
],
"d4": [
"ad-hoc-permanent",
"scheduled-permanent"
],
"d5": [
"whole-model"
],
"d6": [
"dl",
"fm",
"llm"
],
"definition_verbatim": "Adversarial training improves model robustness by incorporating adversarially perturbed examples directly into the optimization objective through a min-max formulation [ref]. The Parametric Update coordinate reflects the defining operational property: adversarial perturbations are generated dynamically within the optimization loop, meaning the gradient signal itself is adversarially shaped at each training step—this is the mechanistic distinction from Data Augmentation, which modifies training inputs as a static pre-processing step independent of the gradient loop. This distinction is taxonomically consequential: adversarial training is a gradient-coupled technique whose computational cost scales with the inner maximization procedure at each step, while Data Augmentation's cost is absorbed in the data-loading stage. Both share Robustness, but only adversarial training carries Parametric Update with an adversarially shaped gradient signal, rather than the Pipeline-Mediated mechanism of augmentation. The DL, FM, LLM ceiling reflects an active research frontier: cross-modal adversarial robustness for vision-language MLLM architectures lacks standardized post-training protocols and established benchmark coverage, making the MLLM tier currently out of scope. Adversarial training universally trades off standard operational performance against robustness [ref]; this Pareto trade-off must be explicitly characterized in the validation record, with the robustness-accuracy frontier documented as a system performance boundary—any deployment threshold revision following adversarial training constitutes a system behavior change requiring separate regression validation.",
"source_ref": "Alignment, Reasoning, and Trustworthiness",
"classification_tension": "",
"summary_editorial": "Trains on deliberately perturbed inputs so the model holds up against adversarial examples rather than breaking on small, crafted changes."
},
{
"slug": "metalrn",
"tech_key": "metaLrn",
"name": "Meta-Learning",
"family": "Training Strategies",
"is_reference_row": false,
"footnote_markers": [],
"footnotes": [],
"notes": [],
"d1": [
"parametric-update"
],
"d2": [
"task-specialization"
],
"d3": [
"task-distribution"
],
"d4": [
"ad-hoc-permanent",
"scheduled-permanent"
],
"d5": [
"partial"
],
"d6": [
"dl",
"fm"
],
"definition_verbatim": "Meta-learning frames adaptation itself as an optimization objective: rather than learning to perform a single task, the model learns an initialization, metric space, or learning procedure that enables rapid generalization to new tasks from minimal data (Task Distribution) [ref]. Its D3 (Data Requirements) coordinate is the defining characteristic: instead of a labeled dataset for one task, meta-learning requires a diverse distribution of training tasks—a data-provenance structure qualitatively different from the task-specific corpora of Full FT (Large Labeled) or PEFT (Small Labeled). Its D6 (Model Type) ceiling at DL, FM encodes a significant structural observation within the taxonomy: at the LLM tier, ICL has architecturally supplanted gradient-based meta-learning as the primary mechanism for rapid few-shot adaptation. The supersession is mechanistically grounded: LLM contain sufficiently rich pre-trained representations that few-shot task generalization can occur through in-context conditioning (Context Injection, Session-Ephemeral) rather than through expensive parametric optimization (Parametric Update, Ad-hoc Permanent, Scheduled Permanent). This supersession relationship dictates that system engineers deploying LLM should route rapid adaptation requirements through prompt-level ICL architectures, reserving formal meta-learning pipelines exclusively for DL and early FM deployments. The D4 (Persistence) coordinate is counterintuitive for a technique framed around adaptability: meta-learning produces a permanent parametric modification (Ad-hoc Permanent, Scheduled Permanent)—not a transient or session-ephemeral one. This means meta-learning demands the same rigorous full-system regression testing and baseline validation protocols as any standard training event, structurally distinct from ICL's zero-artifact execution. Meta-learning requires that the task distribution used for outer-loop training be fully documented as the primary provenance record; the diversity and coverage of this task distribution directly determines the generalization bounds of the resulting meta-initialization, and any modification to the task distribution constitutes a change event requiring re-validation of the adaptation baseline.",
"source_ref": "Training Strategies",
"classification_tension": "",
"summary_editorial": "Trains a model to be good at learning new tasks quickly, rather than good at any one task — \"learning to learn\"."
},
{
"slug": "mtl",
"tech_key": "mtl",
"name": "MTL",
"family": "Training Strategies",
"is_reference_row": false,
"footnote_markers": [],
"footnotes": [],
"notes": [],
"d1": [
"parametric-update"
],
"d2": [
"task-specialization"
],
"d3": [
"task-distribution"
],
"d4": [
"ad-hoc-permanent",
"scheduled-permanent"
],
"d5": [
"whole-model"
],
"d6": [
"dl",
"fm",
"llm",
"mllm"
],
"definition_verbatim": "MTL jointly optimizes shared representations across a fixed task set, exploiting inter-task relatedness to improve generalization on each individual task [ref]. The defining coordinate is Task Distribution: the training corpus spans a predetermined collection of tasks simultaneously available in memory, rather than a single-task labeled dataset (Full FT, , Large Labeled). Its structural distinction from Meta-Learning is that, while both share Task Distribution, MTL's task set is statically fixed at training time rather than dynamically drawn from a distribution. The Whole-Model scope is the critical consequence of the joint optimization: because all task-specific output heads share the same backbone parameters, there is no modular isolation boundary between tasks—any change to the shared encoder propagates across all tasks simultaneously. This means modifying the task set, task-weighting schedule, or encoder architecture requires full-system regression testing across all tasks to verify that positive transfer to new tasks has not induced negative interference on existing ones. At the MLLM tier, MTL is the foundational architecture underlying multimodal instruction tuning—a unified backbone must serve heterogeneous vision, audio, and text task heads simultaneously. MTL requires comprehensive cross-task regression testing at every task-set modification boundary; the inter-task interference structure must be explicitly characterized and documented, as negative transfer to any task constitutes a behavioral regression on that task's validated performance envelope.",
"source_ref": "Training Strategies",
"classification_tension": "",
"summary_editorial": "Trains one model on several related tasks at once so they share representations and reinforce each other."
},
{
"slug": "selfplay",
"tech_key": "selfplay",
"name": "Self-Play",
"family": "Training Strategies",
"is_reference_row": false,
"footnote_markers": [
"\\ddagger"
],
"footnotes": [
{
"marker": "\\ddagger",
"text": "Self-play skips the FM tier because standard FM pre-training and FT pipelines do not employ self-play; it re-emerges at the LLM level through RLHF-adjacent techniques (e.g., self-play FT, SPIN ) that exploit autoregressive generation as an opponent model."
}
],
"notes": [],
"d1": [
"parametric-update"
],
"d2": [
"alignment",
"reasoning",
"task-specialization"
],
"d3": [
"environment-derived"
],
"d4": [
"ad-hoc-permanent",
"scheduled-permanent"
],
"d5": [
"whole-model"
],
"d6": [
"dl",
"llm"
],
"definition_verbatim": "Self-play generates training data by having a model compete against or evaluate itself, producing an auto-curriculum of increasing difficulty without requiring externally provided labeled data. The Environment-Derived coordinate is the defining structural property: the training signal is generated internally through the model's own interactions with copies or historical checkpoints of itself—no human annotation, no static corpus, and no external oracle are involved in data generation. This self-referential data source produces the compound Alignment, Reasoning, Task Specialization profile—the three operationally distinct goals served by self-play at different D6 (Model Type) levels: domain mastery through competitive co-evolution in DL game agents (Task Specialization) [ref], instruction-following improvement through self-generated preference pairs in LLM alignment (Alignment) [ref], and chain-of-thought reasoning cultivation through iterative self-verification in reasoning models (Reasoning) [ref]. This multi-goal profile would be irresolvable in a single-axis taxonomy; the taxonomy's value is precisely in making explicit that the same D1 (Mechanism) and D3 (Data Requirements) regime serve fundamentally different D2 (Goal) depending on the deployment context—and that each dictates entirely distinct evaluation metrics, regression test suites, and operational baselines. Compared to Curriculum Learning, which also generates structured training sequences, self-play mainly differs on D1 (Mechanism) (Parametric Update vs. Pipeline-Mediated) and D3 (Data Requirements) (Environment-Derived vs. Pipeline-Dependent)—Curriculum Learning sequences existing data, while self-play generates new data through model-environment interaction. The DL scope (skipping FM) reflects that standard FM FT pipelines do not typically employ self-play, which re-emerges only as a specialized adaptation mechanism for alignment and reasoning at the LLM tier. Because the specific D2 (Goal) being pursued dictates the validation protocol, any self-play deployment must explicitly document which goal dimension applies, providing a clear audit trail for the intended-purpose and risk-rationale in compliance documentation.",
"source_ref": "Training Strategies",
"classification_tension": "",
"summary_editorial": "Improves a model by having it compete against or critique itself, generating its own training signal without external labels."
},
{
"slug": "currilrn",
"tech_key": "curriLrn",
"name": "Curriculum Learning",
"family": "Training Strategies",
"is_reference_row": false,
"footnote_markers": [],
"footnotes": [],
"notes": [],
"d1": [
"pipeline-mediated"
],
"d2": [
"task-specialization"
],
"d3": [
"pipeline-dependent"
],
"d4": [
"pipeline-dependent"
],
"d5": [
"pipeline-dependent"
],
"d6": [
"dl",
"fm",
"llm",
"ml",
"mllm"
],
"definition_verbatim": "Curriculum learning structures the ordering of training examples rather than their content, presenting data in an easy-to-hard sequence to improve optimizer trajectory and final convergence [ref]. Its Pipeline-Mediated classification is the decisive taxonomic property: curriculum learning never executes a gradient update on the model—it operates strictly as a data-sequencing function upstream of the optimizer. The parametric modification that results from training on a curricularized dataset is performed entirely by the downstream training method; curriculum learning contributes only the temporal ordering of the input stream. This indirect mechanism is the structural reason for Pipeline-Dependent and Pipeline-Dependent: the D4 (Persistence) and D5 (Scope) profile of the resulting model are fully inherited from whichever downstream training method consumes the sequenced data. Compared to Data Augmentation, Data Augmentation synthetically extends the training set (via transformation), while curriculum learning operates on the existing dataset without modification—it changes the sampling order, not the examples. Compared to Active Learning, both share Pipeline-Dependent, but Active Learning introduces a Distributional Gap Bridging component by querying strategically chosen unlabeled samples, while Curriculum Learning's Task Specialization reflects that it optimizes the convergence trajectory rather than the distributional coverage. Because curriculum learning does not independently modify model parameters, it must be documented as a named algorithmic dependency of the primary training event it serves—specifying the difficulty-scoring methodology, sequencing protocol, and any pacing functions—rather than as a standalone change-control entry; the downstream training event inherits the full parametric and structural documentation burden.",
"source_ref": "Training Strategies",
"classification_tension": "",
"summary_editorial": "Orders training examples from easy to hard, so the model builds up competence in stages rather than facing everything at once."
},
{
"slug": "activelrn",
"tech_key": "activeLrn",
"name": "Active Learning",
"family": "Training Strategies",
"is_reference_row": false,
"footnote_markers": [],
"footnotes": [],
"notes": [],
"d1": [
"pipeline-mediated"
],
"d2": [
"distributional-gap-bridging",
"task-specialization"
],
"d3": [
"pipeline-dependent"
],
"d4": [
"pipeline-dependent"
],
"d5": [
"pipeline-dependent"
],
"d6": [
"dl",
"fm",
"llm",
"ml",
"mllm"
],
"definition_verbatim": "Active learning strategically selects the most informative unlabeled samples for oracle annotation, maximizing model improvement per labeled example by targeting the data points the model is most uncertain about [ref]. Like Curriculum Learning, it carries Pipeline-Mediated: the model's parameters are updated by a downstream training method, not by the active learning mechanism itself. However, Active Learning diverges from Curriculum Learning on D2 (Goal): where curriculum learning sequences existing labeled data for convergence optimization (Task Specialization only), active learning's query strategy inherently introduces a Distributional Gap Bridging component—by selectively targeting uncertainty-rich, informationally dense regions of the unlabeled distribution, it constructs a labeled dataset whose coverage is deliberately biased toward the model's current capability boundaries. This D2 (Goal) distinction is operationally significant: curriculum learning does not change the statistical composition of the labeled dataset; active learning does, and the algorithmic selection strategy is the direct determinant of the distributional coverage of the resulting training corpus. The Pipeline-Dependent and Pipeline-Dependent coordinates confirm that persistence and structural scope are fully inherited from the downstream training method—but the labeled dataset itself, as an artifact of the active selection process, constitutes an independent data-provenance record that must be documented and audited separately. The active learning acquisition function must be explicitly documented and version-controlled as a system configuration artifact; the selection strategy directly defines the boundaries of the model's validated operational envelope, and any change to the acquisition function constitutes a data-pipeline modification requiring re-evaluation of the distributional coverage of the resulting labeled corpus before downstream training commences.",
"source_ref": "Training Strategies",
"classification_tension": "",
"summary_editorial": "Lets the model choose which examples get labelled next, concentrating scarce annotation effort where it will help most."
},
{
"slug": "dataaug",
"tech_key": "dataAug",
"name": "Data Augmentation",
"family": "Data-Centric and Privacy-Preserving Methods",
"is_reference_row": false,
"footnote_markers": [],
"footnotes": [],
"notes": [],
"d1": [
"pipeline-mediated"
],
"d2": [
"robustness",
"task-specialization"
],
"d3": [
"pipeline-dependent"
],
"d4": [
"pipeline-dependent"
],
"d5": [
"pipeline-dependent"
],
"d6": [
"dl",
"fm",
"llm",
"ml",
"mllm"
],
"definition_verbatim": "Data Augmentation increases training data diversity through synthetic transformations of existing examples, without acquiring new labeled data [ref]. Its Pipeline-Mediated classification reflects its structural position: augmentation operates strictly upstream of the gradient loop, modifying training inputs before they reach the optimizer—it executes no forward pass on the model and computes no gradients. This upstream positioning means the technique shares Pipeline-Dependent and Pipeline-Dependent coordinates with all other boundary-extension methods (Curriculum Learning, Active Learning, Semi-SL), reflecting that the D4 (Persistence) and D5 (Scope) of any resulting model change are entirely determined by the downstream training method that consumes the augmented data. The compound Robustness, Task Specialization profile reveals a structurally important internal distinction within augmentation: standard transformations (flips, crops, synonym substitution) diversify the input distribution to improve generalization (Task Specialization), while adversarially constructed perturbations specifically target distributional robustness (Robustness) [ref]. The latter variant overlaps in goal with Adversarial Training, but differs critically on D1 (Mechanism): adversarial training generates perturbations dynamically within the gradient loop (dParametric Update with adversarially shaped gradients), while adversarial augmentation applies fixed perturbation transforms as a static pre-processing step (Pipeline-Mediated)—an operationally meaningful distinction because the dynamic coupling in adversarial training produces adaptive perturbations that scale with the model's current vulnerability, while static augmentation perturbations cannot. The augmentation policy—comprising the exact set of transformation functions, their mathematical parameterizations, hyperparameter bounds, and random seed configurations—constitutes a component of the training data specification and must be version-controlled and documented as an immutable dependency of the training event; a change to the augmentation policy redefines the statistical distribution the model optimizes against and requires a new data-provenance record.",
"source_ref": "Data-Centric and Privacy-Preserving Methods",
"classification_tension": "",
"summary_editorial": "Expands the training set with transformed copies of existing data, improving robustness without collecting anything new."
},
{
"slug": "semisl",
"tech_key": "semiSl",
"name": "Semi-SL",
"family": "Data-Centric and Privacy-Preserving Methods",
"is_reference_row": false,
"footnote_markers": [],
"footnotes": [],
"notes": [],
"d1": [
"pipeline-mediated"
],
"d2": [
"task-specialization"
],
"d3": [
"small-labeled",
"unlabeled"
],
"d4": [
"pipeline-dependent"
],
"d5": [
"pipeline-dependent"
],
"d6": [
"dl",
"fm",
"llm",
"ml",
"mllm"
],
"definition_verbatim": "Semi-SL leverages abundant unlabeled data alongside a small labeled set, making it highly valuable in domains where annotation is prohibitively expensive. Its Pipeline-Mediated classification groups it with the other data-pipeline strategies, but Semi-SL occupies a structurally distinct position within this group: unlike Curriculum Learning (which sequences existing labeled data) and Data Augmentation (which transforms existing labeled data), Semi-SL introduces unlabeled data—whose labels were never verified by a human or certified oracle—as a training signal source [ref]. Semi-SL is uniquely complex because it integrates two distinct data sources (Small Labeled, Unlabeled). Because the unlabeled pool lacks ground-truth verification, its inherent sampling bias and distributional coverage directly shape the model’s learned behavior, creating a structural transparency gap that distinguishes Semi-SL from purely supervised pipelines. Compared to SSL/CPT, which shares the Unlabeled dependency and uses self-supervised objectives directly, Semi-SL is Pipeline-Mediated—it does not itself prescribe an optimization objective or update parameters; it delivers a mixed labeled-and-pseudo-labeled dataset to a downstream training method. SSL/CPT is Parametric Update, defining its own objective and directly executing weight modifications. Because the unlabeled data pool's distributional properties directly determine the scope and boundaries of the downstream model's validated capabilities, the unlabeled corpus must be subjected to the same provenance documentation standards as the labeled set—including source identification, collection methodology, distributional characterization, and bias auditing.",
"source_ref": "Data-Centric and Privacy-Preserving Methods",
"classification_tension": "",
"summary_editorial": "Trains on a small labelled set alongside a large unlabelled one, drawing a learning signal from the unlabelled data rather than leaving it unused."
},
{
"slug": "ssl",
"tech_key": "ssl",
"name": "SSL / CPT",
"family": "Data-Centric and Privacy-Preserving Methods",
"is_reference_row": false,
"footnote_markers": [
"\\dagger"
],
"footnotes": [
{
"marker": "\\dagger",
"text": "SSL is classified in its adaptation role (CPT on domain-specific unlabeled data). As a foundational pre-training paradigm, it falls outside the adaptation taxonomy."
}
],
"notes": [],
"d1": [
"parametric-update"
],
"d2": [
"capability-extension",
"distributional-gap-bridging"
],
"d3": [
"unlabeled"
],
"d4": [
"scheduled-permanent",
"unbounded-cumulative"
],
"d5": [
"partial",
"whole-model"
],
"d6": [
"dl",
"fm",
"llm",
"mllm"
],
"definition_verbatim": "SSL/CPT occupies a structurally dual role in the taxonomy (marked dagger in Table ): foundational pre-training from scratch is explicitly out of scope, but applying SSL as an adaptation technique to an already-deployed architecture is within scope. The critical coordinate that distinguishes SSL-as-adaptation from all other pipeline-mediated data strategies is Parametric Update [ref]: unlike Data Augmentation or Semi-SL, which modify the data pipeline and delegate parameter updates to a downstream method, CPT directly defines its own optimization objective and formally executes GD on model weights. The Unlabeled coordinate is shared with Semi-SL, but the mechanism is structurally distinct: Semi-SL uses unlabeled data to construct pseudo-labeled training examples for a downstream training method; CPT uses unlabeled data as the direct training signal for its own self-supervised objective, making it a non-supervised parametric adaptation technique in the taxonomy. The D4 (Persistence) coordinate bifurcates based on execution mode: Scheduled Permanent when CPT is applied as a discrete offline batch on a fixed domain corpus; Unbounded Cumulative when applied as continuous domain-specific streaming—the latter creating an open-ended modification trajectory that breaks the assumptions of static regression testing and demands dynamic behavioral monitoring. At the LLM and FM tiers, the primary adaptation strategy leveraging SSL objectives is CPT: applying self-supervised objectives (e.g., next-token prediction) to unlabeled, domain-specific corpora to internalize novel vocabulary, syntax, and factual structures without requiring supervised labels. CPT functions as an initialization phase preceding SFT or PEFT, providing a domain-enriched representational baseline that ensures the model possesses foundational domain competency before task-specific supervision is applied [ref].",
"source_ref": "Data-Centric and Privacy-Preserving Methods",
"classification_tension": "SSL/CPT. Has a dual role: pre-training paradigm (out of scope) and adaptation technique (in scope); unlike other data-manipulation techniques, SSL-as-adaptation prescribes its own training objective and receives Scheduled Permanent, Unbounded Cumulative. Machine Unlearning vs.\\ Knowledge Editing ( vs. ). share D1 and D4 but differ on D2: unlearning targets Remove Knowledge while editing targets Knowledge Update. The boundary blurs when correction effectively requires erasing the old association. Activation Steering vs.\\ TTA ( vs. ). Both produce Transient adaptations but differ on model type (LLM, MLLM vs.\\ DL, FM), suggesting transient adaptation bifurcates at the FM/LLM boundary. Machine Unlearning and Task Arithmetic ( vs. ). Share Parameter Composition (task-vector manipulation): unlearning via negation is structurally identical to Task Arithmetic with a minus sign, raising the question of whether unlearning is a distinct technique or a specific application of Task Arithmetic.",
"summary_editorial": "Continues pre-training on unlabelled data from a specialist domain, so the model absorbs its patterns and vocabulary before any task-specific tuning."
},
{
"slug": "dpft",
"tech_key": "dpft",
"name": "DP-FT",
"family": "Data-Centric and Privacy-Preserving Methods",
"is_reference_row": false,
"footnote_markers": [],
"footnotes": [],
"notes": [],
"d1": [
"parametric-update"
],
"d2": [
"privacy-preservation"
],
"d3": [
"large-labeled",
"small-labeled"
],
"d4": [
"ad-hoc-permanent",
"scheduled-permanent"
],
"d5": [
"partial",
"whole-model"
],
"d6": [
"dl",
"fm",
"llm",
"mllm"
],
"definition_verbatim": "DP-FT, using DP-SGD, incorporates gradient clipping and noise injection during training to provide mathematical guarantees against membership inference and data extraction attacks. DP-FT is a Parametric Update dedicated exclusively to the Privacy Preservation goal. Unlike FL, which achieves privacy via Distributed data sovereignty, DP-FT operates on centralized data (Large Labeled, Small Labeled) but mathematically protects the records. Like standard parametric updates, DP-FT produces a permanently modified weight checkpoint that can be deployed on a planned cadence or as an event-driven patch, assigning it Ad-hoc Permanent, Scheduled Permanent. It can be applied to the Whole-Model or as DP-PEFT (Partial). Because the DP-SGD algorithm is foundational, its applicability spans deeply into the hierarchy, yielding DL, FM, LLM, MLLM.",
"source_ref": "Data-Centric and Privacy-Preserving Methods",
"classification_tension": "",
"summary_editorial": "Fine-tunes with mathematical privacy guarantees, bounding how much any single training record can influence the resulting model."
},
{
"slug": "fl",
"tech_key": "fl",
"name": "FL",
"family": "Data-Centric and Privacy-Preserving Methods",
"is_reference_row": false,
"footnote_markers": [],
"footnotes": [],
"notes": [],
"d1": [
"parametric-update"
],
"d2": [
"privacy-preservation"
],
"d3": [
"decentralized"
],
"d4": [
"scheduled-permanent"
],
"d5": [
"distributed"
],
"d6": [
"dl",
"fm",
"llm",
"ml",
"mllm"
],
"definition_verbatim": "FL trains a global model by aggregating gradient updates from local models on decentralized datasets, ensuring raw training data never leaves local devices [ref]. Its Parametric Update classification reflects that FL does execute GD and produce formal weight modifications—distinguishing it from pipeline-mediated data strategies—but the critical structural differentiator is Decentralized: no centralized training corpus exists. Each client computes gradients locally on its private data shard and transmits only model updates (weight deltas or gradients) to the server for aggregation; raw data never leaves the originating device. This decentralized data geometry is what justifies Privacy Preservation as a D2 (Goal): FL does not merely apply privacy as a constraint layered onto a task-training objective—privacy preservation is an intrinsic structural property of the data access pattern, not an add-on mechanism. The absence of centralized training data means that data-quality auditing, distributional characterization, and bias assessment must be delegated to client-local verification processes or statistical reporting protocols; the server cannot directly inspect the training corpus, making third-party data-provenance attestation mechanisms a system-engineering requirement rather than an optional audit supplement.",
"source_ref": "Data-Centric and Privacy-Preserving Methods",
"classification_tension": "",
"summary_editorial": "Trains across many devices or institutions without centralising their data — each site computes updates locally and only those are shared."
},
{
"slug": "kd",
"tech_key": "kd",
"name": "KD",
"family": "Efficiency and Composition",
"is_reference_row": false,
"footnote_markers": [],
"footnotes": [],
"notes": [],
"d1": [
"cross-model-transfer"
],
"d2": [
"computational-efficiency"
],
"d3": [
"teacher-derived"
],
"d4": [
"ad-hoc-permanent",
"scheduled-permanent"
],
"d5": [
"surrogate-model"
],
"d6": [
"dl",
"fm",
"llm",
"mllm"
],
"definition_verbatim": "KD transfers a trained teacher model's learned behavioral distribution into a smaller student model, most commonly by training the student to match the teacher's soft probability output distribution, enabling compression of capability without commensurate compression of performance [ref]. Its Cross-Model Transfer classification captures the defining structural property: a complete, separately trained model is the proximate information source for the student's training—not a labeled dataset, not a teacher's parameters directly, but the teacher's output distribution. This cross-model information flow distinguishes KD from all other compression-adjacent techniques: Model Compression operates on a single model's own parameters while KD requires a functioning teacher model to generate training signal, making Teacher-Derived. The Computational Efficiency goal drives the student's architectural reduction (smaller, faster inference). From a validation standpoint, a KD event establishes a strict behavioral dependency chain: verifying the student model requires comprehensive technical documentation of both the teacher model's original training data provenance and the exact algorithmic distillation protocol used to generate the soft targets, ensuring full auditability of the behavioral alignment downstream. The student artifact inherits the teacher's behavioral envelope but introduces a compression-induced approximation gap; this gap must be characterized quantitatively—via behavioral regression testing across the teacher's validated operational envelope.",
"source_ref": "Efficiency and Composition",
"classification_tension": "",
"summary_editorial": "Trains a smaller model to imitate a larger one's outputs, keeping much of the capability at a fraction of the serving cost."
},
{
"slug": "compress",
"tech_key": "compress",
"name": "Model Compression",
"family": "Efficiency and Composition",
"is_reference_row": false,
"footnote_markers": [],
"footnotes": [],
"notes": [],
"d1": [
"parametric-compression"
],
"d2": [
"computational-efficiency"
],
"d3": [
"parameter-only",
"unlabeled"
],
"d4": [
"ad-hoc-permanent",
"scheduled-permanent"
],
"d5": [
"compressed-artifact"
],
"d6": [
"dl",
"fm",
"llm",
"mllm"
],
"definition_verbatim": "Model compression reduces the computational footprint of an already-trained model through structural modification of its parameters, producing a deployable compressed artifact with lower memory and inference cost [ref]. Its Parametric Compression classification explicitly encodes reduction as the defining operation: unlike all Parametric Update techniques that modify parameters to improve capability, Parametric Compression modifies parameters to reduce resource consumption while attempting to preserve capability. Its Computational Efficiency coordinate designates an objective of footprint reduction that inherently lacks a target for intrinsic capability development. Compared to KD, which achieves efficiency through architectural reduction of a student trained from teacher outputs (Cross-Model Transfer, Teacher-Derived), Model Compression achieves efficiency by directly restructuring a single model's existing parameters, i.e., no second model is required. Its D3 (Data Requirements) profile diverges fundamentally based on the chosen algorithmic implementation: Data-free compression methods operate strictly as Parameter-Only interventions; because they consume zero training or calibration data, they completely bypass data-provenance pipelines, making them optimal for strict zero-data-access operational deployments. Conversely, calibration-dependent compression methods require a Unlabeled calibration dataset to calculate activation statistics and minimize quantization error [ref]. The Compressed Artifact is a structurally altered model artifact—not a parameter overlay applied to the original—and must be registered as a distinct model version requiring its own validation suite: the compression operation introduces approximation errors whose distribution across the input space cannot be inferred from the original model's validation record; full behavioral regression testing on the compressed artifact against the original model's validated operational envelope is required.% before deployment equivalence can be asserted. From a regulatory and change-control perspective, this category encompasses two dimensionally distinct approaches. Post-Training Quantization (PTQ, e.g., GPTQ [ref]) operates as a training-data-free (Parameter-Only) algebraic derivation applied after the model is fully trained. In contrast, QAT simulates lower precision during active GD, classifying it as a distinct Parametric Update training event with differing validation and data-provenance requirements.",
"source_ref": "Efficiency and Composition",
"classification_tension": "",
"summary_editorial": "Shrinks a model through quantisation, pruning, or similar, so it fits and runs where the original could not."
},
{
"slug": "taskarith",
"tech_key": "taskArith",
"name": "Task Arith. & Model Merging",
"family": "Efficiency and Composition",
"is_reference_row": false,
"footnote_markers": [],
"footnotes": [],
"notes": [],
"d1": [
"parameter-composition"
],
"d2": [
"capability-extension",
"computational-efficiency"
],
"d3": [
"parameter-only"
],
"d4": [
"ad-hoc-permanent",
"scheduled-permanent"
],
"d5": [
"fused-composition"
],
"d6": [
"fm",
"llm",
"mllm"
],
"definition_verbatim": "Task Arithmetic and Model Merging combine pre-trained and fine-tuned checkpoints into a single multi-capable model through arithmetic operations on parameter space, without any additional training data [ref]. Its Parameter Composition and Parameter-Only classification suggest no gradient computation, no data consumption at merge time, no teacher model inference—only algebraic operations on weight tensors. Machine Unlearning shares Parameter Composition via task-vector negation (subtracting a forget-task vector to remove a learned capability), making Machine Unlearning a targeted polarity reversal of the exact same algebraic operation that Task Arithmetic uses to add capabilities. Compared to MTL, which achieves multi-task capability through joint optimization over a shared task-distribution corpus (Parametric Update, Task Distribution), Task Arithmetic achieves the same compound capability profile through post-hoc parameter composition requiring no additional training and no access to the original task datasets. To mitigate parameter interference and preserve specialized capabilities during the fusion process, practitioners employ advanced mathematical merge operators, including TIES-MERGING [ref], DARE [ref], SLERP [ref], and Fisher merging [ref]. Consequently, the regression validation suites used to verify targeted machine unlearning—specifically, verifying the erasure of a specific capability while ensuring the preservation of adjacent foundational knowledge—are directly applicable to Task Arithmetic (task-vector operations). Because Model Merging bypasses traditional training telemetry, system validation must rely on downstream functional testing and post-merge automated regression benchmarks to verify that parameter-space interference has not degraded core model capabilities.",
"source_ref": "Efficiency and Composition",
"classification_tension": "",
"summary_editorial": "Combines separately tuned models by arithmetic on their weights — adding capabilities, or subtracting them by negating a task vector."
},
{
"slug": "lce",
"tech_key": "lce",
"name": "Long-Context Ext.",
"family": "Efficiency and Composition",
"is_reference_row": false,
"footnote_markers": [],
"footnotes": [],
"notes": [],
"d1": [
"architectural-modification",
"parametric-update"
],
"d2": [
"capability-extension"
],
"d3": [
"small-labeled",
"unlabeled"
],
"d4": [
"ad-hoc-permanent",
"scheduled-permanent"
],
"d5": [
"partial"
],
"d6": [
"fm",
"llm",
"mllm"
],
"definition_verbatim": "Long-Context Extension methods (e.g., PI [ref], YaRN [ref], and PEPE [ref]) extend a language model's operational context window by adapting its positional encoding (e.g., through RoPE [ref] scaling) and, when required, fine-tuning on long-sequence data, enabling the model to process inputs far exceeding its original context length. This operates via a composite D1 (Mechanism): it requires an Architectural Modification to alter the frequency scaling of the RoPE, typically followed by a brief Parametric Update on long documents. The goal is strictly Capability Extension, as it structural expands the model's capacity rather than teaching a new task. It targets the Partial scope and utilizes Small Labeled, Unlabeled long-sequence corpora. Because it permanently alters the positional embedding architecture and corresponding parameters, it produces a versioned checkpoint exhibiting Ad-hoc Permanent, Scheduled Permanent persistence. Unlike techniques that require generation mechanics, context scaling can be applied to base GenFM, establishing its D6 (Model Type) footprint at FM, LLM, MLLM.",
"source_ref": "Efficiency and Composition",
"classification_tension": "",
"summary_editorial": "Extends how much text a model can attend to at once, so longer documents fit in a single pass."
},
{
"slug": "moe",
"tech_key": "moe",
"name": "MoE",
"family": "Efficiency and Composition",
"is_reference_row": false,
"footnote_markers": [
"\\S"
],
"footnotes": [
{
"marker": "\\S",
"text": "Classification tensions involving these techniques are discussed in Supplementary Section S4."
}
],
"notes": [],
"d1": [
"architectural-modification",
"parametric-update"
],
"d2": [
"capability-extension",
"computational-efficiency"
],
"d3": [
"small-labeled",
"unlabeled"
],
"d4": [
"ad-hoc-permanent",
"scheduled-permanent"
],
"d5": [
"modular",
"partial"
],
"d6": [
"fm",
"llm",
"mllm"
],
"definition_verbatim": "MoE Adaptation introduces sparse gating over a population of expert sub-networks or integrating novel expert sub-networks into an existing MoE, enabling a model to route each input to the subset of parameters most relevant to that input's characteristics—achieving scalable multi-task specialization without proportionally scaling inference cost [ref]. This structural distinction necessitates a compound D1 (Mechanism) profile: the topological addition of a new expert constitutes Architectural Modification—physically restructuring the computational graph—while the subsequent tuning of expert weights or the central routing mechanism constitutes Parametric Update. This graph-level intervention is what mandates Modular: the model is a collection of individually addressable expert modules coordinated by a router—each expert constituting an independently auditable behavioral unit whose activation patterns can be tracked per input. Compared to PEFT, which adds lightweight parameter overlays (adapters, LoRA) to a frozen base with Modular, Partial, MoE Adaptation adds structural routing logic that fundamentally changes the model's computational topology—adapter modules are additive overlays; expert modules are gated substitutes for dense computation. The compound D2 (Goal) profile—Capability Extension, Computational Efficiency—reflects a unique operational property: MoE adaptation is the only technique in the taxonomy that simultaneously scales systemic capacity while explicitly preserving fixed inference latency bounds. The Architectural Modification classification has stringent system-engineering implications in the taxonomy: a new model topology is being registered, not a parametric override of an existing one—requiring a complete re-specification of the model's computational graph and expert routing logic audit trail as components of the system configuration record, since any change to router architecture or gating threshold alters which computation paths are activated for which inputs.",
"source_ref": "Efficiency and Composition",
"classification_tension": "MoE adaptation. Straddles architecture and adaptation; the row captures the adaptation act (training routers, adding experts) rather than the architectural pattern. RLVR vs.\\ RLHF/DPO ( vs.\\ ). All share Parametric Update but differ on D2 (Reasoning vs. Alignment, Safety) and D3 (Verifiable Task vs. Preference Pairs). RLVR and DPO share Partial, Whole-Model, while RLHF adds Surrogate Model. The boundary blurs when RLVR induces alignment-like properties or when RLHF pipelines incorporate verifiable components; in production, both are often applied sequentially to the same base model.",
"summary_editorial": "Adapts a model by adding expert sub-networks or training the router that picks between them, so capacity grows without every parameter running on every input."
},
{
"slug": "pe",
"tech_key": "pe",
"name": "PE",
"family": "Inference-Time Adaptation",
"is_reference_row": false,
"footnote_markers": [
"\\S"
],
"footnotes": [
{
"marker": "\\S",
"text": "Classification tensions involving these techniques are discussed in Supplementary Section S4."
}
],
"notes": [],
"d1": [
"context-injection"
],
"d2": [
"behavior-control",
"knowledge-update"
],
"d3": [
"zero-shot"
],
"d4": [
"session-ephemeral",
"version-persistent"
],
"d5": [
"input-output-space"
],
"d6": [
"llm",
"mllm"
],
"definition_verbatim": "PE designs input text to guide a deployed model's output distribution entirely through natural language instructions placed in the context window, without requiring any parameter modification, representing the lowest-overhead adaptation method in the taxonomy [ref]. Its Context Injection classification suggests that it produces no persistent model artifact whatsoever—no checkpoint, no parameter delta, no weight version, no trained embedding. The behavioral effect exists only while the prompt occupies the active context window (Session-Ephemeral) and operates exclusively at the model's input/output interface (I/O Space). Compared to ICL, which shares Context Injection and Session-Ephemeral but adds Few Demonstrations to steer task-specific output patterns through example-based conditioning, PE operates with Zero-Shot through instruction specification alone—relying on generalization from the model's pre-training rather than in-context exemplar matching. This zero-data, zero-gradient profile positions PE as the Behavior Control technique in the taxonomy that imposes the lowest modification overhead but provides the weakest behavioral guarantee: steering is non-parametric and entirely contingent on the model's pre-existing instruction-following capability. Compared to Prompt Learning, which replaces discrete natural-language tokens with continuously differentiable embedding vectors optimized via GD (Parametric Update, Small Labeled), PE operates entirely in human-legible natural language—making it interpretable but non-optimizable. The LLM, MLLM coordinate reflects that effective PE requires a model with emergent instruction-following capability; below LLM scale, models lack the generalization from instruction composition. Because PE produces no auditable architectural or parametric artifact, every deployed system whose behavior is governed by a prompt template must specify that template as a versioned configuration artifact; any change to the prompt—including rewording, structural reordering, or instruction addition—must be treated as a system configuration change requiring re-evaluation of behavioral coverage and output boundary characterization. % before redeployment. The system-level variant—where core system prompts or metaprompts are hardcoded into production deployment infrastructure—exhibits Version-Persistent persistence. This boundary requires rigorous tracking within software configuration management and release engineering protocols, as changes to the global system prompt dictate the functional behavioral envelope of the entire application despite the underlying static model weights.",
"source_ref": "Inference-Time Adaptation",
"classification_tension": "",
"summary_editorial": "Changes what you write in the prompt to steer the model — no weights change, and by default the effect lasts only for that request."
},
{
"slug": "pl",
"tech_key": "pl",
"name": "Prompt Learning",
"family": "Inference-Time Adaptation",
"is_reference_row": false,
"footnote_markers": [],
"footnotes": [],
"notes": [],
"d1": [
"parametric-update"
],
"d2": [
"task-specialization"
],
"d3": [
"small-labeled"
],
"d4": [
"ad-hoc-permanent",
"scheduled-permanent"
],
"d5": [
"partial"
],
"d6": [
"fm",
"llm",
"mllm"
],
"definition_verbatim": "Prompt Learning converts the discrete, human-authored natural-language prompt into a set of continuously differentiable embedding vectors—soft tokens—prepended to the model's input representation and optimized via GD on a task dataset [ref]. %, occupying the structural boundary between context injection and parametric adaptation. It resembles PE in its input-space operation—prepending sequence tokens to the input—but functions mechanically like FT in its optimization: learnable continuous embeddings (soft tokens) are explicitly trained via backpropagation (Parametric Update).% through the frozen base model. Therefore, the parametric modification is confined entirely to the prepended embedding vectors (Partial); all base model parameters remain frozen. The Ad-hoc Permanent, Scheduled Permanent coordinate means trained soft prompt tensors must be registered as distinct parametric artifacts in the system configuration record: deploying a different soft prompt tensor with the same frozen base model constitutes a parametric modification to the deployed system and requires version control and behavioral regression testing against the task coverage specification defined during original validation.",
"source_ref": "Inference-Time Adaptation",
"classification_tension": "",
"summary_editorial": "Learns the prompt itself as trainable parameters rather than writing it by hand — continuous vectors prepended to the input."
},
{
"slug": "apo",
"tech_key": "apo",
"name": "APO",
"family": "Inference-Time Adaptation",
"is_reference_row": false,
"footnote_markers": [],
"footnotes": [],
"notes": [],
"d1": [
"context-injection"
],
"d2": [
"behavior-control",
"task-specialization"
],
"d3": [
"few-demonstrations",
"zero-shot"
],
"d4": [
"version-persistent"
],
"d5": [
"input-output-space"
],
"d6": [
"llm",
"mllm"
],
"definition_verbatim": "APO frameworks, such as DSPy [ref] or OPRO [ref], replace manual prompt engineering with algorithmic search, treating language models as programmable engines. APO utilizes Context Injection within the Input/Output-Space, meaning the underlying model parameters remain completely frozen. However, because the optimization compiles a tested, metric-optimized prompt pipeline, the resulting artifact is classified as Version-Persistent rather than Session-Ephemeral. Data requirements are flexible: APO can utilize Few Demonstrations to bootstrap the optimization, but it can also operate strictly Zero-Shot by relying entirely on programmatic evaluation metrics or an "LLM-as-a-judge" [ref] to score unannotated inputs. Ultimately, this algorithmic search aims for both Task Specialization (optimizing accuracy on a target task) and Behavior Control (systematically enforcing strict output formats, stylistic guidelines, or reasoning templates without altering weights).",
"source_ref": "Inference-Time Adaptation",
"classification_tension": "",
"summary_editorial": "Automatically searches for better prompt wording, optimising the instruction rather than the model."
},
{
"slug": "rag",
"tech_key": "rag",
"name": "RAG",
"family": "Inference-Time Adaptation",
"is_reference_row": false,
"footnote_markers": [],
"footnotes": [],
"notes": [],
"d1": [
"context-injection"
],
"d2": [
"knowledge-update"
],
"d3": [
"external-corpus"
],
"d4": [
"version-persistent"
],
"d5": [
"input-output-space"
],
"d6": [
"llm",
"mllm"
],
"definition_verbatim": "RAG augments LLM generation by retrieving relevant documents from an external knowledge base and injecting them into the input context at inference time [ref]. Its non-parametric nature (Context Injection) and Version-Persistent coordinate confirm that RAG strictly does not modify underlying model parameters: the injected knowledge alters the model's transient output without touching its foundational weights. This fundamentally distinguishes RAG from Knowledge Editing, which permanently rewrites factual associations within the weight space itself. The information source is a separately maintained, indexable knowledge repository whose content is decoupled from the model's weights and can be updated independently (External Corpus)—documents can be added, revised, or removed from the corpus. This decoupling is the sharpest structural contrast with CPT, which internalizes domain knowledge directly into model weights (Parametric Update, Scheduled Permanent, Unbounded Cumulative): CPT makes knowledge intrinsic and persistent at the cost of a full retraining cycle for every knowledge update; RAG makes knowledge external and updateable at inference time at the cost of retrieval latency and context-window occupation. The compound Knowledge Update, Task Specialization profile reflects the dual function: the retrieval mechanism enables the model to answer with temporally current information (Knowledge Update) while the retrieved context shapes the generation to the specific domain and task at hand (Task Specialization). RAG functions as a complement to FT rather than a direct substitute: FT encodes structural behaviors and stylistic patterns directly into parameters, while RAG retrieves localized factual context from an external corpus. From a system engineering and validation perspective, the retrieval corpus constitutes a critical external data dependency. Consequently, the quality, currency, and provenance of the vector database must be rigorously version-controlled; updating the retrieval corpus constitutes a significant system-level configuration change that dictates the model's factual operational envelope and demands end-to-end integration testing, even when the core model parameters remain entirely unmodified.",
"source_ref": "Inference-Time Adaptation",
"classification_tension": "",
"summary_editorial": "Retrieves relevant documents at query time and puts them in the context, so the model can use knowledge it was never trained on."
},
{
"slug": "icl",
"tech_key": "icl",
"name": "ICL",
"family": "Inference-Time Adaptation",
"is_reference_row": false,
"footnote_markers": [
"\\S"
],
"footnotes": [
{
"marker": "\\S",
"text": "Classification tensions involving these techniques are discussed in Supplementary Section S4."
}
],
"notes": [],
"d1": [
"context-injection"
],
"d2": [
"knowledge-update",
"task-specialization"
],
"d3": [
"few-demonstrations"
],
"d4": [
"session-ephemeral"
],
"d5": [
"input-output-space"
],
"d6": [
"llm",
"mllm"
],
"definition_verbatim": "ICL leverages a deployed LLM's pattern-matching capability to perform task adaptation from a small number of input-output demonstrations placed in the context window, with no gradient computation and no parameter modification [ref]. It shares a common structural mechanism—D1 (Mechanism), D4 (Persistence), D5 (Scope), D6 (Model Type)—with PE, but fundamentally differs on D3 (Data Requirements)—the inclusion of Few Demonstrations—which shifts its primary D2 (Goal) focus toward Task Specialization. From a system validation perspective, because ICL leaves the foundational model parameters entirely unchanged, it triggers no artifact-level regression testing. Instead, the specific demonstration sets act as transient input dependencies and must be documented.",
"source_ref": "Inference-Time Adaptation",
"classification_tension": "",
"summary_editorial": "Shows the model a few worked examples in the prompt, adapting it for that request only, with no training and no artifact left behind."
},
{
"slug": "ce",
"tech_key": "ce",
"name": "CE",
"family": "Inference-Time Adaptation",
"is_reference_row": false,
"footnote_markers": [],
"footnotes": [
{
"marker": "Note",
"text": "Highly flexible techniques such as CE (e.g., prompting for step-by-step rationales) and RLHF (e.g., penalizing unexplainable refusals) can dynamically adopt Explainability as a primary goal when engineered specifically to satisfy regulatory transparency mandates."
}
],
"notes": [],
"d1": [
"context-injection"
],
"d2": [
"knowledge-update",
"personalization"
],
"d3": [
"external-corpus",
"user-interaction-data"
],
"d4": [
"version-persistent"
],
"d5": [
"input-output-space"
],
"d6": [
"llm",
"mllm"
],
"definition_verbatim": "CE is the broader architectural discipline of dynamically assembling an LLM's inference context from multiple retrieved, historical, or real-time data sources—encompassing external document retrieval [ref], agent memory [ref], user profiles, tool outputs, and conversational history [ref] (External Corpus, User Interaction Data). CE differs from PE in the structural locus of design: PE optimizes the content of a static input prompt, while CE designs the automated retrieval pipelines and orchestration systems that dynamically populate the model's context window at inference time. CE's External Corpus, User Interaction Data coordinate strictly distinguishes it from PE (Zero-Shot) by explicitly demanding complex, version-controlled retrieval and state-management infrastructure. Crucially, the live inference query serves solely as an execution trigger and data constraint, never as an on-the-fly parametric gradient signal (3tstOnly)—a strict architectural boundary that fundamentally separates CE from TTA. From a system engineering perspective, CE shifts the validation burden away from artifact-level model testing and toward complex state-machine verification. Because the model's operational envelope is dictated by dynamically retrieved external state, regression testing must rigorously validate the orchestration logic, context truncation algorithms, and data pipeline integrity rather than the static underlying model weights.",
"source_ref": "Inference-Time Adaptation",
"classification_tension": "",
"summary_editorial": "Manages everything that goes into the model's context — retrieved documents, tools, history — as a maintained, versioned production asset."
},
{
"slug": "tta",
"tech_key": "tta",
"name": "TTA",
"family": "Inference-Time Adaptation",
"is_reference_row": false,
"footnote_markers": [],
"footnotes": [],
"notes": [],
"d1": [
"parametric-update"
],
"d2": [
"drift-remediation"
],
"d3": [
"test-only"
],
"d4": [
"transient"
],
"d5": [
"partial",
"whole-model"
],
"d6": [
"dl",
"fm"
],
"definition_verbatim": "TTA updates model parameters (Parametric Update) transiently during inference on individual test examples or execution batches (Test-Only), remediating distribution shift (Drift Remediation), systematically reverting those gradient updates before processing the next prediction (Transient) [ref]. TTA is strictly classified at a DL, FM ceiling because executing transient parameter updates via backpropagation is computationally prohibitive at the LLM scale; at the LLM tier, prompt-based adaptation, e.g., PE, ICL, or Activation Steering serves the functionally equivalent role without executing parameter modification. To maintain terminological precision, TTA must be distinguished from TTT [ref]. While both adapt parameters at inference, TTA typically relies on unsupervised entropy minimization to counter distribution shift, whereas TTT executes transient gradient updates against a distinct self-supervised objective formulated directly on the test instance. From a system validation perspective, TTA's transient modification profile technically bypasses the need for persistent, artifact-level regression testing, as the base artifact theoretically remains unaltered. However, the engineering and quality-assurance burden shifts heavily to execution-state management: the inference pipeline must mathematically guarantee and continuously verify the perfect restoration of the base model weights between inference episodes to prevent insidious cross-inference parameter leakage or progressive state corruption.",
"source_ref": "Inference-Time Adaptation",
"classification_tension": "",
"summary_editorial": "Temporarily updates the model's weights at inference using the test input itself, usually to absorb a distribution shift, then reverts them before the next prediction."
},
{
"slug": "ttcompute",
"tech_key": "ttCompute",
"name": "Test-Time Compute Scaling",
"family": "Inference-Time Adaptation",
"is_reference_row": false,
"footnote_markers": [],
"footnotes": [],
"notes": [],
"d1": [
"inference-time-search"
],
"d2": [
"reasoning"
],
"d3": [
"zero-shot"
],
"d4": [
"session-ephemeral"
],
"d5": [
"input-output-space"
],
"d6": [
"llm",
"mllm"
],
"definition_verbatim": "Test-Time Compute Scaling dynamically allocates increased inference-time computation budget (Inference-Time Search) to systematically improve reasoning quality (Reasoning) without modifying any underlying model parameters (Input/Output-Space) [ref]. Unlike TTA, it executes absolutely no gradient updates; unlike PE, its performance gain derives mechanically from internal search algorithms and multi-step generation topologies (Zero-Shot) rather than explicit input phrasing. Its non-parametric D1 (Mechanism) classification and Session-Ephemeral coordinate confirm that it leaves the model architecture mathematically unchanged. Test-time compute scaling serves as the foundational mechanism driving modern reasoning models' characteristic ``extended thinking'' behavior: by generating multiple candidate solutions, applying reward-model reranking, or algorithmically exploring a tree of intermediate reasoning steps, the system intentionally trades inference-time compute latency for output quality [ref]. From a system engineering and validation perspective, test-time compute scaling is strictly an execution strategy rather than an artifact modification. Because the base model weights remain untouched, it bypasses parameter-level regression testing. However, because dynamic search algorithms fundamentally alter the determinism, upper latency bounds, and operational behavioral envelope of the system, this strategy demands rigorous validation through execution-state stress testing.",
"source_ref": "Inference-Time Adaptation",
"classification_tension": "",
"summary_editorial": "Spends more computation at answer time — sampling more, searching further, reasoning longer — without changing a single weight."
},
{
"slug": "calib",
"tech_key": "calib",
"name": "Calibration",
"family": "Calibration, Personalization, and Multimodal Adaptation",
"is_reference_row": false,
"footnote_markers": [],
"footnotes": [],
"notes": [],
"d1": [
"parametric-update"
],
"d2": [
"reliability"
],
"d3": [
"small-labeled",
"unlabeled"
],
"d4": [
"ad-hoc-permanent",
"scheduled-permanent"
],
"d5": [
"modular"
],
"d6": [
"dl",
"fm",
"llm",
"ml",
"mllm"
],
"definition_verbatim": "Calibration post-hoc aligns a model's output confidence scores with its empirical accuracy, correcting the systematic overconfidence or underconfidence produced by standard training objectives that optimize for discriminative accuracy rather than probabilistic faithfulness [ref]. Calibration does not improve the model's task accuracy, it improves the correspondence between the model's stated confidence and its actual accuracy (Reliability). From a system engineering and safety validation perspective, calibration operates not merely as a performance optimization, but as a critical operational safety constraint. Systems integrated into automated pipelines must mathematically bound their uncertainty; uncalibrated confidence scores inherently invalidate downstream threshold-based decision logic and fail-safe routing, mandating rigorous statistical auditing of calibration metrics.",
"source_ref": "Calibration, Personalization, and Multimodal Adaptation",
"classification_tension": "",
"summary_editorial": "Adjusts the model's confidence so its stated probabilities match how often it is actually right, leaving its accuracy unchanged."
},
{
"slug": "crossmodel",
"tech_key": "crossModel",
"name": "Cross-Modal Alignment",
"family": "Calibration, Personalization, and Multimodal Adaptation",
"is_reference_row": false,
"footnote_markers": [],
"footnotes": [],
"notes": [],
"d1": [
"parametric-update"
],
"d2": [
"capability-extension"
],
"d3": [
"paired-multimodal"
],
"d4": [
"ad-hoc-permanent",
"scheduled-permanent"
],
"d5": [
"partial"
],
"d6": [
"fm",
"mllm"
],
"definition_verbatim": "Cross-Modal Alignment trains a model to map representations from divergent data modalities—most commonly vision and language—into a unified, shared embedding space using contrastive optimization objectives [ref]. Its Parametric Update classification applies to the alignment bridge—the projection network [ref], cross-attention module [ref], or contrastive alignment layer [ref]—while the modality-specific encoders may be frozen. Cross-Modal Alignment requires aligned pairs (Paired Multimodal) from two modalities—image-caption pairs, video-transcript pairs, audio-text pairs—where the pairing reflects semantic correspondence across modalities, not merely co-occurrence. The Capability Extension coordinate reflects that Cross-Modal Alignment enables capabilities that are structurally impossible in unimodal models, capabilities that require the bridge, not merely improved unimodal encoders. Its D6 (Model Type) classification at FM reflects the immense architectural scale required: contrastive training across massive, noisy paired corpora is computationally feasible only at the FM and MLLM scale. Since the modification is concentrated in the alignment bridge (Partial), any change to the bridge must be treated as a capability-level modification requiring full re-characterization of the system's multimodal behavioral envelope.",
"source_ref": "Calibration, Personalization, and Multimodal Adaptation",
"classification_tension": "",
"summary_editorial": "Trains image and text representations into a shared space so the model can relate what it sees to what it reads."
},
{
"slug": "modeladapter",
"tech_key": "modelAdapter",
"name": "Modality-Spec. Adapters",
"family": "Calibration, Personalization, and Multimodal Adaptation",
"is_reference_row": false,
"footnote_markers": [],
"footnotes": [],
"notes": [],
"d1": [
"parametric-update"
],
"d2": [
"capability-extension",
"computational-efficiency"
],
"d3": [
"paired-multimodal"
],
"d4": [
"ad-hoc-permanent",
"scheduled-permanent"
],
"d5": [
"modular"
],
"d6": [
"mllm"
],
"definition_verbatim": "Modality-Specific Adapters extend a frozen multi-modal model's capabilities to new modalities or improve its performance on existing ones by inserting lightweight PEFT-style (Parametric Update) modules (Modular) within each modality's dedicated processing pathway, without modifying the shared model backbone or cross-modal bridge [ref]. In contrast to Cross-Modal Alignment, which optimizes the interface between representations to enable semantic integration, Modality-Specific Adapters refine within-modality processing, making them complementary adaptation layers rather than alternatives. Compared to standard PEFT, which applies within a single modality's processing pathway, Modality-Specific Adapters are architecturally instantiated per modality, creating a collection of independently versioned parametric modules whose composition defines the system's active multi-modal capability profile. Their classification strictly at MLLM reflects the architectural prerequisite: independent per-modality adapter injection inherently requires a multi-stream architecture featuring distinct, unimodal computational pathways. The Modular ensures that modality-specific adaptation artifacts can be independently versioned, tested, and rolled back. From a system engineering and validation standpoint, this structural isolation profoundly simplifies configuration management: it mathematically guarantees that high-frequency updates to one modality's adapter (e.g., integrating a new visual encoding standard) can be executed and validated completely independent of the other (e.g., text generation), systematically bypassing the computationally expensive need for end-to-end, multi-stream regression testing.",
"source_ref": "Calibration, Personalization, and Multimodal Adaptation",
"classification_tension": "",
"summary_editorial": "Inserts a small trainable module into one modality's own pathway inside a multimodal model, improving that input type without retraining the shared backbone."
},
{
"slug": "multimodinstr",
"tech_key": "multimodInstr",
"name": "Multimodal Instr. Tuning",
"family": "Calibration, Personalization, and Multimodal Adaptation",
"is_reference_row": false,
"footnote_markers": [],
"footnotes": [],
"notes": [],
"d1": [
"parametric-update"
],
"d2": [
"alignment",
"capability-extension"
],
"d3": [
"paired-multimodal"
],
"d4": [
"ad-hoc-permanent",
"scheduled-permanent"
],
"d5": [
"partial",
"whole-model"
],
"d6": [
"mllm"
],
"definition_verbatim": "Multimodal Instruction Tuning fine-tunes (Parametric Update) a multimodal model on multimodal instruction-response pairs (Paired Multimodal) teaching the model to follow natural-language instructions [ref]. Its D5 (Scope) varies by implementation (Partial, Whole-Model). The D3 (Data Requirements) coordinate of Paired Multimodal establishes a highly distinctive data-provenance dependency: unlike text-only instruction tuning, multimodal datasets require perfectly synchronized, temporally aligned instances across modalities. This severely raises the bar for annotation quality pipelines and introduces strict cross-modal consistency as a mandatory validation criterion. From an engineering and system integration perspective, this technique constitutes a high-impact, Partial, Whole-Model modification event that fundamentally alters the model's core behavioral envelope across all supported modalities simultaneously. Consequently, it mandates exhaustive, cross-modal regression testing to mathematically guarantee that optimization and capability alignment in one specific modality do not destructively interfere with or degrade baseline performance in the others.",
"source_ref": "Calibration, Personalization, and Multimodal Adaptation",
"classification_tension": "",
"summary_editorial": "Instruction-tunes on paired image-and-text examples so a multimodal model follows directions about what it is shown."
},
{
"slug": "munlrn",
"tech_key": "mUnLrn",
"name": "Machine Unlearning",
"family": "Knowledge Modification and Activation-Based Adaptation",
"is_reference_row": false,
"footnote_markers": [],
"footnotes": [],
"notes": [],
"d1": [
"parameter-composition",
"parametric-update"
],
"d2": [
"privacy-preservation",
"remove-knowledge"
],
"d3": [
"forget-set-specification"
],
"d4": [
"ad-hoc-permanent"
],
"d5": [
"fused-composition",
"partial"
],
"d6": [
"dl",
"fm",
"llm",
"mllm"
],
"definition_verbatim": "Machine Unlearning selectively erases learned knowledge, behavioral patterns, or training-data influence (Remove Knowledge) from a deployed model without retraining from scratch, addressing privacy deletion requirements, copyright compliance, and safety-critical knowledge removal (Privacy Preservation) [ref]. Compared to LEACE, which removes a concept's linear influence from activations by modifying the weight matrix's projection geometry (Parameter-Only, applied to a concept direction), Machine Unlearning targets a broader knowledge construct—facts, training data points, behavioral patterns. Its D1 (Mechanism) coordinate spans two distinct mechanistic profiles that must be rigorously distinguished within system configuration management and regression tracking. Gradient-based approaches apply gradient ascent on the target forget set (Forget Set Specification) to mathematically reverse the historical learning signal, operating as Parametric Update with a Partial scope. Conversely, task-vector negation approaches algebraically subtract a localized task vector from the base model's weights to structurally strip out a target capability, operating as Parameter Composition with Fused Composition—mechanistically identical to Task Arithmetic executed with a negation operator. Consequently, they must be governed by the exact same change-control validation logic and interference testing suites. Adversarial membership inference attacks, targeted behavioral probes, and jailbreak-style elicitation of the supposedly forgotten knowledge constitute some available verification tools, and none provides a formal completeness guarantee; any system deploying Machine Unlearning as a compliance mechanism must document this verification gap explicitly and specify the probe suite used to assess forgetting fidelity [ref].",
"source_ref": "Knowledge Modification and Activation-Based Adaptation",
"classification_tension": "",
"summary_editorial": "Removes specific data or knowledge from a trained model, often to meet privacy obligations, aiming to leave it behaving as though it had never seen the material."
},
{
"slug": "ke",
"tech_key": "ke",
"name": "Knowledge Editing",
"family": "Knowledge Modification and Activation-Based Adaptation",
"is_reference_row": false,
"footnote_markers": [
"\\S"
],
"footnotes": [
{
"marker": "\\S",
"text": "Classification tensions involving these techniques are discussed in Supplementary Section S4."
}
],
"notes": [],
"d1": [
"parameter-composition",
"parametric-update"
],
"d2": [
"knowledge-update"
],
"d3": [
"small-labeled"
],
"d4": [
"ad-hoc-permanent"
],
"d5": [
"partial"
],
"d6": [
"llm",
"mllm"
],
"definition_verbatim": "Knowledge Editing modifies a specific factual association encoded in a trained model's parameters without retraining and ideally without disrupting the model's knowledge of logically unrelated facts [ref]. Its D1 (Mechanism) coordinate is highly implementation-dependent and must be explicitly codified in system configuration and artifact versioning pipelines. Closed-form methods treat internal MLP layers as key-value stores and analytically compute rank-one weight updates, operating strictly as Parameter Composition. Conversely, meta-learning and optimization-based methods apply iterative gradient updates focused on the target fact, operating as Parametric Update. Knowledge Editing structurally differs from Machine Unlearning primarily in operational directionality: editing corrects or injects data (Knowledge Update), while unlearning ablates it (Remove Knowledge). While both share a Ad-hoc Permanent persistence profile, their D3 (Data Requirements) differ fundamentally (Small Labeled vs.\\ Forget Set Specification). Crucially, strict sequential scalability bounds remain a limiting operational boundary that must be explicitly quantified in system capability documentation [ref].",
"source_ref": "Knowledge Modification and Activation-Based Adaptation",
"classification_tension": "Knowledge editing. Straddles two fundamentally different structural mechanisms depending on the algorithm deployed. Closed-form editors (ROME [ref], MEMIT [ref]) algebraically rewrite weight matrices (Parameter Composition), whereas gradient-based editors (MEND [ref]) rely on compute-intensive backpropagation (Parametric Update).",
"summary_editorial": "Edits an individual fact inside a model, changing one specific association while aiming to leave unrelated knowledge undisturbed."
},
{
"slug": "actsteer",
"tech_key": "actSteer",
"name": "Activation Steering",
"family": "Knowledge Modification and Activation-Based Adaptation",
"is_reference_row": false,
"footnote_markers": [],
"footnotes": [],
"notes": [],
"d1": [
"activation-space-manipulation"
],
"d2": [
"alignment",
"safety"
],
"d3": [
"small-labeled",
"unlabeled"
],
"d4": [
"transient"
],
"d5": [
"activation-space"
],
"d6": [
"llm",
"mllm"
],
"definition_verbatim": "Activation Steering controls LLM behavior by algebraically injecting computed or learned steering vectors directly into intermediate hidden activations during inference execution, strictly without modifying the underlying stored parameter weights [ref]. A notable advancement in this space leverages SAE to identify and manipulate interpretable latent features, providing a highly precise, semantically grounded source for generating these transient steering vectors [ref]. Its Activation-Space Manipulation classification occupies a mathematically unique structural position: it is mechanistically distinct from both input-layer PE (, which operates exclusively on the discrete input token space) and traditional FT (, which alters permanent weight matrices). The Transient, Activation-Space coordinate profile dictates that activation steering imposes zero permanent structural modification on the base artifact, yet enforces an immediate, highly non-linear behavioral shift during the active inference session. A profound structural insight within this framework emerges when contrasting activation steering with TTA: while both represent transient execution-time adaptations, they physically bifurcate at the architectural scale boundary. TTA executes via transient gradient-based parameter updates at DL, FM, whereas activation steering operates via direct latent-space vector injection at LLM, MLLM.",
"source_ref": "Knowledge Modification and Activation-Based Adaptation",
"classification_tension": "",
"summary_editorial": "Nudges the model's internal activations at run time to push behaviour in a chosen direction, without changing any weights and lasting only for that run."
}
]
}