How the taxonomy is shaped

The model-tier ladder

Moving down the hierarchy adds techniques without removing them — except where a technique is superseded by a more general alternative. Supersession is not binary: a superseded technique often remains technically available.

ML · Machine Learning

8 techniques available

New at this tier (8)
Training, Retraining, Curriculum Learning, Active Learning, Data Augmentation, Semi-SL, FL, Calibration

DL · Deep Learning

28 techniques available

New at this tier (20)
FT (full), FT (partial), PEFT (LoRA, adapters), DA, FSL, CL, TIL, DIL, LP, LEACE, Adversarial Training, Meta-Learning, MTL, Self-Play, SSL / CPT, DP-FT, KD, Model Compression, TTA, Machine Unlearning

FM · Foundation Model

32 techniques available

New at this tier (5)
Task Arith. & Model Merging, Long-Context Ext., MoE, Prompt Learning, Cross-Modal Alignment

LLM · Large Language Model

40 techniques available

New at this tier (14)
ReFT, SFT, RLHF (LLM Alignment), DPO, RLAIF/CAI, RLVR/GRPO, PE, APO, RAG, ICL, CE, Test-Time Compute Scaling, Knowledge Editing, Activation Steering

MLLM · Multimodal Large Language Model

41 techniques available

New at this tier (2)
Modality-Spec. Adapters, Multimodal Instr. Tuning

Computed from D6 membership and curated supersession relations

Do the families hold together?

Not really — and the paper says so. Measuring how well the nine navigational families separate in profile space gives a silhouette of +0.0173 across all rows, or +0.0566 over the post-training techniques alone. Both are near zero; anything below about 0.2 indicates no meaningful clustering.

This is the expected result. The families are navigational groupings to aid discovery, not validated statistical clusters. Techniques share one anchor dimension while differing on others, which is exactly why a single axis cannot hold them.

Computed from the taxonomy data

Similarity projection

Training — Reference BaselineFT (full) — Knowledge Transfer and Task SpecializationFT (partial) — Knowledge Transfer and Task SpecializationPEFT (LoRA, adapters) — Knowledge Transfer and Task SpecializationReFT — Knowledge Transfer and Task SpecializationDA — Knowledge Transfer and Task SpecializationFSL — Knowledge Transfer and Task SpecializationRetraining — Temporal Adaptation and MaintenanceCL — Temporal Adaptation and MaintenanceTIL — Temporal Adaptation and MaintenanceDIL — Temporal Adaptation and MaintenanceSFT — Alignment, Reasoning, and TrustworthinessRLHF (LLM Alignment) — Alignment, Reasoning, and TrustworthinessDPO — Alignment, Reasoning, and TrustworthinessRLAIF/CAI — Alignment, Reasoning, and TrustworthinessRLVR/GRPO — Alignment, Reasoning, and TrustworthinessLP — Alignment, Reasoning, and TrustworthinessLEACE — Alignment, Reasoning, and TrustworthinessAdversarial Training — Alignment, Reasoning, and TrustworthinessMeta-Learning — Training StrategiesMTL — Training StrategiesSelf-Play — Training StrategiesCurriculum Learning — Training StrategiesActive Learning — Training StrategiesData Augmentation — Data-Centric and Privacy-Preserving MethodsSemi-SL — Data-Centric and Privacy-Preserving MethodsSSL / CPT — Data-Centric and Privacy-Preserving MethodsDP-FT — Data-Centric and Privacy-Preserving MethodsFL — Data-Centric and Privacy-Preserving MethodsKD — Efficiency and CompositionModel Compression — Efficiency and CompositionTask Arith. & Model Merging — Efficiency and CompositionLong-Context Ext. — Efficiency and CompositionMoE — Efficiency and CompositionPE — Inference-Time AdaptationPrompt Learning — Inference-Time AdaptationAPO — Inference-Time AdaptationRAG — Inference-Time AdaptationICL — Inference-Time AdaptationCE — Inference-Time AdaptationTTA — Inference-Time AdaptationTest-Time Compute Scaling — Inference-Time AdaptationCalibration — Calibration, Personalization, and Multimodal AdaptationCross-Modal Alignment — Calibration, Personalization, and Multimodal AdaptationModality-Spec. Adapters — Calibration, Personalization, and Multimodal AdaptationMultimodal Instr. Tuning — Calibration, Personalization, and Multimodal AdaptationMachine Unlearning — Knowledge Modification and Activation-Based AdaptationKnowledge Editing — Knowledge Modification and Activation-Based AdaptationActivation Steering — Knowledge Modification and Activation-Based Adaptation

Alignment, Reasoning, and Trustworthiness Calibration, Personalization, and Multimodal Adaptation Data-Centric and Privacy-Preserving Methods Efficiency and Composition Inference-Time Adaptation Knowledge Modification and Activation-Based Adaptation Knowledge Transfer and Task Specialization Reference Baseline Temporal Adaptation and Maintenance Training Strategies

Recomputed, not the projection printed in the paper. Built from the same distance matrix with the paper's own settings, but a different umap-learn build (0.5.12). The distances are verified identical — the raw Gower silhouette reproduces the published +0.0173 exactly — but the embedding is not, because UMAP gives a different layout on a different machine even with the random seed fixed (-0.0851 here against -0.0346 published). Read it for which techniques sit near each other, not for exact positions.
Coordinates as a table
TechniqueFamilyx, y
TrainingReference Baseline-2.81, 2.89
FT (full)Knowledge Transfer and Task Specialization-2.93, 2.52
FT (partial)Knowledge Transfer and Task Specialization-0.59, 3.57
PEFT (LoRA, adapters)Knowledge Transfer and Task Specialization0.18, 3.42
ReFTKnowledge Transfer and Task Specialization0.21, 4.13
DAKnowledge Transfer and Task Specialization-0.98, 3.88
FSLKnowledge Transfer and Task Specialization-0.82, 4.38
RetrainingTemporal Adaptation and Maintenance-2.85, 3.44
CLTemporal Adaptation and Maintenance-2.04, 3.88
TILTemporal Adaptation and Maintenance-1.96, 4.12
DILTemporal Adaptation and Maintenance-2.31, 4.30
SFTAlignment, Reasoning, and Trustworthiness4.27, 0.90
RLHF (LLM Alignment)Alignment, Reasoning, and Trustworthiness3.85, 1.30
DPOAlignment, Reasoning, and Trustworthiness4.31, 1.55
RLAIF/CAIAlignment, Reasoning, and Trustworthiness3.74, 1.05
RLVR/GRPOAlignment, Reasoning, and Trustworthiness4.50, 1.19
LPAlignment, Reasoning, and Trustworthiness0.54, 3.93
LEACEAlignment, Reasoning, and Trustworthiness-0.38, 2.90
Adversarial TrainingAlignment, Reasoning, and Trustworthiness-3.30, 2.78
Meta-LearningTraining Strategies-1.24, 4.49
MTLTraining Strategies-3.46, 3.67
Self-PlayTraining Strategies-3.31, 3.38
Curriculum LearningTraining Strategies0.55, 7.67
Active LearningTraining Strategies0.82, 7.33
Data AugmentationData-Centric and Privacy-Preserving Methods0.86, 7.35
Semi-SLData-Centric and Privacy-Preserving Methods0.41, 7.27
SSL / CPTData-Centric and Privacy-Preserving Methods-1.52, 3.48
DP-FTData-Centric and Privacy-Preserving Methods-1.62, 2.77
FLData-Centric and Privacy-Preserving Methods-2.09, 2.82
KDEfficiency and Composition1.21, 3.41
Model CompressionEfficiency and Composition1.55, 3.53
Task Arith. & Model MergingEfficiency and Composition1.31, 2.97
Long-Context Ext.Efficiency and Composition0.47, 2.75
MoEEfficiency and Composition0.75, 3.04
PEInference-Time Adaptation8.34, 0.59
Prompt LearningInference-Time Adaptation-0.27, 3.88
APOInference-Time Adaptation8.78, 0.77
RAGInference-Time Adaptation8.13, 0.89
ICLInference-Time Adaptation8.69, 0.87
CEInference-Time Adaptation8.44, 1.28
TTAInference-Time Adaptation-1.95, 4.65
Test-Time Compute ScalingInference-Time Adaptation8.50, 0.25
CalibrationCalibration, Personalization, and Multimodal Adaptation0.82, 3.81
Cross-Modal AlignmentCalibration, Personalization, and Multimodal Adaptation1.97, 2.20
Modality-Spec. AdaptersCalibration, Personalization, and Multimodal Adaptation1.89, 2.49
Multimodal Instr. TuningCalibration, Personalization, and Multimodal Adaptation2.64, 1.94
Machine UnlearningKnowledge Modification and Activation-Based Adaptation-0.73, 2.55
Knowledge EditingKnowledge Modification and Activation-Based Adaptation-0.40, 2.73
Activation SteeringKnowledge Modification and Activation-Based Adaptation3.84, 1.53

Pairwise distances are independent of the projection and are shown on every technique page under nearest profiles.