EMEP Component Mapping: Research Dependencies

This page provides the reverse view of the Evidence Map: for each EMEP component, it lists the research papers, methods, and theoretical foundations that the component depends on. This mapping enables reviewers to verify that every component has a grounded scientific or engineering basis, and to identify gaps where a component lacks supporting evidence.

Component-to-Research Table

Component	Research Dependencies	Evidence Type	Notes
ModelRegistry	Version control theory; software artifact management	Engineering assumption	No single foundational paper; depends on MLOps best practices
ModelLoader	Hugging Face Transformers loading patterns; safetensors format	Engineering assumption	Implementation pattern, not research method
ModelCompatibilityAnalyzer	Architecture comparison literature; tensor shape validation algorithms	Engineering assumption	Core logic is structural comparison, not ML-specific
TensorEngine	Task Arithmetic: Ilharco et al. 2022; high-dimensional linear algebra	Established research	Task vector arithmetic is the foundational operation
MergeEngine	Model Soups: Wortsman et al. 2022; Task Arithmetic: Ilharco et al. 2022; SLERP: Shoemake 1985; TIES-Merging: Yadav et al. 2023; DARE: Yu et al. 2023; Evolutionary Model Merge: Akiba et al. 2024	Established research	Implements multiple strategies; each has independent evidence
MergeStrategy	Linear Merge: Model Soups; SLERP: Shoemake 1985; TIES: Yadav et al. 2023; DARE: Yu et al. 2023; Task Arithmetic: Ilharco et al. 2022; Franken-Merge: structural merging heuristics	Established research / Research hypothesis	Franken-Merge is less established than parameter-space methods
CandidateGenerator	Evolutionary Model Merge: Akiba et al. 2024; software engineering patterns	Engineering assumption	Candidate packaging is an engineering concern
EvolutionEngine	Evolutionary Model Merge: Akiba et al. 2024; genetic algorithms literature; multi-objective optimization theory	Established research / Research hypothesis	Core evolutionary mechanics are well-established; LLM-specific application is newer
EvaluationEngine	HELM: Liang et al. 2022; MMLU: Hendrycks et al. 2020; benchmark design literature	Established research	Framework design depends on general benchmarking theory
BenchmarkEngine	HELM: Liang et al. 2022; MMLU: Hendrycks et al. 2020; Big-Bench; other task-specific benchmarks	Established research	Individual benchmarks have independent evidence bases
FitnessEngine	Multi-objective optimization theory; Pareto dominance literature; HELM trade-off analysis: Liang et al. 2022	Established research / Research hypothesis	Pareto selection is established; scalar fitness for LLMs is a hypothesis
ExperimentTracker	Reproducibility literature: Gundersen and Kjensmo 2018; ML experiment management systems	Engineering assumption	Tracking is an engineering practice, not a research method
ArtifactStore	Content-addressable storage theory; cryptographic hashing standards	Engineering assumption	Hash-based integrity is a security best practice
DatasetRegistry	Data versioning literature; benchmark contamination detection: Jacovi, Goldberg, and Tsarfaty 2023	Established research / Engineering assumption	Contamination detection is research; registry design is engineering
QuantizationEngine	GPTQ: Frantar et al. 2022; AWQ: Lin et al. 2023; quantization-aware training literature	Established research	Multiple quantization methods with independent evidence
InferenceBackend	vLLM; TGI; llama.cpp; serving system literature	Engineering assumption	Backend selection is an implementation choice
DeploymentManager	Kubernetes; MLOps deployment patterns; blue-green deployment literature	Engineering assumption	Deployment orchestration is an engineering practice

Dependency Diagram

The following diagram shows how research areas feed into EMEP components. Solid arrows indicate direct implementation dependency. Dashed arrows indicate theoretical inspiration or optional integration.

Gap Analysis

The following components have weaker research foundations and require additional validation:

Component	Gap	Mitigation
ModelCompatibilityAnalyzer	No single paper defines LLM compatibility	Engineering assumption; validation via Integration Tests
CandidateGenerator	Candidate packaging is not a research topic	Engineering assumption; acceptance criteria in Unit Tests
ExperimentTracker	Experiment tracking is practice, not theory	Engineering assumption; aligned with MLflow / Weights & Biases patterns
InferenceBackend	Backend choice is implementation-specific	Engineering assumption; health checks in Performance Validation
DeploymentManager	Deployment orchestration is generic software engineering	Engineering assumption; Runbook defines operational procedures

Not every component requires a peer-reviewed paper. Components that manage data, orchestrate workflows, or interface with hardware are grounded in software engineering best practices. These are tagged as Engineering Assumption and validated through testing and operational review, not scientific experiment.

Evidence Map for the forward mapping (claims to sources)
References for the full annotated bibliography
Decisions for architecture decisions that depend on this mapping
Testing Strategy for how gaps are validated through tests

Project Context

Research Foundation

System Maps

EMEP Component Mapping: Research Dependencies

Component-to-Research Table

Dependency Diagram

Gap Analysis

​Component-to-Research Table

​Dependency Diagram

​Gap Analysis

​Related Pages

Component-to-Research Table

Dependency Diagram

Gap Analysis

Related Pages