> ## Documentation Index
> Fetch the complete documentation index at: https://doc.blueapi.ir/llms.txt
> Use this file to discover all available pages before exploring further.

# Research to EMEP Component Justification Map

> Table and graph mapping each foundational paper to the EMEP components it justifies, from MergeEngine to EvaluationEngine.

Every EMEP component is grounded in published research. This page maps papers to components and visualizes the research-to-engineering dependency graph.

## Paper-to-Component Table

| Paper                                            | Component(s)                                                                              | Justification                                                                       |
| ------------------------------------------------ | ----------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------- |
| Model Soups (Wortsman et al. 2022)               | [MergeEngine](/merge/merge-engine)                                                        | Weight averaging is the baseline merge operation                                    |
| Task Arithmetic (Ilharco et al. 2022)            | [MergeEngine](/merge/merge-engine), [math/task-arithmetic](/math/task-arithmetic)         | Task vector addition and subtraction are core merge primitives                      |
| TIES-Merging (Yadav et al. 2023)                 | [MergeEngine](/merge/merge-engine), [math/ties-math](/math/ties-math)                     | Three-step trimming, sign election, and disjoint merge reduce interference          |
| DARE (Yu et al. 2023)                            | [MergeEngine](/merge/merge-engine), [math/dare-math](/math/dare-math)                     | Random dropping and rescaling control merge capacity                                |
| SLERP (Shoemake 1985)                            | [MergeEngine](/merge/merge-engine), [math/slerp-math](/math/slerp-math)                   | Spherical interpolation preserves angular relationships between checkpoints         |
| Evolutionary Model Merge (Akiba et al. 2024)     | [EvolutionEngine](/evolution/evolution-engine), [CandidateGenerator](/merge/merge-engine) | Genome encoding of merge recipes and NSGA-II search justify the evolution subsystem |
| NSGA-II (Deb et al. 2002)                        | [multi-objective-optimization](/evolution/multi-objective-optimization)                   | Default Pareto-based selection for 2-4 fitness objectives                           |
| NSGA-III (Deb & Jain 2014)                       | [multi-objective-optimization](/evolution/multi-objective-optimization)                   | Reference point selection for experiments with 5+ objectives                        |
| CMA-ES (Hansen 2001)                             | [EvolutionEngine](/evolution/evolution-engine)                                            | Continuous hyperparameter search for merge ratios and scaling coefficients          |
| LoRA (Hu et al. 2021)                            | [lora](/finetuning/lora), [optional-finetuning](/finetuning/optional-finetuning)          | Low-rank adaptation produces task vectors for subsequent merging                    |
| QLoRA (Dettmers et al. 2023)                     | [qlora](/finetuning/qlora)                                                                | Resource-constrained fine-tuning enables task vector production on limited hardware |
| PEFT survey (Han et al. 2024)                    | [peft](/finetuning/peft)                                                                  | Strategy selection guidance for optional fine-tuning pipeline                       |
| Catastrophic Forgetting (McCloskey & Cohen 1989) | [catastrophic-forgetting-controls](/finetuning/catastrophic-forgetting-controls)          | Fundamental problem that merge-time regression tests must detect                    |
| EWC (Kirkpatrick et al. 2017)                    | [catastrophic-forgetting-controls](/finetuning/catastrophic-forgetting-controls)          | Parameter importance weighting as a candidate forgetting mitigation                 |
| LM Evaluation Harness (Gao et al. 2023)          | [EvaluationEngine](/evaluation/framework)                                                 | Task loading, prompt formatting, and metric computation patterns                    |
| HELM (Liang et al. 2022)                         | [benchmark-specification](/evaluation/benchmark-specification)                            | Multi-metric, multi-scenario evaluation design                                      |
| Benchmark contamination (Sainz et al. 2023)      | [benchmark-integrity](/evaluation/benchmark-integrity)                                    | Justifies strict split isolation and contamination detection                        |

## Research-to-Component Graph

```mermaid theme={null}
graph LR
    subgraph Merging["Merging Research"]
        A["Model Soups<br/>Wortsman 2022"]
        B["Task Arithmetic<br/>Ilharco 2022"]
        C["TIES<br/>Yadav 2023"]
        D["DARE<br/>Yu 2023"]
        E["SLERP<br/>Shoemake 1985"]
    end

    subgraph Evolution["Evolution Research"]
        F["Evolutionary Model Merge<br/>Akiba 2024"]
        G["NSGA-II<br/>Deb 2002"]
        H["NSGA-III<br/>Deb & Jain 2014"]
        I["CMA-ES<br/>Hansen 2001"]
    end

    subgraph PEFT["PEFT Research"]
        J["LoRA<br/>Hu 2021"]
        K["QLoRA<br/>Dettmers 2023"]
        L["PEFT Survey<br/>Han 2024"]
    end

    subgraph Forgetting["Forgetting Research"]
        M["Catastrophic Forgetting<br/>McCloskey 1989"]
        N["EWC<br/>Kirkpatrick 2017"]
    end

    subgraph Eval["Evaluation Research"]
        O["LM Eval Harness<br/>Gao 2023"]
        P["HELM<br/>Liang 2022"]
        Q["Contamination<br/>Sainz 2023"]
    end

    subgraph Components["EMEP Components"]
        R["MergeEngine"]
        S["EvolutionEngine"]
        T["EvaluationEngine"]
        U["optional-finetuning"]
        V["catastrophic-forgetting-controls"]
        W["benchmark-integrity"]
    end

    A --> R
    B --> R
    C --> R
    D --> R
    E --> R
    F --> S
    G --> S
    H --> S
    I --> S
    J --> U
    K --> U
    L --> U
    M --> V
    N --> V
    O --> T
    P --> T
    Q --> W
```

## Coverage Check

Every core component has at least one research justification:

* **MergeEngine**: 5 papers (Soups, Task Arithmetic, TIES, DARE, SLERP)
* **EvolutionEngine**: 4 papers (Evolutionary Model Merge, NSGA-II, NSGA-III, CMA-ES)
* **EvaluationEngine**: 2 papers (LM Evaluation Harness, HELM)
* **optional-finetuning**: 3 papers (LoRA, QLoRA, PEFT survey)
* **catastrophic-forgetting-controls**: 2 papers (Catastrophic Forgetting, EWC)
* **benchmark-integrity**: 1 paper (Benchmark contamination)

## Next Steps

See [Evidence Map](/research/evidence-map) for the evidence tags assigned to each paper, or [References](/research/references) for the full annotated bibliography.