Provide theoretical justification for NLP/LLM research methods. Identifies relevant theoretical frameworks, searches literature via paper-finder, constructs formal justifications, and maps experimental results to theoretical predictions. Use when writing a Theoretical Analysis section, seeking theory to explain why a method works, or connecting experimental results to existing theorems/bounds.
Analyzes a research method and its experimental results, identifies relevant theoretical frameworks, searches for supporting literature, constructs theoretical justifications, and generates paper-ready content.
| Variable | Source | Description |
|---|---|---|
method_description | user | Natural language description of the proposed method (mechanism, architecture, training procedure, etc.) |
experimental_results | user | Key experimental findings: main results, ablation studies, analysis plots, observed phenomena |
research_context | user (optional) | Related work, problem setting, or existing draft sections for additional context |
| Variable | Description |
|---|
theory_analysis_doc | Structured theoretical analysis document saved to project directory |
paper_sections | LaTeX-ready paragraphs for Theoretical Analysis / Justification section |
literature_pointers | List of recommended papers/theorems with citation keys and relevance notes |
experiment_theory_map | Table mapping each experimental result to its theoretical explanation |
Extract the following from user input:
Present extraction to user for confirmation before proceeding.
Consult references/nlp_llm_theory_frameworks.md and use domain knowledge to identify candidate frameworks.
For each candidate, provide:
See prompts/identify_frameworks.md for the prompt template.
Present ranked candidates to user. User selects which directions to pursue.
For each selected theoretical direction:
prompts/search_literature.mdPresent literature findings to user. Discuss which papers are most relevant.
Using the selected frameworks and found literature:
Connect method to theory: Formalize how the method relates to the theoretical framework
Build the argument: Structure as:
Assess argument strength:
See prompts/construct_justification.md for the prompt template.
Present theoretical argument to user for interactive discussion and refinement.
For each experimental result (main table, ablation, analysis):
| Experimental Result | Theoretical Prediction | Correspondence | Notes |
|---|---|---|---|
| (result from paper) | (what theory predicts) | confirms / partially confirms / no prediction / contradicts | (explanation) |
Additionally identify:
See prompts/experiment_theory_mapping.md for the prompt template.
Produce three deliverables:
Theory analysis document (theory_analysis.md): Full structured analysis saved to Publication/paper/ or project root
Paper-ready LaTeX sections via prompts/generate_section.md:
Interactive refinement: User can ask follow-up questions, request alternative arguments, adjust formality level, or explore different theoretical angles
For methods with clean mathematical properties (bounds, convergence proofs, algebraic identities):
See references/lean_guide.md for Lean 4 integration details.
This phase is optional and experimental. It is most useful for:
It is NOT recommended for:
This skill operates as an interactive theory advisor. At each phase:
The user controls the depth: they can stop at framework identification (Phase 2) for a quick sanity check, or go all the way to paper-ready sections (Phase 6).