Generate, score, and rank novel CS/ML research hypotheses based on identified literature gaps. Produces 3–5 concrete, differentiated ideas with preliminary novelty assessments.
This skill takes a literature landscape map (from the literature-review skill or a direct description of the field) and generates concrete, original research hypotheses ranked by novelty and feasibility.
Use this skill when:
You should have:
literature-review skill (landscape map + gaps)If no prior literature review is available, the orchestrator will run a quick survey first.
Parse the provided landscape map or gap analysis. Extract:
For each gap, generate 1–2 research hypotheses using this template:
**Hypothesis [N]**: [One-sentence claim]
- **Problem**: [What is missing or broken]
- **Proposed approach**: [Core idea in 2–3 sentences]
- **Motivation**: [Why this might work — theoretical or empirical grounding]
- **Distinguishing factor**: [What makes this different from existing work]
- **Preliminary novelty confidence**: [High / Medium / Low] — [Reason]
Diversity constraint: Hypotheses must differ in at least one of: (a) methodology family (e.g., don't generate 5 fine-tuning ideas), (b) problem framing, (c) target application domain.
For each hypothesis, run a targeted arXiv search to check if an obvious direct solution exists.
Use search_arxiv_papers (via AlphaXiv MCP) with the core claim as the query. This is a quick check — deep validation happens in the idea-critique skill.
Example: For "hypothesis: sparse attention patterns can be learned from dense attention", search:
search_arxiv_papers(query="learning sparse attention patterns", num_results=10)
Score each surviving hypothesis on:
| Criterion | Weight | Description |
|---|---|---|
| Novelty | 40% | How differentiated from existing work |
| Feasibility | 30% | Can be done in 6–12 months with standard compute |
| Significance | 20% | Would matter if it worked |
| Clarity | 10% | Is the claim testable and well-defined |
Rank by weighted score. Present top 3–5.
# Research Hypotheses: [Topic]
Generated: [date]
Based on: [literature source]
## Ranked Hypotheses
### 🥇 Hypothesis 1: [Title]
**Score**: Novelty X/10 | Feasibility X/10 | Significance X/10 | Overall X/10
**Problem**: ...
**Approach**: ...
**Why novel**: ...
**Key risk**: ...
**Recommended next step**: Run `idea-critique` skill to validate novelty
### 🥈 Hypothesis 2: [Title]
...
Save the top 3 hypotheses to the idea store using idea_store with action save (status: draft).