| name | cross-pollination-ideation |
| description | Generate research ideas by finding cross-disciplinary connections, retrieving obscure theorems from distant fields, and identifying analogies between different mathematical domains. Use when Raman asks for new research directions, when ingesting a paper that might connect to other fields, when looking for "bridges" between topics, or when asked to "find connections," "what other fields use this technique," or "cross-pollinate." Implements the cross-pollination techniques from Woodruff et al. (2026, Section 2.2, 4.x) and AI Behavioral Science's framework for studying AI-human strategic interactions (Jackson et al., 2025). |
Cross-Pollination Ideation
Systematic generation of research ideas by finding connections across disciplinary boundaries. Based on the cross-pollination case studies from Woodruff et al. (2026) - where linking Steiner trees to the Kirszbraun Extension Theorem and reframing combinatorics as measure theory led to resolving open problems.
When to use
- Raman asks "what ideas emerge from these papers?" or "give me new research directions"
- A new paper is ingested that uses techniques from a field adjacent to Raman's
- Raman is stuck on a problem and needs a fresh angle
- Exploring whether a technique from one of Raman's topics applies elsewhere
- Enhancing OP3 (Generate new research ideas) from CLAUDE.md
The method
Step 1: Map the technique space
For the source concept or problem, identify:
- The core mathematical structure: What is the essential abstraction? (e.g., "this is really about finding a fixed point," "this is a convex optimization under constraints," "this is about information asymmetry in a network")
- The proof/solution techniques used: What tools does the paper deploy? (spectral methods, measure theory, LP duality, potential functions, coupling arguments, etc.)
- The modeling assumptions: What simplifications does the paper make? Which could be relaxed?
Step 2: Search for analogies
For each core structure identified, search for parallels in Raman's research areas:
| Source domain | Target (Raman's domains) | Bridge question |
|---|
| Technique T in field A | Does T apply to network games? | "What if we use T to analyze equilibrium structure on networks?" |
| Model assumption X | Prospect theory relaxes X | "What happens to the result when agents have biased perceptions?" |
| Optimization method M | Strategic classification uses similar structure | "Can M give tighter bounds on strategic manipulation?" |
| Game-theoretic result R | Information design extends R | "What if the principal can design the information environment?" |
Key bridges for Raman's program:
- Spectral graph theory <-> Network games: Eigenvalues of the adjacency/Laplacian encode equilibrium structure (Ballester-Calvo-Armengol-Zenou pattern)
- Prospect theory <-> Any game-theoretic model: Replace expected utility with CPT - what changes? (Raman's "Architecture of Illusion" pattern)
- Information design <-> Strategic classification: The principal chooses what information to reveal; the agents best-respond strategically
- Mechanism design <-> Network economics: Design network interventions as mechanisms
- Non-normal matrices <-> Network dynamics: Pseudospectra capture transient behavior that normal analysis misses (Raman's non-normal network games idea)
Step 3: The parameterization test
From Raman's own methodology (the "Architecture of Illusion" pattern):
- Find the extreme cases: Is paper X the alpha=0 case? Is paper Y the alpha=1 case?
- Propose the interpolation: What happens for alpha in (0,1)?
- Check if the interpolation is non-trivial: Does the behavior change qualitatively at some critical alpha*? If so, that's a research contribution.
Examples from Raman's work:
- Risk-neutral (alpha=0) vs fully loss-averse (alpha=1) agents in network games
- Full information (alpha=0) vs no information (alpha=1) in Bayesian persuasion on networks
- Myopic (k=0) vs fully strategic (k=inf) in k-hop awareness models
Step 4: The "obscure theorem" retrieval
This is where AI has a genuine comparative advantage (Woodruff et al., 2026). For the stuck problem:
- Restate the problem abstractly: Strip away the domain-specific language. What is the pure mathematical structure?
- Search your knowledge for theorems about that abstract structure from any field:
- Functional analysis (Stone-Weierstrass, Riesz Representation, Hahn-Banach)
- Topology (Brouwer, Kakutani, Borsuk-Ulam)
- Measure theory (Radon-Nikodym, disintegration, optimal transport)
- Combinatorics (Ramsey, Turan, Lovasz Local Lemma)
- Information theory (data processing inequality, Fano's inequality)
- Check if the theorem's conditions are satisfied in Raman's setting
- If they are: Sketch how the theorem resolves the roadblock
Step 5: The Jackson et al. lens (AI Behavioral Science)
For ideas at the intersection of AI and behavioral/social science:
- AI as strategic agent: How do AI agents behave in game-theoretic settings? Can we apply behavioral game theory tools (cognitive hierarchy, level-k) to predict AI behavior?
- AI as tool for behavioral science: Can LLMs simulate participants in network games? Can they predict experimental outcomes in strategic classification settings?
- Human-AI interaction on networks: How does introducing AI agents into a network game change equilibrium behavior? What are the welfare implications?
- Information design for AI: How should a principal design information structures when some agents are AI and some are human (heterogeneous rationality)?
Step 6: Quality filter
Before proposing an idea, run these checks:
- Novelty check: Search the wiki for existing ideas that cover this ground. Search arXiv mentally for whether this has been done.
- Feasibility check: Can Raman actually prove something here, or is this just hand-waving? What would the first lemma be?
- Significance check: Does this connect to an existing open question? Does it have implications beyond a technical curiosity?
- The "why not" test: If this connection is so natural, why hasn't someone done it? (Possible answers: (a) they have and you don't know, (b) there's a technical barrier, (c) the communities don't talk to each other - only (c) is a good sign)
Output format
For each proposed idea:
### Idea: [one-sentence pitch]
**The bridge**: [Source field] -> [Raman's domain]
**The gap**: [What's missing in the literature]
**The key insight**: [Why this connection works]
**First step**: [What the first lemma or result would be]
**Papers it builds on**: [[sources/...]]
**Risk level**: Low (incremental) / Medium (novel combination) / High (speculative)
Propose 3-5 ideas, ranked by feasibility * significance.
Integration with wiki
- Offer to file promising ideas as
wiki/ideas/<slug>.md with status seed
- Update relevant topic MOCs with new cross-references
- If the session produces substantial output, file to
outputs/research-proposals/
- Log to
log.md with op type idea