light-idea-generation

Innovation and Idea Generation

Prerequisites

Confirm two things before starting: (1) whether the literature gap in m01 is clear; (2) whether the data in m02 is sufficient to support it. If the data is insufficient, go back to m02 first and do not come up with fantasy ideas.

Inputs

Project background, existing foundations, data conditions, tech stack and compute, time cycle, goals (top-tier journal/general journal/competition/project/engineering), and constraints.

First determine which level the input belongs to (borrowing the two-level abstraction from AI-Researcher):

• Level 1: A clear direction already exists: the user has given a specific idea → focus on refinement, differentiation, and feasibility verification.
• Level 2: Only a direction/data/references(e.g., "what can be done with this direction") → infer ideas from the literature + data. When initiating at Level 2, for each candidate first fill out aproject initiation card(template templates/idea_card.md, borrowing the idea of the four-piece set from AI Scientist v2 ideation), converging vague ideas into units that can be reviewed field by field by m04. The fields of the card align with what m04 review requires: idea_id / title / one-sentence mechanism / novelty claim / nearest-neighbor work (≥3, with search trace) / differences from them / risks and failure conditions / minimal validation experiment / data feasibility / compute and cost estimate. Multiple cards are aggregated into the standard artifact idea_candidates.md。

Generation Strategy (diverge from multiple angles, then converge)

Diverge: independently generate candidates from multiple angles, avoiding a single line of thinking

1. gap-driven: directly target the literature gap.
2. method-transfer: transfer a mature method from field A to field B.
3. data-driven: infer opportunities from the unique characteristics of the data (exclusive data/new modalities/new annotations). Check the data_type/known biases/license of the db04 dataset card to find data angles others have not used.
4. problem-reframe: redefine the problem or the evaluation method.
5. combination: combine methods, but the mechanism by which 1+1>2 must be explained, not just stacking.
6. theory-gap: supplement theoretical explanation/interpretability/generalization guarantees.
7. efficiency: faster/cheaper/smaller, engineering value.

When divergence is insufficient, additionally use 7 structured stimulation techniques (derived from Scientific Brainstorming): cross-domain analogy, hypothesis reversal ("what if it were reversed/what if resources were unlimited"), scale switching (molecule↔population, millisecond↔millennium), adding or removing constraints, interdisciplinary fusion, and technological extrapolation (what can be done when new technology arrives). For the method-transfer/combination angle, one can imitate ResearchAgent: first extract the core entities of the project's domain, then find neighboring concepts with high co-occurrence with them as sources for transfer/recombination; MAGenIdeas proves that this kind of cross-domain knowledge recombination + iterative retrieval can increase the number of uniquely novel ideas to about 3.4 times.

Converge: filter and rank after divergence

• First use db03 method maturity to filter out directions that are "outdated/already overdone."
• When there are many candidates, usepairwise PK / Swiss roundsto rank and select Top-N (the MAGenIdeas approach); this can widen the gaps better than isolated scoring.
• Use a two-dimensional quick evaluation of "impact × effort" as a convergence funnel, and select Top 3 (the diverge-converge approach).

Novelty verification (don't rely on memory, go look it up)

Before proposing "innovation points / relative to which works," actually search for benchmark works, to avoid "assuming it's new" and citation hallucination.For retrieval, uniformly call the verified scripts of m01(light-literature-search); do not hand-assemble API URLs(hand-assembling URLs easily hits rate limiting/pagination/encoding pitfalls, and becomes disconnected from the single source of truth of m01):

• python light-literature-search/scripts/search_normalize.py: multi-source retrieval and normalization, with backward citation edges included in the first round (OpenAlex / Crossref)—use core mechanism keywords for an initial screening of benchmark works.
• python light-literature-search/scripts/snowball.py: for the one or two most similar papers, do forward citations + backward references "snowballing" to capture the nearest-neighbor works completely.
• Whether OpenAlex requires a key, rate limiting, billing, backoff—thesingle source of truthsee the m01 references "OpenAlex access source of truth" section; this skill does not restate the numbers. Semantic Scholar / arXiv serve as cross-validation sources; endpoints and pitfalls see this skill's references.md。

The retrieval conclusion determines the wording of "innovation points": if highly similar work already exists, go back to divergence and reselect an angle; don't insist it's new.

• Anti-fake diversity among candidates (quantified): when generating multiple candidate ideas at once, compute pairwise semantic similarity to prevent the fake diversity of "reskinning to pad the count"—reuse the method in the m01(light-literature-search) references "SPECTER2 semantic embeddings" section (S2 embedding.specter_v2 768-dimensional vectors for cosine). SPECTER2 cosine values are generally high (measured 0.85~0.93); judge byrelative difference: among a batch of candidates, those whose similarity is significantly higher than the remaining pairs are regarded as variants of the same idea; merge them or re-diverge, don't count them as independent candidates to pad the count. When there is no embedding, degrade to title/abstract text similarity.

Core collision check (highest priority, one-vote veto)

Hard-earned lesson: I once rated an idea of "resampling breaking probability calibration" at novelty ≈70; only after doing the full set of experiments and finishing the paper did I find by searching that Dal Pozzolo 2015 had specifically done the same core, and the novelty was actually 35-45—the core conclusion had been published by predecessors, and submission would surely be rejected with a single sentence "already done." To eradicate this kind of "only discovering the collision after finishing," when proposing an ideamustanswer the following four questions; if any is answered wrong, send it back and reselect the direction,no skipping, no fobbing off with "probably nobody has done it"：

1. Has anyone done the same core? Use the idea'score mechanism/core conclusionas keywords for retrieval (not vague domain words), with at least 2 databases cross-checked.Specifically look for "the most similar paper,"rather than looking for evidence that "proves I'm new"—search with the mindset of "assume someone has already done it, go dig it out." After finding it, compare sentence by sentence: whether its core claim vs my core claim is substantively equivalent.
2. Is what we're doing the current real gap? Clearly state whether this gap is "predecessors didn't do it" or "predecessors did it but with clear shortcomings." For the former, provide negative evidence of "searched X keyword combinations, Y databases, all with no hits"; for the latter, name that prior work, explain its specific shortcomings, and how we fill them.
3. Is there really an innovation point (rather than a reproduction of a known phenomenon / changing the dataset / changing the model)? Honestly distinguish three tiers: ① new phenomenon/new method/new theory (true innovation); ② systematization/quantification/extension of a known phenomenon (incremental; must explicitly say it is incremental); ③ reproduction by merely changing the dataset/model (basically no novelty, unless the reproduction itself reveals an anomaly). Honestly file the idea into one of these tiers.
4. What reasons would reviewers use to reject it? Pre-run the top-3 rejection reasons (most common: a. "the core has already been done by XXX"; b. "the increment is too small / no methodological contribution"; c. "the gap is a pseudo-gap, nobody does it because it's not important"). Each one must either be refutable or be acknowledged and the positioning adjusted. If you can't produce a strong rebuttal, the idea isn't strong enough.

Write the search evidence for these four questions (keywords, databases, HTTP codes, the 3 most similar papers, quantified delta) into the project initiation card,and hand it directly to m04 for review—m04 will re-check the core collision; if the idea's self-report doesn't match reality, it gets a red flag and a score cap.

Every idea must state clearly (none may be missing)

• One-sentence idea。
• Why it's worth doing: motivation + practical/academic significance.
• Innovation point: relative to whichspecificworks (with the real literature found in search), and where the difference lies.
• Why it's stronger: the mechanistic hypothesis by which it may outperform existing methods; and listcompeting explanationsrather than betting on just one (the hypothesis-generation approach).
• What specific problem it solves: quantifiable, falsifiable goals and predictions.
• Data/compute feasibility: whether it can be done with existing conditions.Name specific datasets(aligned with db04, give dataset names) + scale (sample size/data volume) + annotation method, aligned with the criteria of downstream m04 critique dimension 3 (data support); avoid vague statements like "existing data should be enough" (this will be capped at 60 points and red-flagged by m04).
• Output form: paper/competition/patent/system.
• Submission tier: rough positioning as reach/steady/safety (refinement goes to m13).
• Risk: the most likely point of failure. Use the What-If Oracle's "precise IF" style—turn "what if it doesn't work" into actionable conditional statements ("if data volume < N then X fails"), and think through best/most likely/worst/second-order consequences.

Self-check before submitting to m04

Quickly evaluate each idea with the ResearchAgent five dimensions + AI Scientist three dimensions; send low scorers back for rework:

• Five dimensions(1–5): Clarity, Relevance, Originality, Feasibility, Significance.
• Three-dimension quick evaluation(1–10): Interestingness, Feasibility, Novelty (score cautiously and realistically).
• 7 failure modes reverse self-check(derived from the Nature version limitations of The AI Scientist, compiled by ARS): implementation bugs, hallucinated results, taking shortcuts, treating bugs as insights, methodological fabrication, frame-lock (being locked into a single framework), citation hallucination. Focus on preventing frame-lock (don't stubbornly stick to one line of thinking) and citation hallucination (don't fabricate benchmark works).
• Optional multi-perspective mutual critique (Consciousness Council): have the "benchmark faction / feasibility faction / novelty faction / engineering faction" each poke holes, then synthesize.

Output

3–6 layered ideas (high-risk high-reward / steady / safety), ranked by potential, with a comparison table. Label each with maturity and differentiation strength, and include the self-check five-dimension/three-dimension scores.Standard artifact:idea_candidates.md(the handoff artifact to m04; for naming see CONVENTIONS §6.1).

Mandatory linkage

All ideas Mustsend to m04 idea-critique for strict review. Rejected ideas return to this skill with the direction given by m04 to be regenerated, forming a loop. Only ideas that pass proceed to m05. Write to the decision_log of project library db09.

---

For per-tool notes on method sources and real endpoints/review dimensions, see references.md。