novelty-scout

Novelty Scout

Purpose

Find novelty risks, missing related work, missing baselines, and role-model papers before the paper makes strong claims. This skill uses search as evidence for review; it does not decide acceptance and does not prove full-paper novelty by itself.

When To Use

• Before writing Abstract or Introduction novelty claims.
• Before Related Work finalization.
• Before paper-redteam review.
• When reviewers complain that related work is missing.

Inputs

Required:

• target venue and adjacent venues;
• year range;
• working title;
• abstract or project summary;
• contribution bullets;
• method keywords;
• task/dataset/baseline names.

Optional:

• current BibTeX;
• must-cite papers;
• user-provided PDFs;
• known direct competitors.

Primary Source

Use cspapers.org for title/abstract-level CS conference search when the target area is covered. See:

• docs/cspapers-integration.md
• workflows/cspapers-assisted-review.md
• skills/novelty-scout/query-guide.md
• examples/cspapers-hit-classification-template.md

Procedure

1. Generate query families. Use exact title, method+task, dataset+task, central abstract claim, baseline+task, technical phrase, acronym, and acronym expansion.

2. Select venue groups. Use target and adjacent venues. For ML/CV, search CVPR, ICCV, ECCV, ICLR, ICML, and NeurIPS by default.

3. Retrieve top hits. Record title, venue, year, score, index, and abstract when available.

4. Classify hits. Use: direct overlap, same task, same method, same dataset, missing baseline, related work, role model, irrelevant.

5. Compare against bibliography. Mark cited, missing, unknown, or not needed.

6. Translate into writing constraints. Recommend claim weakening, citation addition, baseline addition, or full-paper inspection.

7. Feed high-risk findings into paper-redteam.

Rubric

Hit Class	Meaning	Action
Direct overlap	Same problem and similar method/claim by title/abstract	Inspect full paper; cite/contrast; weaken novelty
Same task	Same task/benchmark with different method	Consider baseline or related work
Same method	Similar technique in different task	Clarify methodological novelty
Same dataset	Uses central dataset/benchmark	Check evaluation comparability
Missing baseline	Likely should be compared in experiments	Add baseline or justify omission
Related work	Contextual but not direct competitor	Cite if relevant
Role model	Strong venue paper to emulate	Use for structure/style only