research-evidence

Research Evidence

Overview

Use this skill as the shared evidence layer for paper writing, related work, review, and rebuttal work. It does not replace domain judgment: it finds and checks evidence, then reports support level, gaps, and risks.

Hard Boundaries

• Use public metadata, open-access PDFs, and user-provided local PDFs only.
• Do not use Sci-Hub or any route that bypasses access control.
• Do not write Markdown, JSON, CSV, or report files unless the user explicitly asks for saved artifacts.
• Run RefChecker in non-LLM mode by default. Enable LLM-assisted extraction or hallucination checks only when the user explicitly requests it and accepts the privacy/cost tradeoff.
• Treat tool output as evidence to inspect, not final truth. Mark weak coverage, metadata mismatches, and unsupported claims clearly.

Venue Defaults

Prefer these venues when the user does not specify otherwise:

• Computer Vision: CVPR, ICCV, ECCV.
• ML/AI: ICLR, NeurIPS/NIPS, ICML.
• Autonomous driving, robotics, and collaborative perception secondary venues: CoRL, ICRA, IROS, AAAI, IJCAI, T-ITS, RA-L.

Do not let secondary venues outrank the primary top-conference set unless the user's task or subfield demands it.

Workflows

Literature Search

1. Translate the research question into compact queries with venue/year terms when useful.
2. Use the dedicated tool environment from references/tooling.md; start with arXiv metadata search for CV/ML topics.
3. Return a compact candidate list with title, source, year/date, identifier, URL, and why each candidate is relevant.
4. Label coverage risk when search sources are unavailable, too broad, too recent, or missing known top-venue work.

Recent Literature / Novelty Risk Audit

Use this workflow when the user asks about novelty, recent work, missing related work, reviewer risk, rebuttal readiness, score prediction, first/SOTA/new benchmark claims, or whether a close paper conflicts with the current claim.

1. Decompose the claim into axes before searching: domain/input, method family, task/output, evaluation protocol, dataset/benchmark, and reliability or failure mode.
2. Expand the search across direct terms, synonyms, neighboring tasks, older terminology, and broad recent-work queries. Do not stop at the user's exact phrasing.
3. For high-stakes checks, cross-check at least two evidence routes: direct arXiv API or official venue metadata; local reference/, PDF, BibTeX, or LaTeX sources; and the local paper-search wrapper as a smoke or secondary check.
4. Classify candidates after inspecting metadata and abstracts: direct competitor, claim limiter, table candidate, prose citation, background, or irrelevant.
5. Report the search matrix, source-coverage limits, and claim-impact verdict. Never answer "not found" without stating which query families and sources were checked.