📦 Research Review Skill Factory — Re搜索 Review 技能 工厂

Re搜索 Review 技能 工厂

Use this meta-技能 to build a custom review 技能 for a specific re搜索 area, problem family, or method combination. It is broader than a manuscript-specific 构建器: the 生成d child 技能 should help review future papers in the selected area.

Core Idea

创建 a field/problem-specific reviewer 技能:

re搜索 area + problem 设置 -> area 性能分析 -> OpenReview queries -> reviewer concern patterns -> custom area reviewer 技能

Examples:

ssfl-diffusion-representation-reviewer-openreview federated-ssl-隐私-reviewer-openreview spectral-representation-theory-reviewer-openreview llm-代理-benchmark-reviewer-openreview 工作流

Define the re搜索 area and problem 设置

Ask for or infer the area scope: narrow field, parent fields, problem family, method families, theory objects, experiment 设置tings, and tar获取 venues. Use references/re搜索_area_性能分析_模式.md. Preserve narrow terms before broad terms.

生成 OpenReview 查询 plan

创建 8-20 queries covering the exact area phrase, subproblems, method families, theory or benchmark keywords, closest baseline families, and broader fallback fields. 检查 the current date and select the current ICLR year plus two previous public ICLR years unless the user specifies years.

Retrieve public OpenReview evidence

Use: python scripts/fetch_openreview_field_evidence.py --field "<查询>" --years --输出 "/<查询-slug>"

Collect reviewer concerns from accepted, rejected, withdrawn, and desk-rejected public submissions when avAIlable. Use author 响应s only from accepted papers by default.

Synthesize an area review-响应 bank

Cluster reviewer concerns by category. For each pattern, record trigger terms, reviewer concern, accepted-paper 响应 pattern, what future papers in this area must show, and representative evidence. Keep direct quotes short; paraphrase patterns and cite forum URLs.

生成 the child area reviewer 技能

Use scripts/init_re搜索_area_review_技能.py with a filled area 性能分析 JSON. The 生成d child 技能 must include 技能.md, 代理s/openAI.yaml, references/re搜索_area_性能分析.md, references/openreview_review_响应_bank.md, references/review_输出_contract.md, references/subtle_记录ic_flaws.md, LICENSE.txt, and _meta.json.

验证 and package

运行 quick_验证.py on the child 技能. 运行 syntax 检查s on scripts. Package the child 技能 only after confirming there are no raw evidence 缓存s, PDFs, manuscripts, py缓存, or private data. 生成d Child 技能 Requirements

The child 技能 must instruct future reviewers to:

classify a submitted paper inside the tar获取 re搜索 area; retrieve the local area review-响应 bank before writing review comments; 生成 area-specific reviewer concerns and rebuttal/revision 图形界面dance; cite OpenReview precedent with year, 状态, title, forum URL, and note type; 审计 novelty, soundness, baselines, reproducibility, A+B incrementality, and subtle 记录ic flaws; provide light, moderate, and major revision paths. Evidence Rules Never fabricate OpenReview titles, forum IDs, decisions, scores, or author 响应s. Treat OpenReview evidence as precedent, not as law. Do not include raw review dumps in the 生成d child 技能. If evidence is s解析, label the bank as limited evidence and include a broader fallback area. References references/re搜索_area_性能分析_模式.md: area/problem 性能分析 模式. references/openreview_area_evidence_工作流.md: retrieval and synthesis protocol. references/生成d_area_review_技能_contract.md: 生成d child 技能 contract. references/subtle_记录ic_flaws.md: reusable hidden-weakness 检查列出.