Use when reranking arXiv papers with an LLM. Applies quality and soft area-mix preferences for relevance, technical depth, novelty, safety, and practical impact.
This skill defines the quality bar for LLM reranking in the daily arXiv pipeline.
LLM role scope: execute Stage-2 scoring/reranking and final metadata-based top-5 decision support; Stage-1 retrieval/filtering is handled by pipeline code.
Tunable Config Knobs
Pipeline sizing parameters are controlled in arxiv_daily_config.json.
stage1_top_per_area: number of papers taken directly from each area's arXiv retrieval order for Stage-2 input.
stage2_per_area_llm_top: number kept per area after area-level LLM scoring.
final_top: final number of papers for deep read and summary.
Reranking Requirements
Always read and use the research interest profile as the primary personalization target.
For stage-2 area-level scoring, use arXiv metadata as main evidence: title, abstract (), categories, and publication context.
相关技能
summary[:1200]
For each area, LLM-score the provided area candidate set and keep top papers by score.
Merge all per-area top-3 candidates, then perform one unified metadata LLM scoring pass and global rerank.
Use the unified global scores/rerank to determine final top 5.
Include candidate-level scores and short reasons in the daily brief.
Prioritize papers that match the research interest and area-specific intent.
Strongly prioritize methodology-first contributions, especially core ML/AI methodology.
Prefer papers with clear algorithmic, modeling, optimization, inference, evaluation, or systems-method novelty.
Prefer methodological and mechanistic depth over benchmark-only gains.
Treat application papers as secondary unless they introduce broadly reusable methods or substantial methodological innovation.
Reward work that improves reasoning, controllability, efficiency, safety, or scientific usefulness.
Reward forward-looking novelty with likely 2-5 year impact.
Deprioritize incremental papers with weak conceptual contribution.
Deprioritize papers with weak grounding, vague claims, or limited technical signal.
Deprioritize application-only work with limited methodological transfer to the broader ML/AI community.
Soft Area-Mix Preferences For Final Top-N
Treat area mix as soft guidance, not hard constraints.
selected_area_index follows the area order in area_search_plan.
Prefer final top-5 to approximately satisfy:
area 1: at most 3, and preferably at least 1
area 2: at most 1
area 3: at most 2, and preferably at least 1
area 4: at most 1
area 5: at most 1
area 2 + area 4: preferably at least 1 in total
If quality/relevance strongly conflicts with area mix, prioritize quality/relevance and explain briefly in reasons.
Final Top-N Deep Read
For final selected top N papers, read full PDF text and produce paper-level summaries.
Highlight parts most related to the research interest profile.
Provide a relevance score and concise rationale for each selected paper.
Output should support generating one unified daily brief in the same day folder as downloaded papers.
The brief should contain both: (1) per-area top-3 final scoring table, (2) final top-N deep-read summaries.
Output Constraints
Select only from provided paper IDs.
Return strict JSON only.
Do not add explanation text outside the JSON response.