Internlm Wildclawbench Wildclawbench

ArXiv Summarizer Orchestrator

Run the full pipeline by composing three sub-skills.

Sub-skill Order

1. arxiv-search-collector
2. arxiv-paper-processor
3. arxiv-batch-reporter

Workflow Parameters

• language: manual language parameter used by all stages. Default is English when omitted.
• paper_processing_mode: subagent_parallel or serial.
• max_parallel_papers: default 5 when paper_processing_mode=subagent_parallel.

Workflow

Stage A: Collection Setup + Query Retrieval

1. Initialize one run with arxiv-search-collector/scripts/init_collection_run.py.
2. Model generates multiple focused queries from original topic and writes a minimal query_plan.json (label + query only).
3. Run arxiv-search-collector/scripts/fetch_queries_batch.py with the plan file (recommended).
4. (Optional fallback) call arxiv-search-collector/scripts/fetch_query_metadata.py manually for one-by-one fetch.
5. Model reads each indexed query list and decides keep indexes.
6. Merge selected items with arxiv-search-collector/scripts/merge_selected_papers.py.
7. If relevance/coverage is still not good, iterate Stage A:
   • generate another query plan with new labels,
   • fetch again,
   • re-merge with --incremental and updated selection-json.
   • set weak labels to empty keep list ([]) to explicitly drop them.

Pass --language <LANG> to collector scripts so all generated markdown files in Stage A follow the selected language. Use serial query fetch in Stage A with conservative controls (for example --min-interval-sec 5, --retry-max 4). Default collector settings already include retries/backoff and run-local throttle state (<run_dir>/.runtime/arxiv_api_state.json), so manual tuning is usually unnecessary. Prefer cache reuse (no --force) unless query parameters changed or data refresh is required.

Output: one run directory with per-paper metadata subdirectories.

Stage B: Per-paper Artifact Download + Manual Summary

For each paper directory, invoke sub-skill arxiv-paper-processor once and let that skill produce <paper_dir>/summary.md.

Recommended pre-step for many papers:

1. Run one batch artifact download before per-paper reading:

python3 arxiv-paper-processor/scripts/download_papers_batch.py \
  --run-dir /path/to/run \
  --artifact source_then_pdf \
  --max-workers 3 \
  --min-interval-sec 5 \
  --language <LANG>

Per-paper execution steps (inside arxiv-paper-processor):

1. If <paper_dir>/summary.md already exists and is complete, skip this paper.
2. If usable source (source/source_extract/*.tex) or PDF (source/paper.pdf) already exists, skip download.
3. If artifacts are missing, download source with arxiv-paper-processor/scripts/download_arxiv_source.py.
4. If source is unusable, download PDF with arxiv-paper-processor/scripts/download_arxiv_pdf.py.
5. Model reads content and manually writes <paper_dir>/summary.md by reference format, in language.

Parallel strategy for many papers:

• Default: paper_processing_mode=subagent_parallel with max_parallel_papers=5.
• Optional: paper_processing_mode=serial to process one paper at a time.
• In parallel mode, run multiple arxiv-paper-processor instances in batches; concurrent papers must not exceed max_parallel_papers.
• Wait for one batch to finish before starting the next batch.
• In serial mode, run exactly one arxiv-paper-processor instance at a time.
• Subagent workers should only own one paper directory each to avoid file conflicts.
• Do not use scripts to auto-compose summary text; scripts are download-only tools.

Output: all paper directories contain summary.md.

Stage C: Bundle + Final Hierarchical Report

1. Run arxiv-batch-reporter/scripts/collect_summaries_bundle.py --language <LANG>.
2. Model reads summaries_bundle.md and writes collection_report_template.md in base dir.
3. In template, each paper leaf entry must include one standalone placeholder line: {{ARXIV_BRIEF:<arxiv_id>}}.
4. Run arxiv-batch-reporter/scripts/render_collection_report.py to generate final collection_report.md.
5. Do not manually paraphrase per-paper conclusion lines in final report; they must come from per-paper summary.md section 10 via script injection.

If language is non-English (for example Chinese), all intermediate markdown files and final reports should follow that language.

Periodic Scheduling

This orchestrator is suitable for cron/scheduled execution in OpenClaw:

• Frequency examples: daily, weekly, monthly.
• For rolling windows, use lookback (1d, 7d, 30d) when initializing runs.

Output Layout

<output-root>/<topic>-<timestamp>-<range>/

• task_meta.json, task_meta.md
• query_results/, query_selection/
• <arxiv_id>/metadata.md + downloaded source/pdf + summary.md
• summaries_bundle.md
• collection_report_template.md
• final rendered collection report (e.g. collection_report.md)

Use references/workflow-checklist.md as execution checklist.

Related Skills

This is the top-level orchestration skill.

Before using it, install and enable these three sub-skills:

• arxiv-search-collector
• arxiv-paper-processor
• arxiv-batch-reporter

Execution order inside this orchestrator:

1. arxiv-search-collector (Stage A)
2. arxiv-paper-processor (Stage B)
3. arxiv-batch-reporter (Stage C)

Source: InternLM/WildClawBench — distributed by TomeVault.