---

name: paper-research description: End-to-end paper research support for arXiv/literature surveys, reproducibility-focused paper shortlisting, and experiment design. Use when you need to (1) search arXiv with complex queries, (2) download PDFs, extract text/sections, and fetch BibTeX, (3) dedupe/cluster results into a structured report, and (4) turn findings into a lit-review plan, benchmark/evaluation suite, and representation/probing experiment checklist (e.g., implicit reasoning, hidden-CoT, multilingual reasoning, cross-lingual alignment).

Paper Research

Overview

Run a fast, reproducible “survey → shortlist → synthesize” loop for research topics, backed by small scripts that fetch arXiv metadata/PDFs/BibTeX, extract text, and generate structured Markdown briefs.

Quick start (recommended workflow)

1. Create a topic workspace directory (keep everything together):
   • Example: notes/implicit-reasoning-survey/
2. Search arXiv and (optionally) download PDFs:
   • Run: python3 scripts/arxiv_survey.py --terms "implicit reasoning" "hidden chain-of-thought" "multilingual reasoning" --max-results 100 --download-pdfs --pdf-dir ./pdfs --out ./arxiv.jsonl
3. Extract text (+ rough sections) from PDFs:
   • Run: python3 scripts/pdf_extract.py --pdf-dir ./pdfs --out-dir ./texts --sections
4. Fetch BibTeX for the found arXiv IDs:
   • Run: python3 scripts/arxiv_bibtex.py --from-jsonl ./arxiv.jsonl --out ./refs.bib
5. Generate a structured research brief (table + clusters + TODO slots for notes):
   • Run: python3 scripts/generate_report.py --jsonl ./arxiv.jsonl --out ./REPORT.md

Then ask Codex to synthesize (taxonomy/benchmarks/experiments) using REPORT.md + your notes.

Workflow decision tree

A) “I need a lit review plan + paper outline”

Do this:

1. Use the scripts to produce REPORT.md (table + clusters) and refs.bib.
2. Build a survey plan as a set of falsifiable questions + “what evidence would change my mind”.
3. Output deliverables (in this order):
   • Lit review plan (subtopics → why → representative papers to read first)
   • Benchmarks/metrics (existing + proposed) aligned to the hypothesis
   • Validation experiments (including representation/probing/interventions)
   • Paper outline + expected contributions

When relevant, include “fastest path to reproduce” (datasets, eval harnesses, probing code).

B) “I need a reproducibility-first shortlist”

Prioritize:

• Open-source repos (training recipe, evaluation harness, probing code)
• Clear protocol (hyperparams, seeds, compute, preprocessing)
• Reusable artifacts (scripts, configs, checkpoints, datasets)

Do this:

1. Run arxiv_survey.py with stricter terms and fewer results (e.g., 30–80).
2. Ask Codex to rank papers in REPORT.md by reproducibility criteria:
   • Code availability, license clarity, dataset accessibility, protocol completeness
3. Produce:
   • Ranked shortlist with repo links (if available)
   • “Reusable parts” per paper (eval harness / probing / training recipe)
   • Minimal reproduction plan (timeboxed: 2h / 1d / 1w)

C) “I need an evaluation suite + detection experiments (multilingual latent reasoning)”

Use this structure:

1. Hypothesis → operational definition (what counts as “English latent reasoning”)
2. Tasks:
   • Multi-step reasoning across languages (same semantics, different surface forms)
   • Translation-free reasoning (language-neutral, symbol-heavy, or synthetic)
   • Controlled prompts enforcing target-language output
3. Metrics that separate reasoning vs fluency:
   • Task accuracy, step-consistency proxies, calibration, controllability, latency
4. Representation-level detection:
   • Layer-wise language ID / probing on activations
   • Activation patching/interventions (swap “language subspace” signals)
   • Forced-language and mixed-language ablations
5. Expected signatures + failure modes (confounds: translation, tokenization, data mixture)

Use assets/experiment_checklist.md as the backbone checklist.

Templates (assets/)

Copy and fill these as working docs:

• assets/research_brief.md → one-topic brief (taxonomy + top papers + open questions)
• assets/paper_comparison_table.md → consistent per-paper extraction fields
• assets/experiment_checklist.md → step-by-step experimental checklist

Scripts

All scripts are pure-Python (stdlib) where possible. pdf_extract.py supports optional extractors; if none are available, it prints a clear install hint.

scripts/arxiv_survey.py

Search arXiv via the official Atom API, write results to JSONL, and optionally download PDFs.

scripts/arxiv_bibtex.py

Fetch BibTeX from arxiv.org for a list of arXiv IDs or a JSONL produced by arxiv_survey.py.

scripts/pdf_extract.py

Extract text from PDFs into .txt and optionally produce rough section splits (heuristics).

scripts/dedupe_jsonl.py

Dedupe a JSONL file by arxiv_id and near-duplicate titles (useful when iterating queries).

scripts/generate_report.py

Generate a structured Markdown report (table + clusters + TODO note slots) from arxiv.jsonl.

References

Read when you need query patterns or a report schema:

• references/arxiv_query_guide.md
• references/report_fields.md

---

Output quality bar (what “good” looks like)

• Prefer explicit assumptions + failure modes over broad claims.
• Prefer checklists and protocols over vague “future work”.
• Always separate: (1) claim, (2) evidence, (3) test that could falsify it.