• Empirical vs. theoretical — papers that observe/measure the phenomenon vs. papers that prove/explain it formally
• Mechanism vs. condition — papers about how something works vs. when/why it emerges
• Method keywords — different communities use different terms for the same concept (e.g., "gradient descent" vs. "meta-optimization" vs. "implicit learning")
• Adjacent formulations — the same idea framed differently (e.g., "in-context learning" vs. "few-shot learning" vs. "learning from demonstrations")

• Synonym substitution: "data pruning" → "data selection", "data filtering", "data curation"
• Specificity adjustment: broaden ("pretraining data quality") or narrow ("perplexity-based data pruning LLM")
• Structural variants: swap word order, add/remove qualifiers, use abbreviations

• Angles: (a) mechanistic/circuit, (b) training dynamics, (c) ICL-as-optimization theory, (d) data/task conditions, (e) formal theory
• Query 1:

  "in-context learning emergence pretraining language model"

  (general)
• Query 2:

  "induction heads formation training transformer"

  (mechanistic)
• Query 3:

  "transformers learn in-context gradient descent meta-learning"

  (optimization view)
• Query 4:

  "pretraining task diversity data structure in-context learning"

  (data conditions)
• Query 5:

  "in-context learning theory linear attention generalization"

  (formal theory)

• Angles: (a) selection methods, (b) quality metrics, (c) scaling effects
• Query 1:

  "data pruning pretraining language model"

• Query 2:

  "data selection pretraining LLM"

• Query 3:

  "training data curation large language model quality"

• Query 4:

  "data quality scoring pretraining scaling"

• Primary:

  scholar_search

  (S2 with automatic arXiv fallback on rate limit)
• Secondary:

  arxiv_monitor --keywords "<variants>" --match-mode flexible

  for broader keyword coverage
• Tertiary (when S2 is rate limited): web search for recent blog posts/surveys that reference papers

• With

  S2_API_KEY

  set (100 req/min): You MAY run multiple

  scholar_search

  calls in parallel.
• Without

  S2_API_KEY

  (100 req/5min, ~1 req/3s): You MUST run

  scholar_search

  calls sequentially, one at a time. Parallel S2 calls without a key will exhaust the rate limit immediately, causing all calls to fail with 429 and fall back to the lower-quality arXiv search. This applies to ALL S2-dependent scripts:

  scholar_search

  ,

  citation_traverse

  ,

  recommend

  ,

  author_search

  ,

  trending

  .
• How to check: Before starting discovery, run

  echo $S2_API_KEY

  or check if the env var is set. If empty, switch to sequential mode.
• arXiv-only scripts (

  arxiv_monitor

  ) are NOT affected by this rule and can always run in parallel with other calls.

1. scholar_search

   automatically falls back to arXiv (built-in since v1.2)
2. Use

   arxiv_monitor --keywords

   with

   --match-mode flexible

   for broader coverage
3. Switch to web search for blog posts, surveys, GitHub repos that reference papers
4. Space S2-dependent calls (

   citation_traverse

   ,

   recommend

   ) at least 5s apart and reduce

   --limit

1. Co-citation on the single highest-cited seed:

   citation_traverse --direction co-citation --limit 15

   — this is the strongest signal for finding closely related work that uses different terminology
2. Forward citations on the top 2 seeds:

   citation_traverse --direction forward --limit 20

   — finds follow-up work
3. Backward citations on 1-2 seeds whose topic coverage differs:

   citation_traverse --direction backward --limit 20

   — finds foundational and adjacent work that seeds build on. Pick seeds from different sub-topics to maximize coverage breadth
4. Recommendations with diverse seeds:

   recommend --positive <seed1>,<seed2>,<seed3>

   — serendipitous discovery of semantically related work not connected by citations

1. Count how many collected papers address it
2. If a sub-topic has 0-1 papers, run a targeted

   scholar_search

   with a query specific to that angle
3. If targeted search finds new relevant papers, optionally run one more

   citation_traverse

   or

   recommend

   round on the new finds

1. Quick academic search — Try

   scholar_search

   with the exact query
2. If zero results — Broaden the search:
   • Web search: Find GitHub repos, blog posts, or social media that reveal the actual paper title or arXiv ID
   • GitHub search:

     github_search.py --query "USER_QUERY"

     — repos often link to papers
3. Extract identifiers — Actual paper title, arXiv ID, GitHub repo URL, author names
4. Re-enter the appropriate workflow with resolved identifiers

🔍 Disambiguation Report for "deepseek engram"
├── Intent: Track recent advances (ambiguous term)
├── Resolution: "Engram" is a module name from DeepSeek AI
│   ├── Actual paper: "Conditional Memory via Scalable Lookup" (ArXiv:2601.07372)
│   └── GitHub: https://github.com/deepseek-ai/Engram
└── Search Plan:
    ├── scholar_search --query "Conditional Memory Scalable Lookup" --sort-by year
    ├── citation_traverse --paper-id ArXiv:2601.07372 --direction forward
    └── github_search --query "deepseek engram"

"Help me survey CRISPR-based gene therapy for sickle cell disease"

1. Discover: Initial

   scholar_search --query "CRISPR gene therapy sickle cell" --limit 20 --sort-by citations

   → iterative expansion with EXPLORE/EXPLOIT strategy →

   citation_traverse --direction forward

   on seminal papers
2. Evaluate: Review each paper's title + abstract for relevance → filter by abstract quality → prefer top-tier venues → shortlist
3. Read:

   fetch_paper

   for key papers → L2 reading → notes using

   assets/paper-summary-template.md

4. Hand off to

   research-survey

   to synthesize the collected papers into a structured survey report

"Find me the attention is all you need paper" "Find me the original GPT 3 paper"

1. Discover:

   scholar_search --query "Attention Is All You Need"

   — single call, return top result
2. Output: Use Format A (Single Paper Card)

"Give me papers about perovskite solar cell stability under humidity" "Find papers on gut microbiome modulation for autoimmune diseases"

1. Sub-topic decomposition + query reformulation: Identify 3-5 research angles within the topic, generate 4-6 variant queries covering distinct angles (see Step 0)
2. Discover: Run

   scholar_search --query "<variant>" --limit 20 --sort-by relevance

   on each variant. If

   S2_API_KEY

   is set, parallelize these calls; if not, run them sequentially one at a time to avoid rate-limit exhaustion (see "S2 parallelization rule" in Step 0). Also run

   arxiv_monitor --keywords "<variants>" --match-mode flexible

   for additional coverage (arXiv calls can always run in parallel with other non-S2 calls)
3. Citation expansion (if initial results ≥ 3 relevant papers): Follow Mandatory Citation Expansion (Step 0) — co-citation on highest-cited seed, forward on top 2, backward on 1-2 diverse seeds, recommend with 3 seeds
4. Coverage gap check: Review collected papers against identified sub-topics. Run targeted searches for any uncovered angles (see Step 0)
5. Filter: Review all results, deduplicate, keep relevant papers based on title + abstract
6. Output: Use Format B (Paper List Table)

"2012 papers by David Harel" "Papers by David Harel from 2020 to 2022" "Journal articles by David Harel from 2020 to 2022"

1. Parse query: Extract author name, year range, venue type (journal/conference)
2. Discover:

   author_search --name "David Harel" --papers --limit 50 --sort-by year

3. Filter: Year range, venue type (check

   venue

   field), other attributes
4. Output: Use Format B (Paper List Table)

scholar_search --query "<keywords>" --year-min YYYY --year-max YYYY

"What's new in condensed matter physics this week?"

1. Discover:

   arxiv_monitor --categories cond-mat --days 7

   (see

   references/arxiv-categories.md

   for codes) +

   trending --query "topological insulator" --period 30

2. Output: Use Format B (Paper List Table), highlight high-potential papers with TLDRs

"I need a baseline for protein structure prediction with code"

1. Discover:

   scholar_search --query "protein structure prediction" --sort-by citations

2. Evaluate:

   find_code

   on top results +

   sota --task "protein-structure-prediction"

   → pick one with official code + high downloads
3. Output: Use Format C (Baseline Recommendation)

"Read this paper: arxiv.org/abs/2301.12345"

1. Fetch:

   fetch_paper --url "https://arxiv.org/abs/2301.12345"

2. Choose reading depth (see

   references/reading-strategy.md

   ):

1. Take notes using

   assets/paper-summary-template.md

   . Save to

   /artifacts/paper-notes/{paper-id}.md

   .

"Find the latest about deepseek engram"

1. Disambiguate: Follow Step 2 above
2. Discover:

   scholar_search

   with resolved title +

   github_search

   with original term +

   citation_traverse

   on arXiv ID
3. Evaluate: Review results, check code via

   find_code

   or GitHub
4. Read:

   fetch_paper

   for top papers
5. If user wants a survey: hand off to

   research-survey

research-ideation Step 2 needs papers to build a literature tree

1. Disambiguate: Parse the research goal → extract domain + method type
2. Discover: Initial broad search (60 candidates) → iterative expansion up to 15 rounds:
   • EXPLORE: new keyword queries for diverse sub-areas
   • EXPLOIT:

     citation_traverse

     or

     recommend

     on strongly relevant papers
3. Evaluate: Only keep strongly relevant papers. Prefer top-tier venues + 2020+ papers.
4. Deduplicate: Track seen titles and abstracts.
5. Output: 30-50 high-quality papers → feed into novelty tree + challenge-insight tree.

"Find me exactly 15 papers about reinforcement learning from human feedback"

1. Use the user's number as the target
2. Apply the closest profile's quality settings
3. Run iterative collection until target met or max iterations exhausted
4. If not enough, progressively relax relevance standard and inform the user

1. Parse query → extract goal, search terms, key term definitions
2. Define task attributes → identify domain + method type
3. Initial search → scholar_search with broad query
4. Review each paper's title + abstract → judge relevance (keep/reject)
5. LOOP until target met or max iterations reached:
   a. From kept papers, pick the most relevant as "grounding set"
   b. Generate next search query:
      - EXPLORE: new keyword query to broaden coverage
      - EXPLOIT: citation_traverse or recommend on a high-relevance paper
   c. Fetch new papers → review → deduplicate → add to collection
6. Final filter: apply quality checks, take top N

1. Skip papers with very short abstracts (< 20 words)
2. For ideation/survey: prefer top-tier venues and journals in the user's field (e.g., Nature, Science, Cell, Lancet, PNAS for broad science; field-specific top venues like NeurIPS/ICML for ML, Physical Review Letters for physics, JACS for chemistry, etc.)
3. For ideation: prefer 2020+ papers; include older only if foundational

• research-survey: After collecting papers, hand off to research-survey for structured survey report generation (8-section goal-centric synthesis).
• research-ideation: After collecting papers, hand off to research-ideation for idea generation (novelty tree + challenge-insight tree + problem selection + solution design).
• experiment-pipeline: After finding a baseline via Workflow 6, hand off to experiment-pipeline.
• paper-writing: Paper notes serve as input for paper-writing's Related Work section.