Literature Review Agent

Literature Review Agent (Step 3)

Faithful implementation of the Hybrid Literature Agent from PaperOrchestra (Song et al., 2026, arXiv:2604.05018, §4 Step 3, App. D.3, App. F.1 p.46).

Cost: ~20–30 LLM calls. This is one of the two longest steps (the other is plotting). Wall-time floor is set by Semantic Scholar's 1 QPS verification limit.

Inputs

• workspace/outline.json — specifically intro_related_work_plan with the Introduction search directions and the 2-4 Related Work methodology clusters
• workspace/inputs/conference_guidelines.md — used to derive cutoff_date
• workspace/inputs/idea.md, workspace/inputs/experimental_log.md — for framing the Intro and grounding the Related Work positioning

Outputs

• workspace/citation_pool.json — verified Semantic Scholar metadata for every paper that survived verification
• workspace/refs.bib — BibTeX file generated from the verified pool
• workspace/drafts/intro_relwork.tex — drafted Introduction and Related Work sections, written into the template, with the rest of the template preserved verbatim

Two-phase pipeline (App. D.3)

PHASE 1 — Parallel Candidate Discovery
   For each search direction in introduction_strategy.search_directions:
   For each limitation_search_query in each related_work cluster:
     - Use the host's web search tool to discover up to ~10 candidate papers.
     - Run up to 10 discovery queries in parallel (host-permitting).
     - Collect (title, snippet, url) tuples — no verification yet.
   → PRE-DEDUP before Phase 2 (see Step 1.5 below)

PHASE 2 — Sequential Citation Verification (1 QPS, with cache)
   For each candidate (after pre-dedup), sequentially:
     0. Check s2_cache.json first (scripts/s2_cache.py --check).
        If HIT: use cached response, skip live S2 call. No throttle needed.
        If MISS: proceed with live request below.
     1. Query Semantic Scholar by title:
          GET https://api.semanticscholar.org/graph/v1/paper/search?query=<title>
              &fields=title,abstract,year,authors,venue,externalIds&limit=5
        (Public endpoint, no key. Throttle to 1 QPS for live requests only.)
     2. Store the S2 response in cache: s2_cache.py --store.
     3. Pick the top hit. Check Levenshtein title ratio against the original
        candidate title. If ratio < 70: discard.
     4. Bonus: if year and venue exactly align with hints, add a +5 point
        match-quality bonus.
     5. Require: abstract is non-empty.
     6. Require: paper.year (or month if known) strictly predates cutoff_date.
        Months default to day-1: e.g., "October 2024" → 2024-10-01.
     7. If all checks pass, add to verified pool.
   After all candidates are verified, dedup by Semantic Scholar paperId.

The host agent does the LLM/web work; the deterministic helpers in scripts/ do the math.

Step-by-step

0. Derive cutoff_date

Parse conference_guidelines.md for the submission deadline. The paper aligns research cutoff with venue submission deadline (App. D.1):

Encode as YYYY-MM-DD. Months default to day-1 (e.g., 2024-10-01).

1. Phase 1: Parallel Candidate Discovery

From outline.json:

• All introduction_strategy.search_directions (3-5 queries)
• For each cluster in related_work_strategy.subsections:
  • The cluster's sota_investigation_mission becomes a search query
  • All limitation_search_queries (1-3 each)

For each query, use your host's web search tool (e.g., WebSearch in Claude Code, @web in Cursor, the search tool in Antigravity). Collect the top ~10 candidates per query: title, abstract snippet, source URL.

If your host supports parallel sub-tasks, fire up to 10 concurrent search queries. If not, run sequentially — slower but functionally equivalent.

Optional: Exa as a Phase 1 backend

If your host has no native web search, OR you want a research-paper-focused backend with better signal-to-noise, you can use Exa via the bundled scripts/exa_search.py helper. It is opt-in and reads EXA_API_KEY from the environment — the repo never commits a key.

export EXA_API_KEY="your-key-here"   # get one at https://dashboard.exa.ai/
python skills/literature-review-agent/scripts/exa_search.py \
    --query "Sparse attention long context transformers" \
    --num-results 15 \
    --discovered-for "related_work[2.1]"

Output is a normalized candidate list ready to merge into raw_candidates.json. Phase 2 verification (Semantic Scholar fuzzy match, cutoff, dedup) is unchanged. See references/exa-search-cookbook.md for the full recipe, query patterns, cost estimates, and security notes.

Optional: Tavily as a Phase 1 backend

If your host has no native web search, OR you want an LLM-optimized search backend with high relevance scoring, you can use Tavily via the bundled scripts/tavily_search.py helper. It is opt-in and reads TAVILY_API_KEY from the environment — the repo never commits a key.

export TAVILY_API_KEY="tvly-your-key-here"   # get one at https://app.tavily.com
python skills/literature-review-agent/scripts/tavily_search.py \
    --query "Sparse attention long context transformers" \
    --num-results 15 \
    --academic \
    --discovered-for "related_work[2.1]"

Output is a normalized candidate list ready to merge into raw_candidates.json. Phase 2 verification (Semantic Scholar fuzzy match, cutoff, dedup) is unchanged. See references/tavily-search-cookbook.md for the full recipe, query patterns, cost estimates, and security notes.

Combine all discovered candidates into a single working list. Tag each with the originating query ID so you can later attribute it to "intro" vs "related_work[i]".

1.5. Pre-dedup before Phase 2

Always run this before starting Phase 2. Multiple search queries routinely return the same papers (e.g., "Attention is All You Need" appears in almost every NLP discovery query). Verifying duplicates wastes 30-40% of S2 quota at 1 QPS.

python skills/literature-review-agent/scripts/pre_dedup_candidates.py \
    --in workspace/raw_candidates.json \
    --out workspace/deduped_candidates.json
# Prints: "150 candidates → 97 unique (53 duplicates removed)"

Use workspace/deduped_candidates.json as input to Phase 2.

2. Phase 2: Sequential Verification via Semantic Scholar (with cache)

For each candidate in deduped_candidates.json, in sequential order:

Step A — check cache first (no S2 call, no throttle needed):

python skills/literature-review-agent/scripts/s2_cache.py \
    --cache workspace/cache/s2_cache.json \
    --check "<candidate title>"
# exit 0 + prints JSON → use cached response, skip Step B
# exit 1 → proceed to Step B

Step B — live S2 request (cache MISS only, throttle to 1 QPS):

Preferred: use the bundled scripts/s2_search.py helper — it handles auth, retries, and 429 back-off automatically:

python skills/literature-review-agent/scripts/s2_search.py \
    --query "<URL-decoded candidate title>" --limit 5
# If SEMANTIC_SCHOLAR_API_KEY is set the key is forwarded automatically.
# If not, the public unauthenticated endpoint is used (≤1 QPS, still works).

Check whether the key is configured before starting Phase 2:

python skills/literature-review-agent/scripts/s2_search.py --check-key

Fallback: if you prefer your host's URL fetch tool, GET:

https://api.semanticscholar.org/graph/v1/paper/search?query=<URL-encoded title>&limit=5&fields=title,abstract,year,authors,venue,externalIds

Add header x-api-key: <SEMANTIC_SCHOLAR_API_KEY> if the env var is set. Be polite: ≤1 request per second for live requests. Cache hits are free.

Step C — store in cache (after every successful live request):

python skills/literature-review-agent/scripts/s2_cache.py \
    --cache workspace/cache/s2_cache.json \
    --store "<candidate title>" \
    --response '<full S2 JSON response>'

For the top hit:

python skills/literature-review-agent/scripts/levenshtein_match.py \
    --candidate "Original candidate title" \
    --found "S2 returned title"
# prints integer 0-100. Discard if < 70.

Then check the temporal cutoff:

python skills/literature-review-agent/scripts/check_cutoff.py \
    --paper-year 2024 \
    --paper-month 9 \
    --cutoff 2024-10-01
# exit 0 if strictly predates, exit 1 if not

If both checks pass AND the abstract is non-empty, append the paper's full S2 metadata to the verified pool.

3. Dedup and assemble the pool

After all candidates are verified:

python skills/literature-review-agent/scripts/dedupe_by_id.py \
    --in raw_pool.json \
    --out workspace/citation_pool.json

The dedupe script keys on paperId (Semantic Scholar's internal unique ID), falling back to externalIds.DOI, then externalIds.ArXiv, then a normalized title.

The script also computes and writes min_cite_paper_count = floor(0.9 * len(papers)) — the minimum number of papers the writing step must cite (the paper's ≥90% integration rule, App. D.3).

Immediately after dedupe_by_id.py, validate and auto-fix the pool schema:

python skills/literature-review-agent/scripts/validate_pool.py \
    --pool workspace/citation_pool.json --fix
# Catches and fixes authors-as-strings, reports missing required fields.
# Must pass before proceeding to Step 4.

3.5. Cross-index verification (Crossref + OpenAlex)

Semantic Scholar is one index and can return a plausible record for a paper that does not exist, or attach wrong metadata. Re-check every S2-verified paper against two independent indices before building the bibliography — this is the practical defense against hallucinated citations leaking in.

# Optional but recommended: a polite-pool email gives faster, more reliable
# service. The repo never commits an address.
export PAPER_ORCHESTRA_MAILTO="you@example.com"

python skills/literature-review-agent/scripts/cross_verify.py \
    --pool workspace/citation_pool.json --inplace
# Annotates each paper with a `cross_verification` field and writes
# workspace/cross_verification_report.json.
# exit 0 = all corroborated; exit 1 = WARN (something flagged or an index
# was unreachable); exit 2 = usage error.

This is a WARN gate, not a hard gate (like validate_consistency.py): it flags suspicious citations but does not block the pipeline or delete anything. Review the low and conflict tiers in the report:

• high — corroborated by ≥1 external index → keep.
• medium — corroborated but year disagrees → keep, spot-check the year.
• low — not found in Crossref or OpenAlex → review by hand. Note that arXiv-only preprints (no DOI) are a common benign cause; low means "could not corroborate," not "fabricated." S2 already confirmed it exists.
• conflict — pool DOI disagrees with the external DOI → likely wrong record.

Drop only the entries you genuinely cannot corroborate, then re-run dedupe_by_id.py onward. If both indices are unreachable (offline), the script degrades gracefully and the pipeline continues on S2 verification alone.

See references/cross-index-verification.md for the full rationale, confidence tiers, and the arXiv false-positive note.

4. Build the BibTeX file

python skills/literature-review-agent/scripts/bibtex_format.py \
    --pool workspace/citation_pool.json \
    --out workspace/refs.bib

The script generates citation keys deterministically from `firstauthor + year

• first significant word of title(e.g.,vaswani2017attention). It writes out only @article/@inproceedings/@miscentries — never invents fields. It also writes the canonicalbibtex_keyback into each paper record incitation_pool.json`.

Immediately after bibtex_format.py, sync keys in intro_relwork.tex:

python skills/literature-review-agent/scripts/sync_keys.py \
    --pool workspace/citation_pool.json \
    --tex  workspace/drafts/intro_relwork.tex \
    --inplace
# Replaces every \cite{agent_key} with \cite{canonical_bibtex_key}.
# Eliminates citation_coverage gate failures caused by key mismatch.

These two steps replace the manual Python snippets that were previously required. The pipeline is now:

dedupe_by_id → validate_pool --fix → cross_verify --inplace → bibtex_format → sync_keys

5. Draft Introduction + Related Work

This is where you (the host agent) actually write text. Load the verbatim Literature Review Agent prompt at references/prompt.md. Substitute the template placeholders:

Also prepend the Anti-Leakage Prompt from ../paper-orchestra/references/anti-leakage-prompt.md.

Run your LLM with the combined prompt against template.tex. The agent's job is to fill in the empty Introduction and Related Work sections of the template and leave everything else untouched. Output: the full template.tex with those two sections filled. Save to workspace/drafts/intro_relwork.tex.

5b. Append §2 to research_brief.md

After intro_relwork.tex is drafted and before the citation coverage check, append §2 to workspace/research_brief.md (see skills/shared/research_brief_template.md).

Template:

## §2 · Literature Landscape
_Written by: literature-review-agent, Step 3_

**What the literature says about the core claim:** <2-3 sentence synthesis>

**Strongest prior work (must address in the paper):**
- <bibtex_key>: <why this is the strongest comparator or predecessor>

**Gaps confirmed by the literature:** <list>

**Baseline comparisons — verification status:**
| Baseline | In citation_pool? | Confidence tier |
|---|---|---|

**Related Work cluster coverage:**
| Cluster | Papers found | Notes |
|---|---|---|

**Anything the section-writing agent should know:** <important context>

This synthesises what was actually found — not what the outline assumed.

6. Verify ≥90% citation coverage

python skills/literature-review-agent/scripts/citation_coverage.py \
    --tex workspace/drafts/intro_relwork.tex \
    --pool workspace/citation_pool.json
# exit 0 if ≥90% of pool is cited; exit 1 otherwise

If the gate fails, re-prompt the writing step explicitly listing the missing keys and asking the agent to integrate them where contextually appropriate.

Critical rules from the prompt

These are excerpted from references/prompt.md. The host agent MUST honor them on the writing call:

• Cite ONLY from collected_papers. Never invent BibTeX keys, never reference papers not in the pool.
• Cite at least min_cite_paper_count of them in Intro + Related Work combined.
• TIMELINE RULE: Do not treat any papers published after cutoff_date as prior baselines to beat. They are concurrent work only.
• EVALUATION RULE: Do not claim our method beats / achieves SOTA over a specific cited paper UNLESS that paper is explicitly evaluated against in experimental_log.md. Frame other recent papers strictly as concurrent, orthogonal, or conceptual work.
• Output format: return the full code for the updated template.tex, with the two empty sections (Introduction and Related Work) filled in, and all the other code (packages, styles, other sections) identical to the original template.tex.
• Wrap output in ```latex ... ``` fences.
• Do not change \usepackage[capitalize]{cleveref} to cleverref (there is no cleverref.sty).

Degraded mode (no web search)

If your host has no web search tool, switch to degraded mode:

1. If the user has placed a pre-built workspace/inputs/refs.bib in the workspace, load it directly into workspace/refs.bib and skip Phase 1 and Phase 2.
2. Otherwise, emit workspace/drafts/intro_relwork.tex containing the template with two TODO markers in the Intro and Related Work sections, and tell the user the pipeline cannot complete Step 3 without web search.

Resources

• references/prompt.md — verbatim Literature Review Agent prompt from App. F.1
• references/discovery-pipeline.md — Phase 1 + Phase 2 explained in detail
• references/verification-rules.md — Levenshtein cutoff, year alignment, dedup
• references/citation-density-rule.md — the ≥90% integration rule
• references/s2-api-cookbook.md — Semantic Scholar URLs, fields, rate limits
• references/cross-index-verification.md — Crossref + OpenAlex corroboration, confidence tiers, arXiv false-positive note
• references/exa-search-cookbook.md — optional Exa backend for Phase 1 (research-paper-focused web search)
• references/tavily-search-cookbook.md — optional Tavily backend for Phase 1 (LLM-optimized web search)
• scripts/pre_dedup_candidates.py — NEW dedup Phase 1 candidates before Phase 2 (saves 30-40% S2 quota)
• scripts/s2_cache.py — NEW persistent S2 response cache (eliminates re-verification on re-runs)
• scripts/validate_pool.py — NEW validate & auto-fix citation_pool.json schema (authors format)
• scripts/sync_keys.py — NEW sync cite keys in .tex with canonical bibtex_keys after bibtex_format.py
• scripts/levenshtein_match.py — fuzzy title match (ratio > 70)
• scripts/check_cutoff.py — date cmp w/ month → day-1 default
• scripts/dedupe_by_id.py — dedup verified pool by S2 paperId
• scripts/bibtex_format.py — build refs.bib from JSON pool
• scripts/citation_coverage.py — ≥90% citation coverage gate
• scripts/s2_search.py — NEW Semantic Scholar title-search helper; reads SEMANTIC_SCHOLAR_API_KEY from env (optional — falls back to unauthenticated)
• scripts/exa_search.py — optional Exa Phase 1 backend (reads EXA_API_KEY from env)
• scripts/tavily_search.py — optional Tavily Phase 1 backend (reads TAVILY_API_KEY from env)
• scripts/crossref_client.py — NEW Crossref title/DOI lookup for cross-index corroboration (no key; reads CROSSREF_MAILTO / PAPER_ORCHESTRA_MAILTO)
• scripts/openalex_client.py — NEW OpenAlex title/DOI lookup for cross-index corroboration (no key; reads OPENALEX_MAILTO / PAPER_ORCHESTRA_MAILTO)
• scripts/cross_verify.py — NEW cross-corroborate the S2-verified pool against Crossref + OpenAlex; flags hallucinated citations (WARN gate)
• skills/shared/research_brief_template.md — NEW §2 schema; append after intro_relwork.tex is drafted