Paper Navigator

Find and read academic papers. Route by intent, judge by relevance.

        User
         │
         ▼
   ┌── Router ──┐
   │            │
   ▼            ▼
 POINT      LIST/ITERATIVE
(1 paper)   (Probe + up to 3 paper rounds:
             R2 breadth / R3 deepen / R4 close)

The agent does relevance judgment — no LLM-as-judge is called, no numeric scoring. You author the rubric, you triage each paper, you rank by relevance.

Setup

Scripts at skills/paper-navigator/scripts/. Run via python skills/paper-navigator/scripts/<name>.py.

arXiv access (arxivmonitor, scholarsearch fallback) uses the DeepXiv SDK: pip install deepxiv-sdk, then deepxiv token once to provision a free API token (saved to ~/.env). The skill reads the token from DEEPXIVAPITOKEN/DEEPXIV_TOKEN in the environment, or from ./.env / ~/.env.

Full env-var list: references/env-vars.md.

Five Red Lines (always)

1. Track history. Don't re-run a query you already ran. Empty result → change angle, not synonyms.
2. Search a gap, not a vibe. Every query maps to one missing piece of information. No stacked-keyword bags.
3. One query = one concept. Split comparisons (A vs B), multi-property asks, and multi-year spans into separate calls.
4. Never hallucinate. Every fact (title, author, year, citation count, content) comes from a tool result.
5. Quote-or-zero. When you claim a paper meets a criterion, quote a ≤80-char span from its abstract / tldr / snippet. No quote → do not claim the paper meets that criterion. (This guards against hallucination; it does not drive a numeric score.)

Router

Default to LIST when unsure. Don't add survey / review to LIST queries — it down-ranks the canonical research papers the user wants.

Output format follows the caller: the Output column above is the structured form (skill callers). Direct user calls default to Narrative — see Step 6 "Output mode".

Ambiguous query (project nickname, codename, single capitalized word with zero hits) → run scholar_search exact + web/GitHub search first to resolve identifiers, then re-route.

POINT branch (known paper)

Paper Card:

📄 **<Title>**
Authors: <First Author> et al. | Year: <Y> | Venue: <V>
Citations: <N> | ID: <ArXiv:xxxx.xxxxx> | DOI: <...>
TLDR: <one sentence>

Stop here. Do not chain to citation expansion unless asked.

LIST / ITERATIVE branch — 6 steps

Step 1: Parse intent

State in one sentence: the research object (specific technique / concept) and the constraints (domain, task, recency, exclusions). Confirm the router branch. When the user gives only a bare noun (concept / model / algorithm / benchmark name) with no direction, default intent is "trace the lineage" — foundations, evolution, current state — not applications or a generic survey.

Step 2: Author the RUBRIC (via think_tool)

Emit a structured block before any search. It persists across rounds and every later step references it.

RUBRIC for "<user query verbatim>"
Branch: LIST | ITERATIVE
Criteria (2–4, atomic; mark each [core] or [secondary]):
  C1 [core]     <what the paper MUST do/be — one sentence>
  C2 [core]     <...>
  C3 [secondary] <...>
Named entities to preserve verbatim: [<ent1>, <ent2>, ...]
Angle tags (3–5 sub-topic axes): [<tag1>, <tag2>, <tag3>]
Recency signal: [none | recency cue — defined in Step 3]
Disqualifiers: [<auto-reject if abstract shows this>]

Rules:

• Criteria atomic (one condition each), non-redundant. Mark each [core] (must-have) or [secondary] (nice-to-have) — this guides relevance ranking, no weights, no math.
• Named entities = proper-noun / technical-term anchors from the user's query. Every entity appears verbatim in ≥1 query across Probe + R2.
• Angle tags = sub-topic axes (method, task, dataset, evaluation, domain, …). No two queries in the same round share a tag.
• Recency signal = whether the user query carries a recency cue (see Step 3 recency trigger). Drives --year-min and the Step 6 recency tie-break.
• Disqualifiers = "specifically X, not Y" exclusions. Tripping a disqualifier → Irrelevant.

For ITERATIVE, criteria can be lighter (e.g. covers topic + is canonical); disqualifiers may be empty.

Step 3: Search — Probe then Multi-round (Breadth → Deepen → Close)

Probe first to grasp the user's need comprehensively, then run up to 3 paper rounds (R2 breadth / R3 deepen / R4 close). Decide round-by-round whether more rounds are needed (Step 5 gate) — do NOT fix a round count. There is no R5; the cap is 3 paper rounds.

Probe (prerequisite — NOT counted in the 3 rounds, 2 parallel queries): establish a comprehensive understanding of the need and lift named entities / angle gaps.

• Q-broad — canonical phrasing of the topic (angle: general)
• Q-narrow — a specific mechanism / sub-question / method (angle: tagged)

python scripts/scholar_search.py --query "<Q-broad>"  --limit 15 --sort-by relevance --output /tmp/pool.jsonl --append
python scripts/scholar_search.py --query "<Q-narrow>" --limit 15 --sort-by relevance --output /tmp/pool.jsonl --append

--output --append auto-dedupes by paperId across rounds (built into the script), so a paper found by two queries is written once. Read /tmp/pool.jsonl to inspect (Step 4 triage). From Probe titles + tldrs, lift:

• recurring named entities (algorithm / benchmark / dataset / model names),
• angle gaps (Step-2 tags not seen),
• vocabulary from adjacent communities.

R2 — Breadth (default — skipped only via the Step 5 early exit; 2–3 parallel queries). Broad academic queries on the core object + canonical terms lifted from Probe. The main axis is the user's core object + canonical mapping; Probe only supplies high-confidence supplements.

R3 — Targeted deepening (optional, when Step 5 gate says CONTINUE, 2–3 parallel queries). Driven by the gap→strategy map below — fill the specific gap Step 4 triage exposed.

R4 — Closing (optional, 2–3 parallel queries). Fill whatever key gap remains after R3: a missing representative work, a strong baseline, a counter-example, or an uncovered sub-direction.

Gap→strategy map (R3/R4 are driven by the Step-4 triage gap, not vibes):

Per-query rules:

• 3–6 words preferred (English academic terms); <3 over-recalls, >6 dilutes ranking.
• Use academic terms (mechanism, benchmark); no how it works phrasing.
• Do NOT add survey / review / tutorial terms — they bias results toward review papers and crowd out the research papers the user wants. Only use them when the user explicitly asks for a survey/review.
• Bare entity names; no paper / pdf / arxiv / original.
• Split comparisons / multi-property; if short of 2–3 queries, fill with upper-topic or representative method.
• Time intent goes into --year-min/max params, never year words in the query.
• Forbidden: "…", (..), OR, AND, |, site:, filetype:.
• No two queries in one round may share >60% of content tokens (after stop-words).

Recency trigger (query layer). When the user query contains a recency signal (最新 / 近年 / 近期 / 近两年 / 前沿 / SOTA / latest / recent / state-of-the-art), set --year-min to last 2 years from R2 onward — Jan 1 of the year before the current system year (e.g. system date 2026-08 → --year-min 2025). Do not extend to 3–4 years.

Without S2APIKEY: swap scholarsearch for arxivmonitor --keywords "<variant>" --match-mode flexible --days 3650.

Citation expansion (ITERATIVE, or LIST after ≥3 All-core/Partial seeds):

python scripts/citation_traverse.py --paper-id <SEED> --direction co-citation --limit 15 --output /tmp/pool.jsonl --append
python scripts/citation_traverse.py --paper-id <SEED> --direction forward --limit 20 --min-citations 20 --year-min 2022 --output /tmp/pool.jsonl --append
python scripts/recommend.py --positive <SEED1>,<SEED2> --limit 15 --output /tmp/pool.jsonl --append

Step 4: Triage — All-core / Partial / Irrelevant

After every round, classify each new paper and stamp a per-criterion mask (✓ / ~ / ✗) over the RUBRIC — no numeric scoring. The mask is what makes conjunctive queries ("papers satisfying A and B") terminate correctly: only a paper with ✓ on every [core] criterion is All-core. Emit a think_tool block:

TRIAGE round=<n>  query="<q>"
  All-core   (k): <paperId> "<title-≤60>" Y=<year> · [C1✓ C2✓ (C3~)]   every [core] ✓
                 C1: "<≤80-char quote>"
                 C2: "<≤80-char quote>"
  Partial    (k): <paperId> "<title>" Y=<year> · [C1✓ C2✗]              some [core] ✓
                 C1: "<≤80-char quote>"
  Irrelevant (k): <paperId> "<title>"                                   no [core] ✓, or trips disqualifier — drop

✓ = abstract/tldr clearly supports. ~ = partial / inferable. ✗ = no support or contradicts. [secondary] criteria don't set the tier but still get a mask symbol.

Rules:

1. Dedup across rounds by paperId first, then normalised title. Keep the stronger mask.
2. Disqualifier check beats all other matches → Irrelevant.
3. Re-diagnose gaps: note any [core] criterion with 0 ✓ across All-core+Partial, and any angle tag with 0 All-core/Partial → that's the next refine target (feeds the Step 3 gap→strategy map).
4. No fabrication: missing abstract → stamp ~, do not infer from training data.

The per-criterion quotes collected here are exactly what the Step 6 Rank-1 bar and the structured LIST template cite — do not skip them.

Snippet upgrade for borderline papers (abstract silent on a [core] criterion): batch-fetch real body text:

python scripts/snippet_search.py --query "<criterion phrase>" \
  --paper-ids "CorpusId:1,CorpusId:2,..." --limit 50

Step 5: Saturation Gate

After Probe and each round, decide CONTINUE vs STOP by whether key gaps remain — not by counting papers.

Early exit after Probe (single-recommendation / conjunctive queries only, K=1–2): for "is there a paper that …?", "recommend a paper", "what's the canonical X", or conjunctive "papers satisfying A and B" queries, if an All-core paper already covers every [core] criterion, STOP without running further rounds. This preserves the fast path where one probe hit settles a POINT-like query (≈2 queries), which matters under keyless S2 rate limits. For broader question shapes ("find papers about …" K=3–5, or 30+-paper ITERATIVE), do not use this early exit — apply the full STOP conditions below.

STOP when ALL hold:

• ≥1 All-core paper exists — a single paper with ✓ on every [core] criterion. This is required so conjunctive queries like "papers satisfying A and B" can't pass on two different papers that each cover only one side, AND
• every angle tag has ≥1 All-core/Partial paper, AND
• no key claim rests on a single source,
• OR further rounds stop surfacing anything new (empty recall / all duplicates).

CONTINUE to the next round (R3 / R4) otherwise, driven by the gap→strategy map:

• 0 All-core papers → fill the [core] criterion still ✗.
• An angle tag has 0 All-core/Partial → open that angle.
• A key claim rests on a single source → multi-source corroboration.

Re-decompose (rubric is wrong) if R2 returns 0 All-core AND 0 Partial across the board: report the strongest Partial candidate(s) + ask the user to relax a criterion.

Round caps: LIST and ITERATIVE up to 3 paper rounds (R2/R3/R4). POINT is a single fetch (no multi-round). If still not saturated at the cap, go to Step 6 and report which criteria / angle tags were not covered.

The gate is mechanical about gaps — do not skip rounds because "the results look right"; do not run extra rounds once the STOP conditions hold. The single-All-core-suffices shortcut applies only to the K=1–2 early exit above, not to broader queries.

Step 6: Rank and Output

Gather: every All-core and Partial paper from across all rounds (dedup by paperId). Drop Irrelevant.

Rank by relevance — the model judges, no numeric score. Order the gathered papers by how directly each answers the user's question: All-core before Partial; a paper satisfying every [core] criterion ranks above one satisfying only some. This is a judgment call, not a formula — do not compute a weighted_total.

Recency-aware tie-break. When the RUBRIC flagged a recency signal (cues defined in Step 3), break ties / near-ties in favor of the more recent paper (year DESC), applied after relevance. Recency rides behind relevance, never ahead of it.

K (soft ceiling — prevents pool-dumping):

K is a soft guide, not a formula. For broad "survey / categorize the field" queries the dominant failure mode is dumping the whole accumulated pool (often 30–50 papers, most off-criterion) into the output, burying the few on-criterion papers. Rank by relevance, keep the top K, and move the rest to an "Also relevant (not ranked)" list — never pad the ranked list with weak papers to reach a count.

Rank-1 quality bar. For single-recommendation queries ("is there a paper that …?", "recommend a paper", "what's the canonical X") the bolded top-1 must be an All-core paper — clearly satisfying every [core] criterion with a quote. Rank 1 carries disproportionate weight in user perception; fronting a Partial paper at top-1 reads as a confident wrong answer. If no paper clears the bar, lead with "No fully-matching paper found" and present the strongest near-miss honestly with its gaps.

If no All-core paper survives after the round cap, report "no fully-matching paper found", list strongest Partial candidates + their gaps, stop.

Output mode (caller inference). Pick the format by what the request demands, not by an explicit flag:

• Structured (formats below) — when the request demands machine-consumable output: a ranked list/table, per-criterion evidence, or hand-off to a downstream skill. Default for skill callers (research-survey, research-ideation, paper-writing, experiment-pipeline).
• Narrative (see "Narrative output" below) — the default for direct user calls: natural-language questions with no structured-output demand.

Structured output formats (skill callers):

LIST (shortlist with evidence):

**Top matches:**
- **<paperId>** "<Title>" — <Authors> et al., <Year>, <Venue>, cited by <N>. <URL>
  - C1 [core]: "<quote>"
  - C2 [core]: "<quote>"
  - C3 [secondary]: "<quote>"

**May also be relevant:**
- <paperId> "<Title>" — <Authors> et al., <Year>, cited by <N>. <URL> (Partial: only C3)

ITERATIVE (ranked table):

| # | Title | Authors | Year | Venue | Cited by | Link |
|---|-------|---------|------|-------|----------|------|
| 1 | …    | … et al. | 2024 | NeurIPS | 1234 | <URL> |

POINT: Paper Card (above).

Narrative output (direct user callers).

Deliver structured knowledge, not a search trace. Strip process words before output (Probe, R2/R3/R4, All-core/Partial, "rounds done") unless the user explicitly asks for a trace.

• Information-first, not list-first. Unless the user only wants a paper list, do not collapse the answer into "title + one-line contribution". Build the cognitive structure the user needs (timeline / topic grouping / comparison / mechanism breakdown / evidence verification / reading path), then place papers into it as evidence nodes.
• One main form + 2–3 auxiliary forms. The main form carries the answer's logic (timeline, topic grouping, comparison, mechanism breakdown, evidence grading, reading path, mini-survey); 2–3 auxiliary forms (paper card, table, evidence grading, annotated bibliography, reader payoff) aid readability. Do not stack every form.
• Intent → form (condensed): latest/SOTA → status-judgment + table; origin/foundation → timeline + source-paper analysis; A-vs-B → conclusion + dimension comparison; mechanism → mechanism breakdown + evidence interleaving; benchmark/data → verification + evidence table; landscape → topic grouping + reading path.
• Citation rules: a core paper shows Title (Venue Year) on first mention (never author-only like Zhang et al.); every cited paper carries its returned [N]. Tables MAY contain [N] markers. At the end of the answer, list every cited paper by number in IEEE style (with citation count appended): [N] A. Author et al., "Title," Venue, Year, cited by N. [Online]. Available: URL. IEEE rules: author names as Initial. Surname (e.g. A. Vaswani); join multiple authors with commas and and; ≥3 authors → A. Firstauthor et al.; title in double quotes; then Venue, Year; append cited by N (the tool's citationCount); end with [Online]. Available: URL for the link. Omit a field only if the tool genuinely did not return it (never fabricate); authors and citation count must appear whenever returned.
• Length: single-point 600–1200 words; comparison/retrieval 1200–2500; landscape/timeline 2000–3500. Do not sacrifice evidence structure for brevity.
• Quote-or-zero still applies — every claim a paper is used to support is still backed by a ≤80-char quote (Red Line 5, anti-hallucination); the narrative just renders it as [N] instead of showing the raw quote.

Pre-output checklist (mandatory). Before emitting the answer, verify each box.

• Pool gathered from every round's triage, deduped by paperId, Irrelevant excluded.
• Ranked by relevance (judgment, not a numeric score) — All-core before Partial.
• Recency tie-break applied when a recency signal is present (year DESC after relevance).
• Rank-1 clears the bar for single-recommendation queries (clearly satisfies every [core] criterion with a quote) — or you've reported "No fully-matching paper found".
• Every cited paper has ≥1 supporting quote for the claim it's used for (quote-or-zero, Red Line 5 — anti-hallucination).
• Output ≤ K (soft cap); surplus relevant papers sit in "Also relevant (not ranked)", not the ranked list.
• Narrative mode only — every cited paper appears in the end-of-answer numbered reference list in IEEE style with authors + citation count ([N] A. Author et al., "Title," Venue, Year, cited by N. [Online]. Available: URL), and the body strips process words (Probe / R2 / All-core/Partial).

If any box is unchecked, return to Step 6 — do not output.

Tool Cheat Sheet

All discovery scripts: --limit N, --json, --output FILE, --append; accept S2 / arXiv / DOI / CorpusId IDs. --output --append auto-dedupes by paperId across rounds (within-batch + cross-file), so the pool stays clean.

Rate limits

Global S2 pacer + circuit breaker (5 failures → 60s cooldown). Retries: 3s / 6s / 12s / 24s / 48s.

Without S2APIKEY: use scholarsearch (arXiv fallback) + arxivmonitor. Skip citationtraverse / recommend / snippetsearch — they're S2-only; do not retry.

References

References are self-contained. Don't chain between them — return here to re-route.

Hand off to