Paper Collection Process

→

STAGE 1 · CANDIDATE GENERATION

recall first
1.1 · LLM RETRIEVAL AGENT

REFINE THE AREA CARD

An agent repeatedly searches, inspects the retrieved evidence, and revises the next search.

LLM
Retrieval Agentconstructs Ci, examines results, and revises Ci+1
Area Card Cifirst-order componentslexical + semantic queriesexclusions + suspected gapssemantic cutoffs
C₁ → C₂ → C₃   ⟹   R₃
1.2 · CITATION EXPANSION

EXPAND THE CANDIDATES

Diverse seeds expose older foundations and later extensions missed by direct retrieval.

← references
older work
citing papers →
later work
R = R₃ ∪ E₁ ∪ E₂
→

STAGE 2 · CONTRIBUTION FILTERING

precision recovery
2.1 · LEARNED CHEAP FILTER

REMOVE OBVIOUS NON-MEMBERS

Train complementary relevance models, fuse their scores, and calibrate a conservative cutoff.

Training judgment samplesampled contribution labels
Calibration sampleheld out from training
train two relevance views
lexical
model
embedding
model
score fusion
choose operating point
Calibrated conservative cutoff
preserve likely-relevant and uncertain papers
fused score + threshold τ ↓
Filtered population RF ⊆ R
2.2 · LLM CONTRIBUTION JUDGE

JUDGE AND PUBLISH

An LLM agent applies the Area Brief directly to every survivor; scores do not decide membership.

paper p
main contribution
Area Brief B
scope contract
LLM
LLM Contribution Judgecompares p with B and records the disposition
include
Area
include
component
exclude /
unresolved
Area-specific set PAaccepted papers + Contribution Tags