A newly published paper has no incoming citations, but it does have a reference list. Scientific retrieval systems that use the citation graph mostly ignore that asymmetry. They either score each paper as an isolated piece of text, or they pass information along citation edges in a way that favors the most-cited papers. We propose to place each new paper in the graph through the papers it cites. Each reference is weighted by how specifically that reference belongs to a research topic. A cited classic that every field cites says almost nothing about where the new paper belongs. A paper cited heavily by one subcommunity, but little elsewhere, says a great deal. We call this weight topic lift. It is a per-topic version of popularity correction. It discounts generic hubs while still rewarding papers that are authorities within one domain.
The opportunity follows from three findings. Graph-aware retrieval helps scientific question answering. Graph message passing, the step in which each node updates its representation by averaging its neighbors', amplifies popularity. Standard popularity fixes treat all high-degree nodes alike, so they penalize genuine niche experts along with celebrities. Our claim is that one quantity resolves all three. It is the pointwise mutual information between a cited paper and the topic of the papers that cite it. This quantity can be computed for a paper published this morning, and it should beat both text-only retrieval and globally degree-normalized graph retrieval on the papers that matter most for literature search: recent ones nobody has cited yet.
Core Claim
Let $\theta_i \in \mathbb{R}^K$ be the topic mixture of paper $i$, inferred from its text. Let $P(k \mid j)$ be the fraction of citation mass reaching paper $j$ from citers in topic $k$, where each citer contributes its own topic mixture as soft counts. Let $P(k)$ be topic $k$'s share of all citation mass. The weight on the edge from a new paper $i$ to a reference $j$ is
$$w_{ij} = \max\Big(0,\ \sum_{k=1}^{K} \theta_{ik}\,\log\frac{P(k \mid j) + \epsilon}{P(k) + \epsilon}\Big),$$
where $\epsilon$ is a smoothing constant for references with few citations. The new paper's graph representation is $g_i = \sum_j w_{ij} h_j / \sum_j w_{ij}$, where $h_j$ is the existing representation of reference $j$.
The same lift, evaluated at the query's topic mixture rather than $\theta_i$, replaces raw citation count wherever a graph retriever would otherwise use authority. That substitution stops old celebrity papers from crowding out uncited candidates.
The claim has a sharp contrast with inverse-degree correction. Take two references with equal citation counts, one cited uniformly across fields and one cited mostly from a single subfield. Inverse-degree weighting treats them identically. Topic lift separates them.
Relation to Prior Work
Three papers motivate the proposal by showing the gap. CG-RAG builds a graph of text chunks linked by citations and uses a Graph Transformer to mix BM25 and dense scores across neighbors. It reports large gains, but it operates over a fixed graph and applies no popularity correction. How Graph Convolutions Amplify Popularity Bias for Recommendation? proves that a node's pull on the learned parameters grows as roughly $d^{3/2}$ in its degree $d$, so graph convolutions amplify popularity rather than merely reflect it. Together these imply that extending CG-RAG to new papers would steer them toward famous neighbors. SciNet shows that current literature agents miss most disruptive work and rarely connect papers into valid citation chains. It also shows that supplying ground-truth citation structure improves synthesis quality by 25.3%.
Two lines constrain the design. Global correction is the first. RawlsGCN makes the graph operator doubly stochastic, meaning every row and column sums to one, which equalizes each node's influence. LinkBERT samples linked documents in inverse proportion to in-degree during pretraining. PageRank, HITS and a unified framework for link analysis normalizes each node's outgoing votes. All three define popularity by degree alone. They are the exact approaches our ablation must beat, and they explain why a uniform penalty is the default.
Domain-specific authority is the second line. An Influence-Receptivity Model for Topic based Information Cascades gives each node a per-topic influence vector estimated from cascade timestamps. A Framework for Modeling Influence, Opinions and Structure in Social Media shows that globally central nodes often lack authority in niche communities. Both establish that authority is topic-dependent. Neither can score a node with no observed history, which is the case we target. PReP is the closest conceptual relative. It separates a node's background visibility from true relatedness in typed networks. However, it is transductive and cannot score entities with no paths, whereas we exploit the outgoing reference edges that new papers already have.
On the cold-start side, CaseLink shows that unseen legal cases can be placed inductively through statute nodes, with a regularizer against over-connection. That reliance on a curated statute taxonomy is precisely what scientific corpora lack. We substitute text-inferred topics plus references. PAM and DisCo handle cold items by leaning on content features and by keeping semantic and behavioral spaces distinct. DisCo's finding that forcing text and interaction signals into one space is harmful leads us to keep the text score and the graph term as separate inputs to the ranker, not a fused embedding. Finally, LitSearch finds that indexing full text hurts retrieval. We therefore use only titles, abstracts and resolved reference lists, never body text.
Research Program
We will build the method in three pieces.
The first is a topic model over abstracts. We will cluster dense abstract embeddings into $K$ soft topics, with $K$ swept, and compare the result against LDA. The aim is to check that the method does not hinge on one topic extractor.
The second is a topic-lift table computed from citation edges available before a time cutoff. The table is refreshed incrementally, because each new citation updates only $P(k \mid j)$ for the cited paper.
The third is a retriever whose final score is a learned combination of a text relevance score and a graph score. The text score comes from the same dense encoder used by the strongest text baseline. The graph score is the similarity between the query and the topic-lift aggregate $g_i$, plus a query-conditioned authority term. Candidates are thus scored by their text and by where their references place them, and authority is measured relative to the query's topic rather than globally.
The analysis has two parts. The first stratifies results by the target paper's citation count, so that gains on uncited papers cannot hide losses on established ones. The second examines interdisciplinary papers, where a single dominant topic is wrong and soft mixtures must carry the load. This is the likeliest failure mode, and we will report it separately.
Baselines
The naive reference point is text-only dense retrieval over titles and abstracts. This means GritLM-7B, which LitSearch reports as its top model with 74.8% recall@5 on specific questions, alongside BM25. Cold-start papers are, after all, just text, so this baseline is the real bar. It will be run with the same encoder our method uses for its text score, the same corpus, and the same index. We count as beating it a statistically significant gain in recall@5 on cold-start targets, with no significant loss on warm targets.
The strongest existing graph method is a CG-RAG-style retriever. CG-RAG reports Hit@1 of 0.961 against ColBERT's 0.913 on PubMedQA-1k augmented with a 7,849-paper citation graph, and QA accuracy of 0.835 against 0.769 for the best hybrid on PapersWithCodeQA. To make it inductive, we will let new papers connect through their outgoing references using its standard neighbor aggregation. It will receive the same edges, the same text encoder, and the same hyperparameter search budget as our method. We aim to reproduce its reported PubMedQA-1k number before comparing. Beating it means higher recall on cold-start targets and a lower share of top-10 slots taken by papers in the top citation decile, at equal or better overall Hit@1.
Global popularity correction is represented by doubly stochastic normalization from RawlsGCN, inverse-in-degree weighting as in LinkBERT, and post-hoc subtraction of the popularity direction as in DAP from How Graph Convolutions Amplify Popularity Bias for Recommendation?. Each is swapped in for topic lift inside our pipeline, with everything else held fixed.
The ablation that isolates the key ingredient sets $K=1$, which removes topics and leaves only degree information. We also compare against uniform reference weights. If topic lift does not beat inverse degree on references matched for citation count, the core claim fails.
Why not just…?
Why not just rely on a strong text encoder, since new papers are fully described by their abstracts? This is the most serious threat, given GritLM's 74.8% recall@5 on LitSearch. We will report results separately for queries where the text retriever already succeeds and for those where the target ranks outside its top 20. Topic lift is only justified if it recovers a meaningful share of the second group. If the gain concentrates in no subgroup, we will report that the graph adds nothing for cold-start papers.
Why not just divide by citation count? Inverse-degree weighting, used by LinkBERT and in spirit by RawlsGCN, penalizes a subfield's foundational paper exactly as much as a field-wide classic. The influence-receptivity evidence of An Influence-Receptivity Model for Topic based Information Cascades says that this distinction carries signal. The $K=1$ ablation on degree-matched references tests this directly.
Why not just append the titles of a paper's references to its abstract and embed that? This simple approach could plausibly capture much of the benefit. It is also exposed to the dilution LitSearch documents for full text, because generic classics in the list add noise. We will run it with the same encoder. Topic lift must beat it, especially on papers whose reference lists are dominated by widely cited works.
Why not just boost recent papers? A recency prior raises exposure for all new papers regardless of relevance. We will include it as a baseline and measure precision among the boosted papers, not only their exposure.
Proposed Experiments
We will build temporal cold-start splits on three corpora. On LitSearch (64,183 ACL and ICLR papers, 597 queries; code at princeton-nlp/LitSearch), citation edges come from S2ORC. For every target published after a cutoff, all incoming citations are deleted, while its reference list is kept. We apply the same protocol to the PubMedQA-1k citation graph and to PapersWithCodeQA from CG-RAG. The ego-centric tasks of SciNet, where reported recall@50 is below 5%, serve as a stress test for surfacing non-famous work.
Metrics are recall@5 and recall@20 stratified by citation tier, Hit@1 for the question-answering sets, and the share of top-10 slots held by top-decile papers relative to their share among relevant documents.
Two checks are decisive. The first: on cold-start targets the text retriever misses, topic lift must beat both the CG-RAG-style inductive baseline and the appended-reference-titles baseline. The second: on references matched for citation count, swapping topic lift for inverse degree must measurably lower recall. If either check fails, the specific claim about domain-conditioned popularity correction does not hold, whatever the aggregate numbers show.
References
- CG-RAG: Research Question Answering by Citation Graph Retrieval-Augmented LLMs
- How graph convolutions amplify popularity bias for recommendation?
- SciNet: Evaluating AI Agents in Relation-Aware Scientific Literature Retrieval
- RawlsGCN: Towards Rawlsian Difference Principle on Graph Convolutional Network
- LinkBERT: Pretraining Language Models with Document Links
- PageRank: HITS and a Unified Framework for Link Analysis
- An Influence-Receptivity Model for Topic Based Information Cascades
- A Framework for Modeling Influence, Opinions and Structure in Social Media
- PReP: Path-Based Relevance from a Probabilistic Perspective in Heterogeneous Information Networks
- CaseLink: Inductive Graph Learning for Legal Case Retrieval
- Online Item Cold-Start Recommendation with Popularity-Aware Meta-Learning
- DisCo: Towards Harmonious Disentanglement and Collaboration between Tabular and Semantic Space for Recommendation
- LitSearch: A Retrieval Benchmark for Scientific Literature Search
- S2ORC: The Semantic Scholar Open Research Corpus