Most retrieval systems score each paper or passage as if it stood alone. When systems do use the citation graph, they usually rank with a random-walk prestige score such as PageRank, which rewards papers that many others cite. That reward has a known failure. On the SciNet benchmark, the best retrieval agents find relevant papers with over 60% consistency, but only about 14% of the multi-hop chains they return are actually connected. A citation chain of this kind ("which line of work led from method A to method B") is a lineage path. We argue that two cheap signals already present in the data are being thrown away. The first is where in the citing paper a citation appears: a reference in the methods section usually means "we build on this," while one in the introduction usually means "this is background." The second is the shape of a paper's neighborhood across one, two and three citation hops, which separates a focused cluster member from a hub that everyone cites in passing.
The proposal turns both signals into a single unsupervised ranking walk over the citation graph. Citation edges are typed by the section they appear in, and walking along each type is weighted separately. The walk's restart distribution is corrected by an entropy-based neighborhood signature from which the part explained by citation count has been removed. In two sentences: follow methods-section citations more than introduction citations, and stop favoring papers just because they are heavily cited. The concrete test is whether doing so closes part of SciNet's path-connectivity gap without any neural message passing.
Core Claim
The claim is that the section in which a citation occurs is the main missing edge attribute for lineage retrieval, and that walking the graph with section-specific weights recovers connected multi-hop paths far better than a flat citation graph. A secondary, separately testable claim is that a multi-scale ego-network entropy signature (a measure of how spread out or concentrated a paper's 1-, 2- and 3-hop neighborhood is) adds information beyond what degree already gives, once degree has been regressed out. Correcting for degree alone is the obvious cheaper fix. We therefore design the entropy component so that it can only help if it carries signal that degree does not, and we report it as falsified if it does not.
Relation to Prior Work
The closest existing work is MRGSEM-Sum. It builds a multi-relational sentence graph, uses structural entropy minimization to choose clusters, and applies lead bias to compress multi-document summaries. Three choices separate us from it. It works on sentences inside a summarization input set, not on a corpus-scale citation graph. It applies position as a post-hoc heuristic rather than as an edge type. It uses entropy for clustering rather than as a node signature in ranking. PositionRank supplies the template we extend: it feeds a $1/pos$ position prior into PageRank's restart distribution over a single document's word graph and reports a 29.09% relative MRR gain on the Nguyen keyphrase set. We move the positional signal from nodes within one document onto edges between documents, and we replace a fixed decay with weights per rhetorical section. Prior positional schemes such as DeepRank and the legal tail bias of CaseSummarizer motivate this choice: they show that a fixed decay does not transfer across genres.
On the graph-retrieval side, CG-RAG builds chunk-level citation graphs and trains a Graph Transformer to combine lexical and dense scores. Asai et al. train a recurrent retriever to beam-search Wikipedia hyperlinks. LitFM counters popularity bias with LLM-generated pseudo-queries and topic-controlled reranking. All three are supervised or LLM-dependent, and none distinguishes citations by where they occur. They are our strongest competitors, not our method. The entropy signature reuses the quadratic, linear-time approximation of von Neumann entropy from VNEstruct, which was shown to compete with GIN on MUTAG at a fraction of the cost. Depth-based traces (Bai and Hancock) motivate computing that entropy at several radii rather than one.
MrSE acts as a constraint. On ACM, NMI drops from 48.36% to 3.16% when a multiplex graph is flattened. That result argues for keeping section-typed edges as separate relations rather than summing them into one weight, so we keep per-type transition matrices. Finally, Tixier et al. show that high-degree nodes are often non-cohesive hubs and propose $k$-core filtering. This is a direct competitor to our entropy correction and appears as a baseline.
Research Program
Building the section-typed graph. We build a citation multiplex over papers with available full text. Each citation occurrence is assigned one of a small number of section types: introduction/background, related work, methods, experiments/results, and discussion. Types come from parsed section headers. When headers are missing, we fall back to the occurrence's relative position in the document, bucketed into five bins. One pair of papers can be linked by several typed edges.
Defining the walk. The ranking is a personalized random walk. From paper $u$, the walk moves to cited paper $v$ with probability proportional to $\sum_s w_s\, c_s(u,v)$. Here $c_s(u,v)$ is the number of times $u$ cites $v$ in section type $s$, and $w_s \ge 0$ is a per-type weight. The walk also traverses edges in reverse, with separate weights, so that it can reach later work. With some probability it restarts at a query-seeded distribution.
Correcting the restart distribution. For each paper $v$, we compute the approximate von Neumann entropy $S_r(v)$ of its $r$-hop ego network for $r \in \{1,2,3\}$, where the ego network at radius $r$ is the subgraph within $r$ citation hops of $v$. For $r=3$ on large graphs, we use sampled neighborhoods. We regress each $S_r$ on log-degree across the corpus and keep the residual $\rho_r(v)$, the part of the neighborhood shape that citation count does not explain. The restart distribution then becomes
$$\pi(v) \propto q(v)\,\exp\Big(\sum_{r} \beta_r\, \rho_r(v)\Big),$$
where $q(v)$ is a lexical-plus-dense relevance score of $v$ for the query and the $\beta_r$ are a handful of scalars.
Tuning budget. The $w_s$ and $\beta_r$ total about a dozen parameters. They are set by grid search on a small development split, and every baseline receives the same tuning budget. Path reconstruction follows the walk's highest-probability nodes and links them with a beam search, scored by the product of edge probabilities.
Analysis. Beyond end-task metrics, we will measure two things. First, how the probability mass that the tuned walk sends through each section type differs between lineage queries and topical queries. Second, whether degree-residualized entropy separates gold path nodes from non-path nodes, measured by AUC. We will also check how much of that separation $k$-core number already provides.
Baselines
The naive reference point is dense-plus-lexical retrieval with no graph at all. It uses the same $q(v)$ that seeds our walk, here a hybrid of BM25 and a dense retriever. This isolates what any graph use buys. The SciNet paper reports that current agents fall below 5% Recall@50 on disruptive-paper retrieval and reach roughly 14% connectivity on path-wise tasks, and these are the reference points we must clear on the path-wise subset.
The graph-only reference is personalized PageRank over the flat, unweighted citation graph, seeded by the same $q$. The flat graph is equivalent to setting all $w_s$ equal and all $\beta_r = 0$. Our method beats this baseline only if the tuned section weights and the entropy term improve path connectivity at matched recall of individual relevant papers.
The strongest cheap anti-popularity competitors are three variants of that walk. The first applies a degree-normalized restart (restart probability divided by $\deg(v)^\gamma$, with $\gamma$ tuned). The second restricts the walk to the subgraph above a tuned $k$-core threshold, following Tixier et al.. The third places a PositionRank-style $1/pos$ weight on each citation occurrence in place of section types. Each gets the same grid-search budget as our parameters. These baselines isolate the two key ingredients. If the $1/pos$ edge weight matches section types, then the rhetorical role adds nothing beyond position. If degree normalization or $k$-core filtering matches the entropy term, then the entropy signature is redundant.
The strongest existing supervised system is CG-RAG, which reports Hit@1 = 0.961 on PubMedQA-1k against 0.913 for ColBERT, and 0.835 accuracy on PapersWithCodeQA against 0.769 for the best hybrid baseline. On those two benchmarks we will compare against these reported numbers using the same corpora and splits, and we will rerun CG-RAG if code is available. We do not expect an unsupervised walk to beat a trained Graph Transformer on PubMedQA-1k, where the score is already near ceiling. On these benchmarks, success means coming within about two points of Hit@1 at substantially lower training cost. On PapersWithCodeQA, whose questions more often hinge on which method built on which, we expect to outperform the 0.769 hybrid baseline and to close part of the gap to 0.835. For LitFM, no result was found on Lacuna for SciNet or our QA benchmarks. We include its diversity-guided retriever only if it can be reproduced under our compute budget.
Why not just…?
Why not just use a strong dense retriever or ColBERT and skip the graph? A dense retriever scores papers one at a time, so it cannot return a connected chain except by accident. SciNet's finding that agents are consistent but disconnected is exactly this failure mode. CG-RAG's improvement over ColBERT on PubMedQA-1k (0.961 versus 0.913 Hit@1) shows that graph context helps even on single-answer questions. Our no-graph baseline measures the gap directly on path tasks, where we expect it to be largest.
Why not just put all candidate papers into a long-context LLM? The LLM still has to decide which links are real, and LitFM reports that GPT-4o hallucinates citations and trails a structure-aware model by 8.0% on Medicine citation link prediction. We will test this directly by giving a long-context model the top-$N$ papers from our seed retriever and asking it to reconstruct paths. Connectivity is scored against the true graph. If the LLM matches our walk at comparable cost, the walk is not justified for path tasks.
Why not just divide PageRank by degree, or filter by $k$-core, instead of computing entropy? This is the real threat to the entropy component. At radius 1, the VNEstruct approximation is itself largely a function of degrees, so we only use degree-residualized entropy at radii 2 and 3. Even so, the residual may carry nothing that $k$-core number lacks. The degree-normalized and $k$-core baselines, with equal tuning, settle the question. If they match the entropy term, we drop it and report the section-typed walk as the contribution. The section-typing claim does not depend on the entropy result.
Why not just compute the disruption index directly for SciNet's disruptive-paper task? SciNet's ground truth for that task is defined by a local citation-neighborhood statistic. Any method with graph access can compute that statistic directly, which would make a win on that task close to circular. We therefore report the direct disruption-index computation only as a sanity ceiling and make no claim on the node-level disruption task. Our claims rest on path-wise connectivity and research QA.
Proposed Experiments
The primary experiment uses SciNet's path-wise queries, restricted to those whose gold path papers have full text with parseable references (for example, open-access biomedical and arXiv subsets). We report coverage explicitly, because a subset biased toward open-access fields could inflate gains. The metrics are SciNet's connectivity and consistency, plus recall of individual gold papers at 50, so that connectivity gains cannot come from returning fewer, safer papers. A second experiment runs the same walk on the PubMedQA-1k and PapersWithCodeQA citation corpora used by CG-RAG, reporting Hit@1, accuracy and query latency against CG-RAG's reported 404 ms.
We will report three decisive checks. The first is a factorial ablation, section typing (off, $1/pos$, typed) crossed with restart correction (none, degree-normalized, $k$-core, entropy residual), all at equal tuning budget. Its outcome tells us which ingredient carries the gain. The second is a transfer test. Section weights tuned on biomedical queries are applied unchanged to computer-science queries, and vice versa. If the weights must be retuned per field, the rhetorical-role claim is weaker than stated and closer to the genre-specific decay heuristics it aims to replace. The third is a robustness test in which section labels are replaced by the position-bin fallback for all papers. This shows how much the method depends on clean full-text parsing, which is its main practical limitation.
The idea is supported if typed edges raise path connectivity well above the flat-walk baseline and SciNet's reported ~14% at matched paper recall, and if the gain survives cross-field transfer. The idea fails if $1/pos$ edge weights match typed edges and degree normalization matches the entropy term. In that case the result is a useful negative: plain positional and degree heuristics are enough, and lineage retrieval needs neither rhetorical roles nor entropy signatures.
References
- SciNet: Evaluating AI Agents in Relation-Aware Scientific Literature Retrieval
- MRGSEM-Sum: An Unsupervised Multi-document Summarization Framework based on Multi-Relational Graphs and Structural Entropy Minimization
- PositionRank: An Unsupervised Approach to Keyphrase Extraction from Scholarly Documents
- DeepRank: A New Deep Architecture for Relevance Ranking in Information Retrieval
- CaseSummarizer: A System for Automated Summarization of Legal Texts
- CG-RAG: Research Question Answering by Citation Graph Retrieval-Augmented LLMs
- Learning to Retrieve Reasoning Paths over Wikipedia Graph for Question Answering
- LitFM: A Retrieval Augmented Structure-aware Foundation Model For Citation Graphs
- Ego-based Entropy Measures for Structural Representations on Graphs
- A Graph Kernel from the Depth-Based Representation
- Multi-Relational Structural Entropy
- GoWvis: A Web Application for Graph-of-Words-based Text Visualization and Summarization