SciAtlas: Panoramic Science Network

• SciAtlas is a large-scale, multi-disciplinary academic knowledge graph that integrates 43.30M papers, 157M entities, and 3B triplets to create a navigable scientific evolution network.
• It employs a neuro-symbolic retrieval framework combining keyword, semantic, and title matching with graph reranking to surpass traditional keyword-based searches.
• SciAtlas supports diverse scientific tasks, including literature review, trend synthesis, and academic trajectory exploration for both human researchers and AI agents.

SciAtlas is a large-scale, multi-disciplinary, heterogeneous academic resource knowledge graph designed as a panoramic scientific evolution network. It is presented as a response to the contemporary “information explosion,” in which scientific knowledge is abundant but fragmented, weakly organized, and difficult to navigate through structural reasoning. By integrating 43.30M papers from 26 disciplines, together with 157M entities and 3B triplets, SciAtlas is intended to provide a topological cognitive substrate for automated scientific research, allowing retrieval to move beyond keyword overlap or vector similarity toward deterministic association discovery (Qiao et al., 20 May 2026).

1. Problem setting and conceptual definition

SciAtlas is motivated by a diagnosis that current academic retrieval systems are limited in three different ways. Traditional keyword matching is fast and interpretable, but it is shallow and fails when the relevant literature is conceptually related rather than lexically similar. Dense or vector retrieval improves semantic matching, but remains a feature-space similarity operation that does not explicitly model citation topology, author and institution structure, disciplinary hierarchy, or multi-hop associations across heterogeneous entities. Agentic deep-research frameworks attempt to compensate for this, but are described as prone to logical hallucinations and high inference cost / response latency (Qiao et al., 20 May 2026).

Within this framing, SciAtlas is not merely a citation graph and not merely a semantic search layer over paper embeddings. It is defined as a paper-centric but heterogeneous scholarly graph that connects papers through semantic, conceptual, disciplinary, and social pathways simultaneously. The stated goal is to dismantle the “knowledge island” phenomenon by making the latent structure of science explicit and traversable. The resulting system is intended for both human researchers and AI agents, especially in workflows involving literature retrieval, synthesis, idea positioning, author discovery, and trajectory analysis (Qiao et al., 20 May 2026).

The paper’s conceptual claim is that scientific automation requires a structured world model rather than stronger language modeling alone. In that sense, SciAtlas is proposed as an infrastructural layer: a graph-based representation of scientific evolution that can anchor retrieval and reduce reliance on unguided multi-step LLM search (Qiao et al., 20 May 2026).

2. Graph corpus, entity schema, and construction pipeline

SciAtlas spans 26 disciplines. The largest reported disciplinary shares are Medicine: 18.56%, Social Sciences: 10.70%, Engineering: 9.43%, Biochemistry, Genetics and Molecular Biology: 6.44%, and Computer Science: 6.29%; these five together account for 51.43% of total paper volume (Qiao et al., 20 May 2026).

The graph contains 9 categories of entity nodes. The reported counts are as follows.

Entity type	Count
Paper	43.30M
Author	109.70M
Keyword	3.76M
Institution	0.12M
Topic	4.52K
Subfield	252
Field	26
Source	0.28M
Domain	4

The relation inventory includes (Paper, CITES, Paper): 213.88M, (Paper, HAS_KEYWORD, Keyword): 101.38M, (Paper, HAS_TOPIC, Topic): 105.89M, (Author, AFFILIATED_WITH, Institution): 195.94M, (Author, AUTHORED, Paper): 149.00M, (Author, COAUTHOR, Author): 2.06B, (Keyword, COOCCUR, Keyword): 60.37M, (Field, DOMAIN_OF, Domain): 26, (Paper, RELATED_TO, Paper): 68.38M, (Subfield, FIELD_OF, Field): 252, (Topic, SUBFILED_OF, Subfield): 4.52K, and (Paper, PUBLISH_IN, Source): 40.90M. The paper notes a minor inconsistency in its own schema description, stating 11 relationship types in one place and 12 categories of relational edges elsewhere (Qiao et al., 20 May 2026).

The primary data source is OpenAlex, described as an open scholarly resource library with over 480M academic publications. Construction proceeds by extracting entity types and attributes, normalizing and deduplicating many names after standardization, filtering out non-English papers and papers with very short abstracts, and discarding entities missing critical attributes, especially papers without PDF URLs. The authors explicitly state that they do not deduplicate authors because of name ambiguity (Qiao et al., 20 May 2026).

A major design choice is the keyword layer. Rather than relying on OpenAlex Concept entities as the main conceptual substrate, SciAtlas generates denser conceptual nodes using Qwen3-30B-A3B-Instruct-2507. For each paper abstract, the model outputs 3–8 core keywords, each with an importance score, and the prompt explicitly prefers canonical, reusable, high-level keywords over paper-specific branding or verbose phrases. Keywords co-occurring in the same paper are connected by COOCCUR edges, with co-occurrence frequency as edge weight (Qiao et al., 20 May 2026).

SciAtlas also stores embeddings for paper title, paper abstract, and keyword text, using bge-large-en-v1.5. The vector indexes are 1024 dimensions with cosine similarity. Deployment is in Neo4j, with full-text indexes over paper title and abstract and vector indexes for title, abstract, and keyword embeddings (Qiao et al., 20 May 2026).

The graph is organized at four stated levels. At the semantic level, CITES and RELATED_TO link papers directly. At the conceptual level, HAS_KEYWORD and keyword COOCCUR connect papers through reusable concepts. At the direction level, Domain → Field → Subfield → Topic imposes disciplinary hierarchy. At the social level, AUTHORED, COAUTHOR, and AFFILIATED_WITH connect papers through researchers and institutions. This layered organization is central to the system’s claim of supporting cross-disciplinary and multi-hop reasoning (Qiao et al., 20 May 2026).

3. Neuro-symbolic retrieval and graph reranking

The main algorithmic contribution of SciAtlas is a neuro-symbolic retrieval framework based on tri-path collaborative recall + graph reranking. The three recall paths are keyword matching, semantic matching, and title matching (Qiao et al., 20 May 2026).

In the keyword path, an LLM extracts a set of query keywords with normalized importance scores,

$\mathcal{K}=\{(k_i, s_i^\text{LLM})\}_{i=1}^m,$

where $k_i$ is the $i$-th extracted keyword and $s_i^\text{LLM} \in [0,1]$. Exact matches are scored by

$\text{score}_{exact}(k_i, g)=s_i^\text{LLM},$

while vector matches use cosine similarity above a threshold $\theta_{kw}=0.7$,

$$\text{score}_{vec}(k_i, g)=s_i^\text{LLM}\cdot\text{sim}(k_i, g).$$

The final seed weight for matched keyword node $g$ is

$$w_g^{kw}=\max_i \left( \mathbb{1}[k_i=g]\cdot s_i^\text{LLM}, \mathbb{1}[\text{sim}(k_i,g)\ge \theta_{kw}]\cdot s_i^\text{LLM}\text{sim}(k_i, g)\right).$$

Only the top-3 vector matches per keyword are retained (Qiao et al., 20 May 2026).

In the semantic path, the query embedding retrieves top-60 papers by title embeddings and top-60 by abstract embeddings, after which bge-reranker-large reranks them and the system keeps top-15 for title and top-15 for abstract. The paper states that title and abstract are weighted 0.4 and 0.6, respectively (Qiao et al., 20 May 2026).

In the title path, titles are extracted from the input paper and its references using GROBID. The LLM assigns a confidence $c_j$ to each title $k_i$0, keeps the top-10, and computes a fuzzy similarity

$k_i$1

where $k_i$2 is based on LCS and $k_i$3 is Jaccard overlap over tokens. Matches below $k_i$4 are discarded, and at most top-5 papers are retained per title (Qiao et al., 20 May 2026).

Semantic and title candidates are then merged into a unified paper seed score,

$k_i$5

with defaults $k_i$6 and $k_i$7, and

$k_i$8

The system then performs 2-hop subgraph propagation, treating edges as undirected during propagation and keeping at most 500 nodes per hop per entity type (Qiao et al., 20 May 2026).

Paper importance is defined as

$k_i$9

where $i$0 is the paper citation count and $i$1 is total citations in the subgraph. Seed paper weight is

$i$2

with default $i$3. The initial seed distribution is

$i$4

This initializes graph propagation over a typed, weighted local subgraph (Qiao et al., 20 May 2026).

Edge weights are relation-specific. For example, HAS_KEYWORD uses

$i$5

with default $i$6, while CITES, RELATED_TO, and AUTHORED use default coefficients $i$7, $i$8, and $i$9. COAUTHOR and COOCCUR further scale by capped logarithmic functions of collaboration and co-occurrence counts (Qiao et al., 20 May 2026).

Graph reranking is implemented as Random Walk with Restart. For node $s_i^\text{LLM} \in [0,1]$0, transition probability is

$s_i^\text{LLM} \in [0,1]$1

With $s_i^\text{LLM} \in [0,1]$2, the iteration is

$s_i^\text{LLM} \in [0,1]$3

and stops when

$s_i^\text{LLM} \in [0,1]$4

with $s_i^\text{LLM} \in [0,1]$5 or $s_i^\text{LLM} \in [0,1]$6. Final paper ranking uses

$s_i^\text{LLM} \in [0,1]$7

with defaults $s_i^\text{LLM} \in [0,1]$8, $s_i^\text{LLM} \in [0,1]$9, $\text{score}_{exact}(k_i, g)=s_i^\text{LLM},$0, and

$\text{score}_{exact}(k_i, g)=s_i^\text{LLM},$1

The system returns top-20 papers, along with score breakdowns and path-based explanations (Qiao et al., 20 May 2026).

4. Modes of scientific use

SciAtlas is presented as infrastructure for several downstream tasks rather than as a single retrieval engine. The paper highlights literature review, automated research trend synthesis, idea positioning, idea grounding, idea evaluation, idea generation, related author retrieval, and academic trajectory exploration (Qiao et al., 20 May 2026).

For literature review, SciAtlas retrieves a relevant paper set for a research direction and can then pass that set to an LLM-based survey generator. The stated advantage is topology-aware retrieval with configurable emphasis on paper importance, venue prestige, author authority, or institution reputation (Qiao et al., 20 May 2026).

For trend synthesis or prediction, the graph retrieves representative papers in a direction, often with stronger citation weighting, after which papers are sorted chronologically and summarized into stage-wise accounts of methodological shifts, emerging topics, open gaps, and future directions. The paper gives an example on biologically plausible learning in spiking neural networks, dividing the area into four temporal stages from 2006–2025 (Qiao et al., 20 May 2026).

For idea grounding, the system takes an idea or paper draft, retrieves highly relevant papers, and then supports paragraph-level evidence retrieval using refined LLM-generated claims or queries. The paper’s example uses InnoEval, grounding a claim about the weakness of LLM-as-a-Judge against supporting evidence from another paper. For idea generation, the authors suggest relaxing constraints on distant nodes so that retrieval becomes more exploratory and diverse; the example given is “Federated and Privacy-Preserving Knowledge Editing” (Qiao et al., 20 May 2026).

The same graph reranking machinery can be redirected from papers to authors by replacing paper score extraction with author score extraction,

$\text{score}_{exact}(k_i, g)=s_i^\text{LLM},$2

The authors further note that AUTHORED edge weights can be adjusted by author order to emphasize first or last authors (Qiao et al., 20 May 2026).

For academic trajectory exploration, SciAtlas aggregates an author’s papers and supports clustering and LLM summarization into a trajectory report. The example in the paper describes a researcher moving from knowledge-enhanced NLP, to reasoning/planning/agentic AI, to model analysis, control, and alignment (Qiao et al., 20 May 2026).

Operationally, the system is currently accessible primarily through Neo4j, and the authors state that they have released interfaces for KG retrieval and various downstream tasks in the GitHub repository https://github.com/zjunlp/SciAtlas. They also plan CLI tools, packaged retrieval capabilities, and agentic skills for one-click use in automated scientific discovery settings (Qiao et al., 20 May 2026).

5. Empirical status, runtime, and acknowledged limitations

SciAtlas provides detailed dataset statistics, an explicit retrieval formulation, and qualitative application examples, but its empirical validation is limited. The paper claims that “The entire retrieval process can be completed within 2 minutes, significantly shorter than LLM-based deep research frameworks, while still delivering high-relevance results with in-depth topological reasoning.” This is the principal quantitative efficiency statement in the paper (Qiao et al., 20 May 2026).

The paper is explicit about what it does not provide. It does not include formal benchmark datasets for retrieval evaluation, baseline comparisons with BM25 or dense retrievers, retrieval metrics such as Recall@$\text{score}_{exact}(k_i, g)=s_i^\text{LLM},$3, nDCG, MRR, or MAP, quantitative hallucination measurements, ablation studies over the three recall paths, or task-specific downstream benchmark results. The downstream tasks are described by the authors as running examples and remain at the qualitative analysis level (Qiao et al., 20 May 2026).

Several additional limitations are stated. Access is mainly through Neo4j, and secondary development requires graph queries. The current graph is largely paper-centric, centered on papers, keywords, authors, institutions, and topics. The authors state that future versions should incorporate atomic knowledge, theorems and standards, experimental experiences, datasets, and code. Updates currently rely largely on periodic manual script execution, though the construction pipeline supports both OpenAlex daily API endpoints and periodic OpenAlex changefiles every two months, and recommends GROBID for papers absent from OpenAlex (Qiao et al., 20 May 2026).

These omissions do not negate the architectural claim, but they delimit its present status. SciAtlas is therefore best characterized as a large-scale, explicitly formulated scholarly knowledge-graph system with a qualitative application portfolio rather than as a fully benchmarked retrieval platform (Qiao et al., 20 May 2026).

6. Atlas lineage and significance

SciAtlas belongs to a broader class of systems in which an atlas is no longer a static reference object but an active computational substrate. In neuroanatomy, the Active Atlas combined a 3D anatomical reference model with learned texture detectors, automatically aligned itself to new specimens, transferred annotations, and updated its anatomical model and classifiers iteratively (Chen et al., 2017). In diffusion MRI, atlas-powered deep learning used an atlas mean, atlas variability map, and alignment error map as explicit inputs to a predictive model, making the atlas a voxelwise structured prior rather than a post hoc reference (Karimi et al., 2022). In microscopy, a probabilistic atlas for cell identification framed atlas-based labeling as a collaborative, uncertainty-aware, incrementally trained system that returns marginal probabilities over cell identities rather than a single forced answer (Bubnis et al., 2019). In social science, ExAtlas transformed archives of experiments into an atlas of links, conflicts, and bridgeable gaps, using local composition to reconcile inconsistent findings and propose bridge experiments (Zhang et al., 26 May 2026).

This suggests that SciAtlas extends an already visible atlas trajectory into the scholarly-information domain: the atlas becomes a structured, updateable, multi-relational intermediate representation for discovery, explanation, and coordination, rather than a static index or repository. In SciAtlas, the objects being atlased are papers, authors, institutions, keywords, and disciplinary hierarchies; the resulting atlas is a scientific evolution network rather than an anatomical, imaging, or experimental landscape (Qiao et al., 20 May 2026).

Its stated significance lies in this infrastructural role. The graph is intended to help agents and researchers retrieve better evidence, reason over explicit links, discover associations across disciplines, and reduce costly blind search. The strongest claims are architectural and conceptual rather than benchmark-based: SciAtlas proposes that automated scientific research should be grounded in a knowledge topology, with semantic retrieval, graph structure, and explanation paths integrated into a single system (Qiao et al., 20 May 2026).