SCoR: A Hierarchical Framework for Forecasting Relations Between Scientific Concepts

Jingze Wang; Fred Sun; Shangqi Guo

cs.CLAug 27, 2026

SCoR: A Hierarchical Framework for Forecasting Relations Between Scientific Concepts

Authors: Jingze Wang, Fred Sun, Shangqi Guo

Organizations: Center for Brain-Inspired Computing Research, Tsinghua University

Abstract

Anticipating emerging research directions is a critical goal of AI-assisted science. Existing methods mainly predict which concepts will co-occur in future papers, but co-occurrence captures shared attention rather than the scientific meaning of a connection, such as whether one method uses, combines, replaces, or contradicts another. We formulate research-direction discovery as hierarchical scientific-relation forecasting over a shared candidate-pair space, comprising three temporally aligned tasks: first co-occurrence, first scientific-relation formation, and relation type at formation. We construct SCoR-Graph from 187,848 cs.CV papers published between 2017 and 2026, yielding 270,687 consolidated concepts, 7.45 million co-occurrence edges, and 615,036 typed, directed relation edges. From cutoff-specific graph snapshots, we derive SCoR-Bench, a leakage-audited benchmark for these three capabilities, with expert-verified gold labels for the entire relation-type test set. We further introduce HiSCoR, a task-adapted model family that models relation emergence as a temporally evolving, hierarchically constrained process by encoding pre-cutoff event histories and conditioning relation formation on future co-occurrence. On the held-out 2025-2026 window, HiSCoR achieves an AUROC of 0.9515, a 2.4% relative improvement over the strongest temporal-graph baseline, and improves population-AUPRC by 14.0%; its relation-type variant achieves a Macro-AUROC of 0.7795. Ablations show that semantic, co-occurrence, and typed-relation views provide complementary predictive evidence. SCoR advances research-direction forecasting from predicting which concepts will co-occur to anticipating whether and how evidence-backed scientific relations will emerge.

Figures & tables

Figure 1: Nested event spaces for scientific-relation forecasting. (a) Within the candidate space $\Omega$ , pairs that co-occur in the target window $\mathcal{H}_{t}$ form $C^{\star}$ ; relation-positive pairs form the nested subset $\mathcal{R}$ , on which alone relation types are defined. $T_{0}$ – $T_{2}$ evaluate first co-occurrence, first relation formation, and relation type. (b) Both example pairs co-occur, yet only one forms an evidence-backed relation: optical flow is explicitly combined with a VLM in their joint paper ( combines ), whereas CNN and NeRF merely appear in the same paper, so no relation type is defined.

Figure 2: End-to-end overview of SCoR. Literature is converted into SCoR-Graph , whose semantic, co-occurrence, and typed-relation views support the SCoR-Bench tasks; HiSCoR couples target-window co-occurrence with relation formation through a hierarchy-constrained dual tower, with relation type evaluated conditional on formation.

Figure 3: HiSCoR for relation formation. A static S/C/R backbone estimates $P(C^{\star})$ and $P_{\mathrm{static}}(R\mid C^{\star})$ ; the event pathway adds a gated residual with training-only ranking supervision. Their product preserves $P(R)\leq P(C^{\star})$ by construction.

Figure 4: Held-out test ROC curves. T0/T1 use population weights; T2 averages one-vs-rest curves. Insets show low FPR; circles mark operating points; bold/dashed curves denote final HiSCoR/EF. Table 1 reports three-seed means and PR metrics.

T0 First co-occurrence
Model	AUROC	Pop. AUPRC	Brier $\downarrow$	AP-Lift
Marwitz mixture	0.9296	$2.97{\times}10^{-3}$	$6.02{\times}10^{-2}$	69.4 $\times$
DyGFormer	0.9273	$3.62{\times}10^{-3}$	$6.92{\times}10^{-2}$	84.5 $\times$
GraphMixer	0.9348	$2.09{\times}10^{-3}$	$6.07{\times}10^{-2}$	48.8 $\times$
HiSCoR-Cooc
w/o Temp.	$\underline{0.9420}$	$\underline{3.90{\times}10^{-3}}$	$\underline{7.00{\times}10^{-5}}$	$\underline{91.1{\times}}$

Table 1: Held-out test performance. Three-seed means; best/runner-up are bold/underlined, and blue rows are HiSCoR.

Model	$K=100$	$K=1$ K	$K=10$ K
Marwitz concat	7.00 / 0.015	6.00 / 0.126	2.97 / 0.626
GraphMixer	5.67 / 0.012	5.23 / 0.110	4.09 / 0.863
HiSCoR-EF	0.00 / 0.000	0.73 / 0.015	0.76 / 0.160
HiSCoR	12.00 / 0.025	9.00 / 0.190	4.85 / 1.022

Table 2: Future-blind T1 candidate reranking. Precision@K / end-to-end Recall@K (%).

Figure 5: Matched HiSCoR-family diagnostics. (a) Task-specific static/temporal test comparisons (T1: HiSCoR-EF); (b) validation S/C/R ablations; and (c) HiSCoR-State test factorization controls. Bars show three-seed means and sample SDs; these are precursor, not final-model, ablations.

Figure 6: Matched diagnostic evidence. (a) HiSCoR-EF validation AP-Lift for Temporal/ S/C/R (Brier:

1.46\times 10^{-5}\!\to 3.94\times 10^{-6}

). (b) Frozen-Test paired

\Delta

AUROC and hierarchy validity around HiSCoR-State A0; X3 violates the hierarchy in 53.06%. All are precursor variants.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 7: Original HiSCoR-State architecture. Shared semantic (

\mathcal{S}

), co-occurrence (

\mathcal{C}

), and typed-relation (

\mathcal{R}

) field encoders produce a normalized pair state for two task-specific MLP towers. The auxiliary tower estimates target-window co-occurrence

P(C^{\star})

, while the conditional tower estimates

P(R\mid C^{\star})

. Their product gives

P(R)=P(C^{\star})P(R\mid C^{\star})

, structurally guaranteeing

P(R)\leq P(C^{\star})

. This static hierarchy-consistent model serves as the backbone for the temporal-event extensions described in the main text.

Statistic	Value	Statistic	Value
Papers	187,848	Coverage	2017-01-03 to 2026-06-25
Surface concepts	290,197	Active surface concepts	290,001
Merged surface concepts	19,510	Final graph nodes	270,687
Blocked ambiguous acronym families	975	Unique co-occurrence edges	7,452,716
Co-occurrence paper events	9,199,848	Unique typed relation edges	615,036
Underlying relation pairs	580,615	Relation-evidence events	1,306,721

Appendix

Table S1: Final SCoR-Graph construction and identity-model record. U/C/R/X denotes uses / combines / replaces / contradicts . Identity rows report separately held-out test performance; the parentheses in the final two rows give selected-threshold precision/recall.

Split	Cutoff	Horizon end	Master rows
Train-1	2022-06-25	2023-06-25	42,121,503
Train-2	2023-06-25	2024-06-25	51,506,057
Validation	2024-06-25	2025-06-25	63,476,038
Test	2025-06-25	2026-06-25	75,632,357

Appendix

Table S2: Final SCoR-Bench chronological splits and frozen test views. Master rows are the final benchmark rows before task-specific evaluation views. U/C/R denotes uses / combines /rare ( replaces or contradicts ) T2 positives.

Task	Model	AUROC / M-AUROC	AUPRC / M-AP	Brier / M-F1	AP-Lift / rare-F1
T0	HiSCoR-Cooc w/o Temp.	$0.9420{\pm}0.0012$	$0.003901{\pm}0.000898$	$(7.0034{\pm}0.7520){\times}10^{-5}$	$91.1{\pm}21.0$
T0	HiSCoR-Cooc	$0.9495{\pm}0.0005$	$0.010275{\pm}0.001495$	$(4.2937{\pm}0.0106){\times}10^{-5}$	240.0
T1	HiSCoR-EF w/o Temp.	$0.9409{\pm}0.0020$	$0.000249{\pm}0.000152$	$(1.3296{\pm}0.0387){\times}10^{-5}$	$130.7{\pm}79.5$
T1	HiSCoR-EF	$0.9436{\pm}0.0022$	$0.001168{\pm}0.000326$	$(3.2213{\pm}0.0748){\times}10^{-6}$	$612.8{\pm}170.9$
T1	HiSCoR	$0.9515{\pm}0.0001$	$0.003859{\pm}0.000673$	$(4.8725{\pm}0.3562){\times}10^{-6}$	$2025.2{\pm}353.1$
T2	HiSCoR-Type w/o Temp.	$0.7754{\pm}0.0011$	$0.5425{\pm}0.0032$	$0.5034{\pm}0.0083$	$0.0927{\pm}0.0017$

Appendix

Table S3: Final HiSCoR-family frozen-test summary (mean

\pm

sample SD over seeds 17, 23, and 42). For T0/T1, metric columns are AUROC, population AUPRC, Brier, and AP-Lift. For T2, they are Macro-AUROC, Macro-AP, Macro-F1, and rare-class F1, respectively. Checkpoints were selected only on validation data.

Scientific Discovery

Scientific discovery uses computational methods to find hypotheses, laws, mechanisms, and materials from data, literature, and experiments.

See all 85 papers

Sep 16, 2026

Time-Aligned Evolving Concept Graphs for Scientific Relation Forecasting

Forecasting scientific relations can guide discovery by identifying promising connections before they emerge. Existing approaches often model concept semantics and graph structure separately or summarize semantics over coarse historical snapshots, leaving semantic representations potentially misaligned with rapidly evolving graph evidence. We propose a time-aligned evolving concept graph framework that jointly models semantic and structural evolution. Its core idea is to treat dated papers as shared update events, reconstructing semantic and structural states from the same publication history through each prediction time. Pair-level fusion combines these states to forecast first co-occurrence, relation formation, and conditional relation type. Holding architecture and training fixed, refreshing context alongside graph updates improves mean relation AUPRC by 16.6% over frozen context. On a graph built from 187,848 papers with 270,687 concepts and 7.45 million co-occurrence links, the complete framework improves mean relation AUROC from 0.9290 for the strongest evaluated baseline to 0.9722, with mean population-weighted AUPRC 0.005778.

Fred Sun, Jingze Wang, Minkun Xu +1

Center for Brain-Inspired Computing Research, Department of Precision Instrument, Tsinghua University, Beijing, China · Guangdong Institute of Intelligence Science and Technology, Zhuhai, China

AI-Assisted Scientific Research Temporal Link Prediction Dynamic Graph Learning

Ask PDF

Feb 24, 2026

PreScience: A Dataset and Benchmark for Scientific Forecasting

Can AI systems trained on the existing scientific record forecast the advances that will follow? We introduce PreScience, a dataset and benchmark for scientific forecasting built around 98K recent AI research papers, together with companion papers covering author publication histories and citation links, yielding 502K papers in total. The resulting paper records include titles, abstracts, disambiguated author identities, influential references, topic labels, citation trajectories, and metadata snapshotted to respect temporal cutoffs. We instantiate seven exemplar tasks: five paper-anchored tasks -- contribution generation, collaborator prediction, prior work selection, citation count prediction, and future combination prediction -- and two aggregate topic trend forecasting variants. We develop baselines ranging from simple heuristics and embedding methods to frontier language models and agentic systems, and introduce LACER, an LLM-based metric for evaluating similarity of generated contribution descriptions that agrees better with human judgments than existing metrics. Finally, we compose task models to generate a 12-month synthetic corpus and find that the resulting papers are systematically less diverse and less novel than human-authored research from the same period. We release the PreScience dataset (https://huggingface.co/datasets/allenai/prescience) and code (https://github.com/allenai/prescience).

Anirudh Ajith, Amanpreet Singh, Jay DeYoung +7

Allen Institute for Artificial Intelligence, Seattle, WA, USA · Knowledge Lab, University of Chicago, Chicago, IL, USA · School of Computer Science and Engineering, Hebrew University of Jerusalem, Israel +1

AI for Science Forecasting Benchmarks

Ask PDF

Oct 18, 2026

Bridging Concepts: Multi-Agent Reasoning over Scientific Concept Graphs for High-Impact Research Prediction

Within a research community, new topics often emerge when previously unconnected concepts are first studied together. We construct scientific concept graphs (SCGs)--domain knowledge graphs whose nodes are fine-grained research concepts extracted from paper text by LLMs and whose edges capture their co-occurrence in papers--and study the task of forecasting high-impact concept pairs, i.e., discovering concept pairs that are currently unconnected but will have high impact when connected. To this end, we introduce new benchmark datasets constructed from CIKM, KDD, and SIGIR conference papers, with varied prediction settings, including impact forecasting for emerging concepts unseen during training. We propose MiroGraph, a Multi-agent reasoning framework with scientific concept Graphs, in which a feature designer agent designs predictive features for a given concept pair from structural, semantic, and trend perspectives, and then three LLM agents compute the actual feature values from the enclosing subgraph of the concept pair. Using these feature values, a classifier designer agent constructs a classifier that predicts the final impact labels. MiroGraph outperforms various impact forecasting and temporal link prediction methods.

Donggyu Yoon, Minsung Hwang, Juneha Baek +2

School of Computing, KAIST · College of Business, KAIST · Department of AI Computing, KAIST

Knowledge Graphs LLM Reasoning with Graphs Temporal Link Prediction Multi-Agent Reasoning

Ask PDF

Abstract

Scientific Discovery

Explore similar work

Time-Aligned Evolving Concept Graphs for Scientific Relation Forecasting

PreScience: A Dataset and Benchmark for Scientific Forecasting

Bridging Concepts: Multi-Agent Reasoning over Scientific Concept Graphs for High-Impact Research Prediction