Many scientific advances come from cross-pollination, carrying an idea from one field over to a problem in another. Yet retrieving papers across fields remains underexplored, with existing benchmarks confined to a narrow range of fields. We study cross-pollination retrieval, the problem of identifying transferable ideas across research fields, and introduce CPR-Dataset, a large-scale dataset constructed by mining cross-topic citations. Its evaluation split CPR-eval has 1,300 queries in 26 fields, and its training split CPR-train has 10,573 pairs. On CPR-eval, the best of eleven existing retrievers reaches a Recall@10 of only 0.26, whereas an LLM reranker reaches 0.81. Standard contrastive fine-tuning on CPR-train raises Recall@10 to as high as 0.70. The gain carries over to domains held out from training and to analogy and reasoning-intensive retrieval. On AnaloBench, it raises nDCG@10 by up to 33% over the untrained retriever, whereas training on three existing retrieval datasets lowers it. On BRIGHT, it raises nDCG@10 by up to 70%, more than twice the gain from those datasets. CPR-Dataset is available at https://huggingface.co/CPR-Dataset, and the code at https://github.com/SeokwonJung-Jay/Cross-Pollination-Retrieval.
Figures & tables
Figure 1: Queries and their relevant papers stay in a narrow region of science in existing benchmarks, whereas those of our Cross-Pollination Retrieval benchmark span all four domains. Each panel shows the 100 pairs farthest apart on a PCA map of paper embeddings (E5-base-v2 ( Wang et al., 2024 ) ) in SciDocs ( Cohan et al., 2020 ) , LitSearch ( Ajith et al., 2024 ) , Same Problem, Different Field ( Kulikowski, 2026 ) , or ours, with background papers and ellipses colored by OpenAlex domain. Lines join each query (filled dot) to its relevant paper (open circle).
Figure 2: Example of a cross-pollination pair. The query describes the problem of Sohl-Dickstein et al. (2015) , which introduced diffusion models. A same-topic paper, generative adversarial nets ( Goodfellow et al., 2014 ) , cannot be transferred to it. A paper from nonequilibrium statistical physics ( Jarzynski, 1997 ) can: it moves a system from an initial to a final equilibrium state, as the base paper turns the data distribution into a tractable one. The query and titles are shortened.
Figure 3: Existing retrievers fail at cross-pollination retrieval, and training on CPR-train helps beyond CPR-eval. (a) Recall@10 on CPR-eval, averaged within each retriever family; the dotted line marks random ranking. Dense scientific and Dense reasoning denote retrievers trained for scientific literature search and for reasoning-intensive retrieval. (b) Relative gain of Qwen3-Embedding-4B over its untrained version after one epoch on the same number of pairs from each dataset (Recall@10 on CPR-eval, nDCG@10 on AnaloBench and BRIGHT); for existing datasets, the mean over three random samples. Colors are scaled within each row.
Figure 4: Construction pipeline of CPR-Dataset. Top: cited papers on other topics are filtered for analogy and transfer, and those rated useful for the generated query become positives (Section 3.2 ). Bottom: hard negatives are candidates pooled by three retrievers and rated not useful (Section 3.3 ).
Split
Queries
Base papers
Pos.
Neg.
CPR-eval
1,300
1,080
1
150
CPR-train
10,573
8,137
1
9
Table 1: Split sizes of CPR-Dataset. Number of queries and of the distinct base papers they come from, with positives (Pos.) and hard negatives (Neg.) per query.
Domain
Health
Life
Physical
Social
Avg.
Method
R@1
R@10
R@1
R@10
R@1
R@10
R@1
R@10
R@1
R@10
LLM reranker †
MiMo-V2.6-Pro
0.60
0.86
0.60
0.80
0.56
0.77
0.58
0.86
0.58
0.81
Sparse
BM25
0.03
0.16
0.04
0.16
0.04
0.16
0.07
0.19
0.04
0.16
Table 2: Recall@1 and Recall@10 of retrievers and an LLM reranker on CPR-eval. Columns show each OpenAlex domain of the base paper and all queries. Bold marks the best retriever per column. Random ranking gives 0.01 and 0.07. † For reference: the LLM reranker reorders the same 151 candidates in a tournament.
Figure 5: Retrievers trained on different data. Each retriever is evaluated before training and after training for one epoch on the same number of pairs (10,573), randomly sampled from each existing dataset; for the existing datasets, bars show the mean over three random samples and whiskers one standard deviation. Qwen3-0.6B and Qwen3-4B denote Qwen3-Embedding-0.6B and 4B, and Nemotron-8B denotes llama-embed-nemotron-8b. BRIGHT covers its StackExchange subsets, and AnaloBench averages stories of 1, 10, and 30 sentences.
Figure 6: Robustness of the difficulty of CPR-eval. (a) Recall@10 divided by its random expectation in the same pool: 10/151 with all candidates, and 10/101 after removing the 50 candidates that share an OpenAlex topic with the base paper. (b) Recall@10 on the original pools and on pools rebuilt with one of the three pooling retrievers left out. (c) Recall@10 of BM25 and Gemini Embedding with the original query and with queries rewritten by GLM-5.3; the dashed line marks the LLM reranker on all 151 candidates, and the dotted line marks random ranking.
Figure 7: Recall@10 of Qwen3-Embedding-4B on domains unseen in training. The retriever is evaluated on one OpenAlex domain before training, after training on CPR-train without that domain (5,896 pairs), and after training on all of CPR-train (10,573 pairs). Details and results for Qwen3-Embedding-0.6B are in Appendix L .
Appendix figures & tables19 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 8: An analogous pair mined by our pipeline. q : Solving the Latin Square Completion Problem by Memetic Graph Coloring. d : Precoloring Extension. I. Interval Graphs. Object mappings and shared relations are those extracted by the analogy judge (Appendix B.1 ), with object names shortened.
Query
Positive
Hard negative
Object mappings (query side ↦ positive)
How to define a representative conformation tensor average for turbulent viscoelastic flows that respects the non-Euclidean geometry of positive-definite tensors
A Differential Geometric Approach to the Geometric Mean of Symmetric Positive-Definite Matrices
Numerical Simulation of the Planar Extrudate Swell of Pseudoplastic and Viscoelastic Fluids
conformation tensors ↦ symmetric positive-definite matrices; geometric mean conformation tensor ↦ Riemannian geometric mean
How to understand the role of personal and social factors in the development of social entrepreneurship in Pakistan
Economic Action and Social Structure: The Problem of Embeddedness
Shaker A. Zahra: Pioneering Entrepreneurship Scholar
social entrepreneurs in Pakistan ↦ individuals undertaking economic action; social capital and networks ↦ structures of social relations
How to cost-effectively obtain large-scale sequence information for upland cotton flower development when the genome sequence is unavailable
An EST Resource for Tilapia Based on 17 Normalized Libraries
Improvement of Cotton Fiber Quality by Rabbit Keratin Gene Transformation
cotton normalized cDNA library ↦ tilapia normalized cDNA libraries; cotton EST reads ↦ tilapia EST reads
How to understand positive and negative responses to death portrayed in animated Disney films
The Development of Coping
Frozen in Time: How Disney Gender-Stereotypes Its Most Powerful Princess
film characters facing death ↦ children facing real-life stressors; coping portrayed in the films ↦ families of coping strategies
Appendix
Table 3: More analogous pairs mined by our pipeline. Each hard negative shares an OpenAlex topic with the base paper and was rated 1 (related but not usable) by the judge of Step N2. Two of the object mappings extracted by the analogy judge (Appendix B.1 ) are shown for each pair; queries, titles, and object names are shortened.
Task
Items
Agreement
Cohen’s κ
1. Transfer verification (P2)
50
0.92
0.84
2. Usefulness of positives (P3)
25
0.92
0.84
3. Usefulness of candidates (N2)
25
0.96
0.92
Appendix
Table 4: Agreement between the LLM judges and a human annotator. Usefulness scores are binarized at 2.
Domain
Field
Subfield
Topic
Health Sciences
Dentistry
Periodontics
Oral microbiology and periodontitis research
Nursing
Nutrition and Dietetics
Child Nutrition and Water Access
Life Sciences
Biochemistry, Genetics and Molecular Biology
Genetics
Genetic Associations and Epidemiology
Immunology and Microbiology
Immunology
Immunotherapy and Immune Responses
Physical Sciences
Computer Science
Computer Vision and Pattern Recognition
Advanced Neural Network Applications
Mathematics
Statistics and Probability
Advanced Causal Inference Techniques
Appendix
Table 5: Examples of the OpenAlex hierarchy. Each row is a topic assigned to papers in CPR-eval, with the subfield, field, and domain it belongs to. OpenAlex has 4 domains, 26 fields, 252 subfields, and about 4,500 topics.
Table 14
Figure 9: All query and relevant paper pairs of each benchmark. The same map as Figure 1 , with every pair instead of the 100 with the longest lines.
Domain
Health
Life
Physical
Social
Avg.
Method
R@5
MRR
R@5
MRR
R@5
MRR
R@5
MRR
R@5
MRR
LLM reranker †
MiMo-V2.6-Pro
0.80
0.68
0.76
0.67
0.73
0.63
0.81
0.67
0.77
0.66
Sparse
BM25
0.09
0.06
0.12
0.07
0.12
0.07
0.12
0.10
0.11
0.07
Appendix
Table 8: Recall@5 and MRR@10 of retrievers and an LLM reranker on CPR-eval. Columns show results by the OpenAlex domain of the base paper and the average over all queries. The best retriever in each column is in bold. MRR denotes MRR@10. Random ranking gives 0.03 and 0.02. † Shown for reference: the LLM reranker reorders the same 151 candidates in a tournament.
Method
R@1
R@5
R@10
MRR@10
LLM reranker †
MiMo-V2.6-Pro
0.58 [0.55, 0.61]
0.77 [0.74, 0.79]
0.81 [0.79, 0.84]
0.66 [0.63, 0.68]
Sparse
BM25
0.04 [0.03, 0.05]
0.11 [0.10, 0.13]
0.16 [0.14, 0.19]
0.07 [0.06, 0.09]
BGE-M3 (sparse)
0.03 [0.02, 0.04]
0.08 [0.06, 0.09]
0.11 [0.09, 0.12]
0.05 [0.04, 0.06]
Dense (general)
Appendix
Table 9: Results on CPR-eval over all 1,300 queries with 95% confidence intervals. Brackets give the 2.5th and 97.5th percentiles over 2,000 bootstrap resamples of the queries. † Shown for reference: the LLM reranker reorders the same 151 candidates in a tournament.
Model
Training data
R@1
R@5
R@10
MRR@10
Qwen3-Emb-0.6B
Before training
0.04 [0.03, 0.05]
0.08 [0.06, 0.09]
0.12 [0.10, 0.14]
0.06 [0.05, 0.07]
CPR-train
0.24 [0.21, 0.26]
0.41 [0.39, 0.44]
0.50 [0.47, 0.53]
0.31 [0.29, 0.33]
MS MARCO
0.05 [0.03, 0.06]
0.09 [0.08, 0.11]
0.12 [0.11, 0.14]
0.07 [0.05, 0.08]
ReasonIR data
0.03 [0.02, 0.04]
0.07 [0.06, 0.09]
0.11 [0.10, 0.13]
0.05 [0.04, 0.06]
SPECTER data
0.02 [0.01, 0.03]
0.06 [0.05, 0.07]
0.09 [0.08, 0.11]
0.04 [0.03, 0.05]
Qwen3-Emb-4B
Before training
0.06 [0.04, 0.07]
0.13 [0.11, 0.15]
0.17 [0.15, 0.19]
0.09 [0.07, 0.10]
Appendix
Table 10: Retrievers trained on each training set, evaluated on CPR-eval over all 1,300 queries. Each model is trained for one epoch on 10,573 pairs with hard negatives and evaluated with a maximum length of 1,024 tokens. Brackets give 95% bootstrap confidence intervals. Bold marks the best value for each model, including the model before training. For MS MARCO, ReasonIR data, and SPECTER data, the values are from one sample (seed 42); Table 15 gives the means over three samples.
Model
Training data
Bio.
Earth
Econ.
Psy.
Rob.
Stack.
Sus.
Avg.
Qwen3-Emb-0.6B
Before training
0.113
0.211
0.158
0.136
0.111
0.120
0.116
0.138
CPR-train
0.233
0.222
0.215
0.248
0.094
0.177
0.141
0.190
MS MARCO
0.103
0.234
0.185
0.172
0.123
0.118
0.109
0.149
ReasonIR data
0.129
0.130
0.148
0.114
0.050
0.147
0.083
0.114
SPECTER data
0.131
0.192
0.183
0.170
0.098
0.187
0.107
0.153
Qwen3-Emb-4B
Before training
0.125
0.246
0.136
0.181
0.109
0.148
0.111
0.151
Appendix
Table 11: nDCG@10 on the seven StackExchange subsets of BRIGHT (all 749 queries; each subset is searched over its own corpus). Bio.: biology, Earth: earth science, Econ.: economics, Psy.: psychology, Rob.: robotics, Stack.: Stack Overflow, Sus.: sustainable living. Avg. is the mean over the seven subsets. Bold marks the best value for each model, including the model before training. For MS MARCO, ReasonIR data, and SPECTER data, the values are from one sample (seed 42); Table 15 gives the means over three samples.
T2 nDCG@10
T1 accuracy
Model
Training data
S1
S10
S30
Avg.
S1
S10
S30
Avg.
Qwen3-Emb-0.6B
Before training
0.336
0.307
0.298
0.314
0.631
0.582
0.573
0.595
CPR-train
0.450
0.393
0.364
0.402
0.686
0.644
0.627
0.652
MS MARCO
0.369
0.264
0.259
0.297
0.653
0.541
0.546
0.580
ReasonIR data
0.265
0.257
0.264
0.262
0.553
0.502
0.522
0.526
SPECTER data
0.209
0.235
0.238
0.227
0.480
0.504
0.517
0.501
Appendix
Table 12: Results on AnaloBench. T2: retrieving analogous stories from a pool of stories (nDCG@10); T1: choosing the analogous story among four options (accuracy). S1, S10, and S30 are the story lengths in sentences; Avg. is their mean. Bold marks the best value for each model, including the model before training. For MS MARCO, ReasonIR data, and SPECTER data, the values are from one sample (seed 42); Table 15 gives the means over three samples.
Qwen3-Embedding-0.6B / 4B
llama-embed-nemotron-8b
Framework
ms-swift, full fine-tuning
NeMo AutoModel 0.8.0, full fine-tuning
Batch size
128 queries (8 GPUs × 4 × 4 accumulation)
128 queries (8 GPUs × 1 × 16 accumulation)
Steps
83 (1 epoch)
83 (1 epoch)
Learning rate
6×10−6 , cosine, no warmup
2×10−6 , 2 warmup steps, cosine to 1%
Optimizer
AdamW ( β = 0.9, 0.95), weight decay 0.1
AdamW ( β = 0.9, 0.999), weight decay 0.01
Gradient clipping
1.0
1.0
Appendix
Table 13: Training hyperparameters. The same settings are used for every training set. Evaluation uses the same maximum length (1,024 tokens) and query prompt, for the trained models and for the models before training.
Training set
Source
Hard neg.
CPR-train
ours
9
MS MARCO
microsoft/ms_marco (v1.1, train)
9
ReasonIR data
reasonir/reasonir-data (HQ, VL)
1
SPECTER data
allenai/scirepeval ( cite_prediction )
1
Appendix
Table 14: Training sets. Each set has 10,573 pairs. Hard neg. is the number of hard negatives per pair.
Training set
CPR-eval
AnaloBench
BRIGHT
Qwen3-Embedding-0.6B
MS MARCO
.127 ± .004
.297 ± .001
.151 ± .006
ReasonIR data
.119 ± .004
.268 ± .010
.119 ± .005
SPECTER data
.092 ± .002
.228 ± .001
.154 ± .001
Qwen3-Embedding-4B
MS MARCO
.162 ± .005
.304 ± .008
.180 ± .008
Appendix
Table 15: Retrievers trained on three random samples of each comparison training set. Mean ± sample standard deviation over seeds 42, 1, and 2, without leading zeros. CPR-eval: Recall@10 over all 1,300 queries. AnaloBench: T2 nDCG@10 averaged over S1, S10, and S30. BRIGHT: nDCG@10 averaged over the seven StackExchange subsets (all 749 queries).
Figure 24
Model
Training data
Health (250)
Life (250)
Physical (500)
Social (300)
Qwen3-Emb-0.6B
Before training
0.13 [0.09, 0.18]
0.10 [0.06, 0.14]
0.12 [0.09, 0.14]
0.13 [0.10, 0.17]
Without the domain
0.47 [0.41, 0.54]
0.42 [0.36, 0.48]
0.33 [0.28, 0.37]
0.50 [0.44, 0.56]
All domains
0.51 [0.45, 0.57]
0.48 [0.42, 0.54]
0.45 [0.40, 0.49]
0.60 [0.54, 0.66]
Qwen3-Emb-4B
Before training
0.18 [0.13, 0.22]
0.16 [0.12, 0.21]
0.16 [0.13, 0.19]
0.18 [0.14, 0.22]
Without the domain
0.71 [0.65, 0.76]
0.62 [0.56, 0.67]
0.54 [0.49, 0.58]
0.72 [0.67, 0.77]
All domains
0.68 [0.63, 0.74]
0.67 [0.61, 0.73]
0.66 [0.62, 0.70]
0.75 [0.70, 0.80]
Appendix
Table 17: Recall@10 on each held-out domain. Without the domain : trained on 5,896 pairs without the evaluated domain. All domains : trained on all 10,573 training pairs, including the evaluated domain. Brackets give 95% bootstrap confidence intervals; the number of queries is shown next to each domain.
CPR-eval (R@10)
AnaloBench (nDCG@10)
BRIGHT (nDCG@10)
Model
Training data
HN
IB
HN
IB
HN
IB
Qwen3-Emb-0.6B
Before training
0.120
0.314
0.138
CPR-train
0.501
0.323
0.402
0.330
0.190
0.176
MS MARCO
0.124
0.093
0.297
0.224
0.149
0.131
ReasonIR data
0.115
0.115
0.262
0.249
0.114
0.082
SPECTER data
0.094
0.094
0.227
0.224
0.153
0.155
Appendix
Table 18: Retrievers trained with hard negatives (HN) and with in-batch negatives only (IB). CPR-eval: Recall@10 over all 1,300 queries. AnaloBench: T2 nDCG@10 averaged over S1, S10, and S30. BRIGHT: nDCG@10 averaged over the seven StackExchange subsets (all 749 queries). The same pairs are used in both settings; IB removes their hard negatives. For MS MARCO, ReasonIR data, and SPECTER data, the values are from one sample (seed 42).
Figure 11: Retrievers trained with in-batch negatives only. The same comparison as Figure 5 , but each retriever is trained without hard negatives. Before-training values are the same as in Figure 5 .
Query (shortened)
Relevant document
Idea that links them
Ranked first by the other retrievers
Rank
BRIGHT, earth science
Is wind just mainly nitrogen particles moving in one direction?
Wikipedia article on Brownian motion
a net drift on top of random thermal motion
Wikipedia article on wind; a page on air molecules hitting a moving car
1 / 636
Why were temperature anomalies negative between 1950 and 1980, although oil consumption had been massive since 1880?
Wikipedia article on global dimming
aerosols from air pollution blocked sunlight
Pages on temperature anomalies and on climate change
2 / 220
BRIGHT, robotics
In ROS2, do I need to rebuild the package after each change to a Python launch file?
A forum answer on symlink-install
linking the installed files to the source, so that edits take effect without compiling again
A page on the ros2 launch command; a tutorial that reruns colcon after adding a launch file
1 / >1000
AnaloBench, stories of 10 sentences
Appendix
Table 19: Queries solved only by the retriever trained on CPR-train. All retrievers are Qwen3-Embedding-4B. Rank : rank of the best relevant document for the retriever trained on CPR-train / the best of the same retriever trained on MS MARCO, ReasonIR data, or SPECTER data (one sample each, seed 42) or before training. BRIGHT ranks the documents of its subset; “>1000” means that the relevant document is not in the top 1,000. AnaloBench ranks a pool of 200 stories.
Retrieved scientific literature can serve as inspiration for both human and AI scientists. Inspiration can take different forms: prior work may directly suggest how to address a problem, or surface directions at different levels of abstraction---zooming out to a more general view or zooming in to a concrete realization. We introduce RATIO (Retrieval Across Typed Ideation Operations), a large-scale benchmark in which relevance is defined by three operations which we name ideation moves: Address retrieves potential approaches for stated problems, Broaden retrieves more general formulations, and Specify retrieves concrete instantiations. RATIO is constructed from millions of full-text scientific papers across CS literature via a general recipe that extends discourse-marker distant supervision---previously used only for classification---to corpus-scale retrieval, combined with extensive LLM and human vetting. Experiments show that operation-specific fine-tuning substantially boosts retrievers and LLM reranking helps further, but much room for improvement remains. Researchers blind to condition found the fine-tuned retriever's candidates more useful as inspirations. RATIO provides a scalable training and evaluation framework for retrieval components that support literature-grounded ideation, opening up new research avenues in scientific inspiration retrieval. We release full data, code, models, and annotations.
Maayan Sharon, Tom Hope
The Hebrew University of Jerusalem · The Allen Institute for AI (AI2)
What makes great scientists great? Even as AI systems start to make progress on open problems, scientists remain far ahead of them at sensing which prior idea, buried in an ever-growing archive of research, a new problem needs. To study this skill, we draw on researchers who know firsthand which earlier work advanced their completed projects, with papers serving as pointers to the ideas within. Using our automated pipeline that makes author annotation scalable, we build ScholarCatalyst by having 184 lead authors of 207 recent computer science papers label which candidates did or could have advanced their project, each with a detailed rationale. We introduce a retrieval task with author-provided judgments: given an initial research question, retrieve these papers from only the literature available when the project began. Agentic search does no better than embedding retrieval (0.42 vs. 0.48 Recall@20) despite calling that same retriever as a tool. Even an agent built on Claude Fable 5.1, which may have seen the completed papers during training, reaches only 0.51 R@20. These results highlight the need for new training recipes that equip models with expert intuition for searching broad corpora. We envision ScholarCatalyst as a step toward scientific agents that can take a half-formed idea and point to the prior research it needs.
Sohyeon Kim, Yoonho Lee, Bo Liu +11
Stanford University · Seoul National University · University of Washington +3
The discovery of novel methodologies for emerging problems is a continuing cycle in ML, often driven by the migration of techniques across domains. Building on this observation, we ask whether current LLM ideation systems benefit from targeted cross-domain retrieval or simply from exposure to diverse mechanisms. We study this question through PaperGym, a three-stage pipeline: (1) tool-augmented seed extraction via read, grep, and bash over an isolated paper environment, (2) cross-domain seed retrieval via paraphrasing across seven ML domains, and (3) method synthesis from retrieved seeds, each scored by rubric-based judges. Tool-augmented extraction improves specificity, and paraphrase-based retrieval broadens domain coverage. In synthesis, cross-domain retrieval receives more pairwise novelty wins than no-retrieval and same-domain baselines, but shows no significant difference from a random diverse-seed control. These findings suggest LLM ideation systems benefit from diverse seed exposure, but do not yet reliably exploit the semantic reason particular seeds were retrieved. We release the seed library, rubric prompts, and run scripts at https://github.com/yunjoochoi/PaperGym
Yunju Choi, Min Song
1Yonsei University, Seoul, Republic of Korea · 2OnomaAI, Seoul, Republic of Korea.