When researchers leave a field, they may take more than their labor with them. A concept that only a handful of people actively work on can stop being connected to the rest of the field once those people move to a hotter topic or to an industry lab. This proposal tests that claim directly. It tracks individual authors as they migrate between research areas, and asks two questions. First, does a field lose cross-topic connections when the people who carried a concept depart? Second, does a concept's chance of entering a new field depend on whether its carriers moved there?
The idea borrows its mechanics from the economics of brain drain, where skilled people leave a periphery for a high-reward center. That literature predicts tipping points and self-reinforcing traps rather than smooth decline. The proposal applies those mechanics to the problem of knowledge silos, where science splits into poorly connected subfields. Existing bibliometric work on fragmentation tracks concepts and leaves people out. Existing work on migration tracks people and leaves ideas out. Joining the two would tell research policy whether funding training or bridging programs heals silos, or only speeds the flight of talent out of donor fields.
Core Claim
We define a concept's carriers in a field as the authors who account for most of its uses there over a trailing window. The central claim is that the departure of carriers causes measurable "edge decay" in the donor field: fewer new co-occurrences between the concept and other concepts, and loss of the concept's bridging position. The departure of carriers also predicts the concept's arrival in the destination field, beyond what the concept's own network position already predicts.
A second, sharper claim comes from the brain-drain model. Survival is not linear in the share of carriers lost. Below some critical fraction, the remaining community sustains the concept. Above it, the concept collapses toward a low-activity state. This mirrors the two stable steady states and the initial-endowment threshold proved in FIXED POINTS IN MULTIPLE EQUILIBRIA PROBABILITY-MIGRATION MODELS, a two-period model in which migration expectations split economies into a low-education trap and a high-education state. Here, "human capital" becomes the active carrier base of a concept within a field.
Where This Sits
Three papers motivate the gap.
Topics, Authors, and Institutions in Large Language Model Research documents the inflow side. In 2023, 49.5% of large-language-model first authors had never published in natural language processing; they came from vision, security, and robotics. That study does not ask what happened to the fields those authors left. That question is our first estimand.
Will this Idea Spread Beyond Academia? tracks 450,000 concepts across papers, patents, and clinical trials. It shows that a concept's co-occurrence neighborhood predicts transfer (AUC about 0.80 for patents). In its conclusion it names researcher mobility as the missing mechanism. We supply that mechanism, and we test whether it adds information beyond the neighborhood.
Can We Stop the Academic AI Brain Drain? argues qualitatively that corporate hiring hollows out academic fields and creates proprietary silos. It offers no graph measurements. We provide them.
The "talent vacuum" finding in Spillover Effects of Foreign Bank Entry in China's Banking Sector plays a constraining role. In that study, entrants absorbed most internationally experienced staff (domestic banks kept only 5.08% of staff with foreign work experience) instead of producing the expected spillovers. Movement of people therefore need not mean movement of knowledge, and our design must allow carriers to move without their concepts following.
Two papers supply methods. Exploring and Verbalizing Academic Ideas by Concept Co-occurrence builds evolving concept graphs and predicts new edges with a pretrained language model. We reuse its graph construction and use it as a baseline. Cao et al. supply concept extraction and the transfer-prediction setup.
Research Program
We build an author-resolved concept graph over arXiv computer science and statistics papers from 2015 to 2024. This extends the 418,000-paper population that Movva et al. filtered. Concepts are extracted with the phrase-mining approach of Cao et al. Fields are arXiv primary categories, and the topic clusters of Movva et al. serve as a robustness check.
For each concept, field, and year we compute three quantities. The first is the carrier set. The second is the share of carriers whose subsequent primary output shifts to another field or to industry-only affiliation. The third is the edge outcome: new co-occurrence edges formed in the donor field over the following three years, and the concept's betweenness position between topic clusters.
Causal design. Departures are endogenous, because people abandon topics that are already dying. The design therefore uses a shift-share pull shock. The rise of large-language-model work in 2022–2023 pulled authors in proportion to their pre-period closeness to language modeling, as measured by their own earlier papers. Closeness predicts departure but should not, by itself, affect a donor concept's trajectory. This exclusion assumption is checked with event-study pre-trends.
The event study compares two groups of concepts that are matched on pre-period frequency, neighborhood, and growth. In one group, carriers were heavily exposed to the pull. In the other, they were not.
Threshold test. We fit concept survival as a function of carrier-loss fraction. A flexible model with a breakpoint is compared against a linear-decay model. We also test whether post-shock activity across concepts is bimodal, which a two-equilibrium mechanism predicts and a smooth one does not.
Destination side. For the destination field, we ask whether a concept's first appearance in a new category is predicted by its carriers having moved there. The comparison is against concepts that arrived without carrier movement, which we read as diffusion through reading rather than through people.
Baselines
Naive reference. A persistence model predicts a concept's future edges and arrivals from its own past activity alone. Every other model must beat it on held-out years.
Strongest existing predictor of transfer. This is the structural-neighborhood model of Will this Idea Spread Beyond Academia?, which reports an AUC of about 0.80 for academic-to-patent transfer. That number comes from a different task, so we re-implement the model on our cross-category arrival task. It gets the same concepts, the same five-year observation window that paper found optimal, and the same hyperparameter search budget as our model. Beating it means a statistically reliable AUC gain from adding carrier-mobility features, with bootstrap intervals over concepts, on years not used for tuning.
Strongest edge predictor. This is the pretrained-language-model link predictor of Exploring and Verbalizing Academic Ideas by Concept Co-occurrence, which reports F1 0.560 on new 2021 edges. We retrain it on our graphs and add carrier-departure features as extra input under an identical compute budget. Beating it means higher F1 on new donor-field edges.
Temporal graph networks. GCN-GAN and EvolveGCN are retained for completeness. No numeric result for them was found on Lacuna; Xu et al. report only that they "struggled to identify any entirely new connections."
Shuffled-carrier ablation. This isolates the proposal's key ingredient. Carrier identities are permuted within field-year while the number of departures per concept is preserved. If the gain survives the shuffle, it comes from field-level attrition, not from who carried what.
Why not just…?
Why not just use the concept's network neighborhood, as Cao et al. do? This is the most serious alternative. Carriers' movements may simply mirror a concept's neighborhood, since people working on bridge concepts are also the ones most pulled toward adjacent fields. Two pieces of the design settle it. The predictive comparison conditions on the full neighborhood model, and the shuffled-carrier ablation checks whether identity matters. If mobility adds nothing over the neighborhood, we report that as a negative result: fragmentation is then a property of idea topology, not of people.
Why not just count how many authors left each subfield? Descriptive attrition counts, in the style of Movva et al., cannot separate "people left, so the topic decayed" from "the topic was decaying, so people left." The pull-shock instrument and the pre-trend checks exist for this reason. If pre-trends differ between exposed and unexposed concepts, the causal claim fails, and we fall back to predictive claims only.
Why not just apply a temporal graph network to co-authorship graphs? Xu et al. found that these models fail to predict genuinely new connections on sparse concept graphs. Our edge outcome is dominated by exactly such new connections. We still include them as a baseline so the failure is shown on our data rather than assumed.
Proposed Experiments and Decisive Checks
Experiment 1: donor-side edge decay. We estimate the event study around 2022–2023 for vision, robotics, speech, and security concepts. A clear result would be a drop in new-edge formation and bridging position for high-exposure concepts relative to matched controls, with flat pre-trends. A null with tight intervals would refute the hollowing claim made by Kunze.
Experiment 2: destination-side arrival. We predict first appearance of concepts in new arXiv categories from 2019 to 2024, training on earlier years and testing on later ones. The model is compared against the neighborhood and language-model baselines above.
Experiment 3: tipping. We test the breakpoint and bimodality predictions. A linear-decay fit that matches the breakpoint model would argue against importing the two-equilibrium brain-drain mechanics into science.
Transport check. As a secondary check on whether the result extends beyond computer science, we link Web of Science authors to USPTO inventors by name and affiliation in the corpus of Cao et al. We then ask whether academics who move into patenting carry their concepts across the 2.7% transfer barrier. Name linking is noisy, so this is reported separately as an exploratory check.
Policy reading. If carriers matter and tipping exists, the policy reading is direct. Programs that train or pull people across fields without retaining a critical carrier mass in donor fields can deepen fragmentation, not heal it. This is the academic analogue of rural training budgets subsidizing urban workforces.
References
- FIXED POINTS IN MULTIPLE EQUILIBRIA PROBABILITY-MIGRATION MODELS
- Topics, Authors, and Institutions in Large Language Model Research: Trends from 17K arXiv Papers
- Will this Idea Spread Beyond Academia? Understanding Knowledge Transfer of Scientific Concepts across Text Corpora
- Can We Stop the Academic AI Brain Drain?
- Spillover Effects of Foreign Bank Entry in China’s Banking Sector
- Exploring and Verbalizing Academic Ideas by Concept Co-occurrence