Search arXiv⌕ Search
Home/Archive/arXiv

arXiv · 2509.05591

Language Model Perplexity Predicts Scientific Surprise and Transformative Impact

Zhen Zhang·James Evans
View original source ↗Read source document ↗Export metadata ↓

Abstract

Scientific breakthroughs typically emerge through the surprising violation of established research ideas, yet quantifying surprise has remained elusive because it requires a coherent model of all contemporary scientific worldviews. Deep neural networks like large language models (LLMs) are arbitrary function approximators tuned to consistently expect the expressions and ideas on which they were trained and those semantically nearby. This suggests that as LLMs improve at generating plausible text, so the perplexity or improbability a text sequence would be generated by them should come to better predict scientific surprise and disruptive importance. Analyzing over 2 million papers across multiple disciplines published immediately following the training of 5 prominent open LLMs, here we show that higher perplexity scores systematically predict papers that receive more variable review ratings, longer editorial delays, and greater reviewer uncertainty. The most perplexing papers exhibit bimodal outcomes: disproportionately represented among the most celebrated scientific achievements and also the most discounted. High-perplexity papers tend to be published in journals with more variable impact factors and receive fewer short-term citations but in prestigious venues that bet on long-term impact. They also generate more interdisciplinary engagement portending long-term influence, and are more likely to have been supported by speculative funders like DARPA versus the NIH. Interestingly, we find the opposite pattern for humanities research, where the least surprising work is the most celebrated and cited. Our findings reveal that computational measures of corpus-wide linguistic surprise can forecast the reception and ultimate influence of scientific ideas, offering a scalable approach to recognize and generate potentially transformative research that challenge conventional scientific thinking.

Explore related subjects

cs.SI

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Zhen Zhang, James Evans. 2025-09-06. Language Model Perplexity Predicts Scientific Surprise and Transformative Impact. https://arxiv.org/abs/2509.05591

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Two Fault Lines: Latent Polarity Geometry in X Community Notes

Community Notes is X's crowdsourced fact-checking system. A note is published beneath the post it corrects only when raters who usually disagree both rate it helpful, a design called bridging. To apply that rule, the system learns who disagrees with whom from the ratings alone, placing every rater and note on one line, the polarity axis. Every scorer in the production pipeline uses a single axis. Refitting the base model these scorers share on the full public data (212.9M ratings, 2.33M notes, 1.07M raters), we find that one axis is too few. The space is at least two-dimensional. The first axis is left/right politics, while the second, which we interpret as trust in institutions, is largely independent of the first. A held-out test confirms that the second axis improves prediction of unseen ratings, while a third adds little. A second rater dimension learned from one set of topics predicts how raters judge COVID and Ukraine notes excluded from the fit, so it does not merely restate subject matter. Among heavily rated notes that barely divide raters politically, the published share falls from 71.5% to 11.7% as second-axis disagreement grows. A one-axis fit records these notes only as weakly polarised and less helpful; the information that raters at one end of the second axis support them is lost. Authors write notes matching their own position on both axes (r = 0.538 and 0.358), and a small minority of raters cast most ratings (Gini = 0.718). Fewer notes are published in the smallest language communities, but the shortfall is in ratings received, not in how the rule treats them. Keeping ratings per note constant, only Hindi stays below the global rate of 10.85%, and Greek moves from 7.76% to 11.68%. We argue for a bridging model with more than one axis of disagreement, and for recruiting raters in the languages the current design reaches least.

cs.SI↗
arXiv2026

Systematic comparison of gender inequality in scientific rankings across disciplines

Participation of Women in academia has grown over recent decades. Yet, it is unclear how this growth translates into representation at the top of academic rankings (measured by scientific productivity and citations). Here, we investigate gender gaps in productivity, citations, and coauthorship networks across 18 fields, using 67.7 million papers published between 1975 and 2020 in the Semantic Scholar Open Research Corpus, with the authors binary gender inferred from names using Genderize and Namsor. We find that women remain consistently underrepresented in top-ranked positions across all fields, even in disciplines where their overall participation is relatively high. We observe that rankings are generally becoming increasingly rigid over time, with fewer researchers entering or leaving top-ranked positions from one year to the next in most fields, although we do not test whether this rigidity contributes to the disparities we document. Across fields, highly productive men receive more citations than the closest available women based on the observed productivity and career stages. However, when top-ranked women are compared with their closest male counterparts in terms of career stage and publication profile (accounting for venue prestige and authorship position), gaps narrow in some fields and in a few others disappear or reverse, showing that, among highly productive researchers, differences in research output alone do not fully account for the citation gaps.

cs.SI↗
arXiv2026

Finding Common Ground: Graded Communal Knowledge in Bluesky Starter Packs

Communication is made possible by common ground---the unspoken knowledge that people share and presuppose of one another, whether that be online or offline. In his conception of common ground, Clark (1996) distinguishes between personal and communal common ground, and asserts that the latter is graded: the more community affiliations two people share, the more common ground they share as well. Social media research has invoked this mechanism to explain how users connect, but it has gone largely untested because community memberships are rarely visible and, where they are, they are coupled to user interactions in a way that leads to conflating effects. To circumvent these challenges, this study repurposes Bluesky starter packs (SPs) as user-curated community affiliation labels. Across 191,648 pairs of users, we show that shared lexical repertoire---our proxy for common ground---grows monotonically with the number of SPs that users share, with users sharing a single pack being roughly twice as similar as equally connected strangers. A semantic renormalization of SP co-membership shows furthermore that it is more so the number of topically \emph{distinct} communities, rather than the raw count, in which common ground is graded. Finally, we show that community co-membership adds to common ground independently of proximity in the Bluesky follow network. These results lead to the conclusion that community membership is a measurable, separable, and semantically structured carrier of common ground. Reading it as such makes common ground observable before an exchange rather than inferred from it, and thus opens the door for large-scale observational approaches to a set of questions that have so far only been posed in the laboratory.

cs.SI↗