Toward a digital twin of U.S. Congress

TL;DR

The paper tackles whether a digital twin of political actors can be realized from publicly accessible social-media data. It develops congressperson-specific language-models trained on a daily-updated Twitter corpus (the Nomic Congressional Database), confirms that generated tweets are largely indistinguishable from real posts using a Turing-test-like framework, and derives predictive inferences via a data kernel perspective space (DKPS) and a flip-score metric. Key findings show that virtual tweets not only resemble real ones but also enable robust vote prediction, with a median DKPS-based accuracy of $0.87$ across 13 bills versus $0.62$ for retrieved data, and that the flip-score linearly correlates with observed cross-party voting (R$^{2}>0.8$, $p<0.05$). This work demonstrates a practical digital twin for a political body, offering actionable guidance for resource allocation and policy influence, while acknowledging data-source limitations and the need for multi-modal data and careful interpretation of inferences.

Abstract

In this paper we provide evidence that a virtual model of U.S. congresspersons based on a collection of language models satisfies the definition of a digital twin. In particular, we introduce and provide high-level descriptions of a daily-updated dataset that contains every Tweet from every U.S. congressperson during their respective terms. We demonstrate that a modern language model equipped with congressperson-specific subsets of this data are capable of producing Tweets that are largely indistinguishable from actual Tweets posted by their physical counterparts. We illustrate how generated Tweets can be used to predict roll-call vote behaviors and to quantify the likelihood of congresspersons crossing party lines, thereby assisting stakeholders in allocating resources and potentially impacting real-world legislative dynamics. We conclude with a discussion of the limitations and important extensions of our analysis.

Figures (7)

• Figure 1: An illustration of a system that contains a digital twin for a set of congresspersons.
• Figure 2: High-level characteristics of the Nomic Congressional Twitter dataset from October 10th, 2024. The dataset is updated daily and available at https://atlas.nomic.ai/data/hivemind/.
• Figure 3: Detectability analysis of different generative systems for producing Tweets from U.S. Congressperson Morgan Griffith (left) and the distribution of detectability of different systems for 100 randomly sampled congresspersons (right). The inclusion of previously written Tweets via RAG decreases detectability significantly.
• Figure 4: Two-dimensional Euclidean spaces induced by MDS of the retrieved Tweets (left) and the Data Kernel Perspective Space (right) of the generated Tweets corresponding to 117-HR-1319. Each dot represents a congressperson. Color corresponds to how the congressperson voted on the bill. The geometry of the generated Tweets has more vote-related information than the geometry of the retrieved Tweets. We validate this observation in Figure \ref{['fig:inference']}.
• Figure 5: Average performance of four different methods for predicting the voting behaviors of individual congresspersons on various pieces of legislation. Averages are calculated from 10-fold cross validation. Error bars represent one standard error. The legislation is ordered by the time of the vote in the House. For the "Retrieved" and "Generated" methods we report the average accuracy of the highest performing $k$-nearest neighbor classifiers for $k \in \{1,5,9,19,49\}$. Different bills may have different optimal $k$.