Research

2026 Electoral Systems and Geographically-Targeted Oversight: Evidence from Taiwan Legislative Yuan. Electoral Studies. 99:103026. (with Li Tang)

2025 Catalysts for Progress? Mapping Policy Insights from Energy Research. Energy Research & Social Science. 121: 103955. (with Brian Boyle, Stefan Müller, Sarah King and Robin Rauner)

2024 Electoral Reform and Fragmented Polarization: New Evidence from Taiwan Legislative Roll Call. Legislative Studies Quarterly. 50 (1): 3-21.

2024 (Mis)perception of Party-voter Congruence and Satisfaction with Democracy. Political Science Research and Methods. 13 (5): 885-902. (with Royce Carroll and Li Tang)

2023 The Role of Rituals in Adversarial Parliaments: An Analysis of Expressions of Collegiality in the British House of Commons. (Invited Contribution) Historical Social Research. 48 (3): 209-234. (with David Beck and Thomas Saalfeld)

Peer-Reviewed Articles (in Chinese)

2025 官僚「再詮釋」領導人意識形態初探：以《人民日報》習近平外交思想的評論為例 (Bureaucratic ‘Reinterpretation’ of Leaders’ Ideologies: A Case Study of People’s Daily’s Commentary on Xi Jinping Thought on Diplomacy). 中國大陸研究 Mainland China Studies. (with Yi-Nung Tsai)

• Electoral Reform and Issue Attention in Legislative Oversight: From SNTV to Mixed-Member Majoritarian System in Taiwan. (with Yu-Ceng Liao and Yi-ting Wang) Under Review PDF

  Abstract
  (Conference: 2023 EPSA ⬥ 2025 TPSA)

  This research note examines how Taiwan's electoral reform—from Single Non-Transferable Vote (SNTV) in multi-member district (MMD) to a single-member district (SMD) -- dominant mixed-member majoritarian system (MMM) -- affects how closely legislators align with their party's policy attention. While existing work explains which issues legislators emphasize and where they position themselves, less is known about how electoral systems shape the cohesion of issue attention within parties. We address this gap using 1999–2019 interpellations from Taiwanese legislators annotated with 422 fixed topic keywords. We introduce a new application of Wordfish: instead of estimating ideology, we scale legislators' issue attention and compute their divergence from the party. We find significantly higher intra-party convergence under SMD than SNTV. Local socioeconomic conditions strongly influence attention variation under SNTV but are substantially attenuated after reform. The findings highlight a key institutional trade-off: SNTV incentivizes geographically differentiated agendas, whereas SMD promotes party-aligned priorities.

• Multi-Agent LLM Systems for Synthetic Survey Experiments in Ethically Constrained Settings. (with Linette Lim and Slava Jankin) PDF

  Abstract
  (Conference: 2025 PolMeth Europe ⬥ 2025 CwC-LLM Workshop EPSA ⬥ 2025 APSA)

  Experimental studies of misinformation often face ethical constraints because they expose human participants to false or harmful content. We propose an LLM-based multi-agent framework, implemented in AG2, that reproduces core features of survey experiments in a fully synthetic environment, avoiding direct exposure of human subjects to misinformation. We calibrate a population of 1,140 synthetic agents to rich Taiwanese voter survey data and examine, within this population, susceptibility to misinformation and the effectiveness of fact-checking interventions under randomized assignment. Agents with more pro-China predispositions exhibit higher acceptance of misinformation. Corrective information substantially reduces credibility ratings -- from 2.5 to 1.0 on a five-point scale in the treatment group relative to controls -- yet correlations between political attitudes and misinformation susceptibility persist after controlling for demographics. We make the multi-agent design and implementation fully transparent, releasing code and agent specifications, and argue that such synthetic survey experiments can complement, rather than replace, human-subject studies by enabling pre-testing of experimental designs and exploration of ethically sensitive scenarios.

• Game-Theoretic Multi-Agent Systems with LLMs for Crisis Negotiation and Simulation. (with Shuli Zhang)
  

  Abstract
  (Conference: 2026 CwC-LLM Workshop SPSA ⬥ 11th Behavioral Models of Politics Conference)

  In this note, we integrate game-theoretic modeling with LLM-based multi-agent simulation to study belief-dependent crisis bargaining. Using Russia-Ukraine negotiations as a test case, we model a sequential game where Ukraine, Russia, and the US make strategic choices under uncertainty about adversary trustworthiness. Drawing on foreign policy typologies (Gravelle et al., 2017, 2020), we assign each leader one of four predispositions—Cooperative Internationalism (CI), Isolationism (ISO), Militant Internationalism (MI), and Global Justice (GJ)—while manipulating trust perceptions of Putin. This 2 × 4³ design yields 128 scenarios via AutoGen. Our results partiallyconfirm predictions while revealing asymmetries across actors: all actors conformstrongly to theory under hawkish orientations (ISO/MI), but under dovish orientations (CI/GJ), only Zelensky shows substantial cooperation (negotiating and implementing 75% and 50% of the time, respectively), while Putin never accepts demands and Trump never grants sanctions relief. Our LLM-based multi-agent simulation may help bridge the gap between game-theoretic models and behavioral implementation while providing a computational sandbox for exploring belief-dependent bargaining theories.

• Cross-Lingual Stance Detection in Political Texts: Comparison and Application. (with Stefan Müller)

  Abstract
  (Conference: 2024 ESPA ⬥ 2024 COMPTEXT ⬥ 2026 PolMeth Europe)

  Measuring the stance on specific policies provides valuable insights for understanding policy-making, changes in political preferences, and party competition. In this paper, we fine-tune three multilingual transformer models — Sentence-BERT, Multilingual BERT, and XLM-RoBERTa on annotated texts of stances in more than 53,000 comments on Twitter and more than 67,000 comments to 150 political questions in German, French, and Italian. We benchmark our fine-tuned transformer models against open-source large language models (gpt-oss-120b and LLaMA 3.1) on ground truth annotations, finding that fine-tuned transformers achieve competitive performance with substantially better computational efficiency. After identifying the most suitable model, we validate our approach by applying this fine-tuned transformer to datasets from published journal articles: politicians’ support for the annual budget (Lowe and Benoit 2013), social media posts (Bestvater and Monroe 2023), and stances across policy areas (Green-Pedersen and Little 2023). Our findings demonstrate that fine-tuned multilingual transformers provide a scalable solution for large-scale stance detection in political texts. Drawing from our systematic comparison and validation, we provide methodological guidance for researchers applying stance detection to political texts. We release our fine-tuned models alongside benchmark ground truth data to enable researchers to deploy them directly or adapt them through further fine-tuning to new domains and languages, including those not covered in our German, French, and Italian training set.

• How Do 100+ LLMs Differ Politically? A Multi-Agent Approach to Measuring AI Ideology (with Ting Luo and Slava Jankin)
  

• Knocking on the Wrong Door: Partisan Canvassing and Asymmetric Belief Updating in Britain. (with Li Tang)

• The Rural-Urban Divide in Populist Rhetoric: Cross-National Evidence from European Parliaments. (with Edoardo Viganò)