Selected Publications

1. ICML 2026
   Position: Safe Models Do Not Guarantee Safe Societies: The Case for Sociopolitical Risk

   D. Guzman Piedrahita, D. Banerjee, C. Li, T. J. Zhang, K. Blin, S. Simko, P. S. Pandey, I. Strauss, R. Mihalcea, B. Schölkopf, Z. Jin

   Spotlight · Position Paper Track

   Led a position paper arguing that model-level safety does not by itself address the sociopolitical risk posed by increasingly agentic AI systems. The extended preprint, AI Poses Risks to Democratic and Social Systems, was developed with a senior author and advisor network including Yoshua Bengio, Stuart Russell, Audrey Tang, Bernhard Schölkopf, and Richard Mallah.

   PDF
2. COLM 2025
   Corrupted by Reasoning: Reasoning Language Models Become Free-Riders in Public Goods Games

   D. Guzman Piedrahita, Y. Yang, M. Sachan, G. Ramponi, B. Schölkopf, Z. Jin

   Best Oral Paper, REALM @ ACL 2025 · 95th-percentile reviews · MSc Thesis

   Investigated the behavior of reasoning-augmented language models in public goods games, finding that reasoning can paradoxically lead to free-riding behavior. Built multi-environment, game-theoretic benchmarks to systematically simulate agent interactions and the influence of punishments and rewards.

   arXivPDF
3. EACL 2026
   Democratic or Authoritarian? Probing a New Dimension of Political Biases in Large Language Models

   D. Guzman Piedrahita, I. Strauss, B. Schölkopf, R. Mihalcea, Z. Jin

   Oral presentation

   Proposed a novel methodology to assess LLM alignment with the democracy-authoritarianism spectrum. Found that LLMs generally favor democratic values and leaders, but exhibit increased favorability toward authoritarian figures when prompted in Mandarin.

   arXivPDF
4. EMNLP 2025
   Are Language Models Consequentialist or Deontological Moral Reasoners?

   K. Samway, M. Kleiman-Weiner, D. Guzman Piedrahita, R. Mihalcea, B. Schölkopf, Z. Jin

   Introduced a taxonomy of moral rationales to systematically classify reasoning traces by consequentialism and deontology. Revealed that LLM chains-of-thought tend to favor deontological principles, while post-hoc explanations shift toward consequentialist rationales.

   arXivPDF
5. Preprint
   When Ethics and Payoffs Diverge: LLM Agents in Morally Charged Social Dilemmas

   S. Backmann, D. Guzman Piedrahita, E. Tewolde, R. Mihalcea, B. Schölkopf, Z. Jin

   Introduced MoralSim (Moral Behavior in Social Dilemma Simulation) to evaluate how LLMs behave in the prisoner's dilemma and public goods game under morally charged contexts. Showed substantial variation across models in both their general tendency to act morally and the consistency of their behavior across game types.

   arXivPDF
6. Preprint
   Robustness of Misinformation Classification Systems to Adversarial Examples Through BeamAttack

   A. Fazla, L. Krauter, D. Guzman Piedrahita, A. Michail

   Introduced BeamAttack, a novel algorithm for generating adversarial examples in NLP via beam search. Used a masked language model to predict contextually plausible word replacements, improving the coherence of the generated adversarial examples.

   arXivPDF
7. BSc Thesis
   LSTM-based Time Series Forecasting for Air Quality

   D. Guzman Piedrahita

   Evaluated LSTM neural networks for air-quality time-series forecasting in Lombardy, Italy. Employed surrogate models, specifically LIME, to assess feature importance and enhance interpretability.

   University of Bergamo