Dylan Van Mulders, Matthias Bogaert, Dirk Van den Poel
Forming a governing coalition is messy negotiation, driven by both concrete policy goals and stubborn ideological commitments. That makes it tempting to simulate with LLM agents, except the politeness and neutrality drilled into these models by RLHF makes them bad at holding a firm partisan line. The authors combine fine-tuning, preference optimization, and retrieval so each agent adopts an aggressive party persona while staying anchored to its actual manifesto.
They run this on the 2019 Flemish election, with party agents negotiating under a formateur, and add tooling to trace every clause in the final deal back to its manifesto source and score which party shaped the outcome. Across three runs they get a stable winner and ranking, and manifesto-grounded clauses predicted what really happened while hallucinated ones did not.
This describes the abstract, so read the paper for how the tracing works and how far to trust the simulation.
The formation of political coalitions is a complex negotiation driven by both concrete policy objectives and deep-seated ideological convictions. While Large Language Models (LLMs) open new avenues for computational political science, the neutrality and helpfulness biases instilled by Reinforcement Learning from Human Feedback (RLHF) prevent them from sustaining steadfast partisan behaviour. We present a multi-agent framework that reconciles factual grounding with…
Revisiting data-driven dynamic security assessment with a tabular foundation model
arXiv (cs.AI) · July 17, 2026Rethinking Quantum Continual Learning with Quantum Fisher Information
arXiv (cs.LG) · July 17, 2026CLaC@FinMMEval 2026 Task 3: Sentiment-Augmented Deep Reinforcement Learning for Active Trading -- An Alpha-Reward Approach
arXiv (cs.LG) · July 17, 2026Constrained Hebbian Learning Supports Efficient Representational Allocation under Structural Constraints