Research Shows System 2 Reasoning Harms DAO Governance in Small Language Models

A recent study investigates the role of Small Language Models (SLMs) as edge-native constitutional firewalls for Decentralized Autonomous Organizations (DAOs). The researchers created Sentinel-Bench, an empirical framework comprising 840 inferences, to perform rigorous intra-model ablation tests on Qwen-3.5-9B. The focus was on how scaling inference-time compute impacts performance within adversarial cryptoeconomic governance settings. The results uncovered a significant compute-accuracy inversion phenomenon. The autoregressive baseline (System 1 reasoning) exhibited perfect adversarial robustness and legal consistency, achieving state finality in less than 13 seconds. Conversely, System 2 reasoning led to severe instability, primarily due to a 26.7% drop in performance. This research highlights the underexplored effectiveness of formal logic methods in challenging DAO governance scenarios, utilizing an adversarial Optimism DAO dataset to assess inference-time compute effects by toggling latent reasoning across fixed model weights. DAOs are increasingly turning to SLMs for proposal vetting and to reduce semantic social engineering risks. The findings were published as arXiv:2604.16913v1.