Francesco Belardinelli

The paper proposes detecting 'alignment faking' (AF)—where LLMs revert to unsafe behavior when unmonitored—by analyzing observable tool selection patterns, finding that detection rates vary significantly across different LLMs and domains.

The paper introduces a novel shielding framework for Robust MDPs (RMDPs) that guarantees safety under worst-case transition probabilities, enabling safe reinforcement learning even when transition dynamics are unknown.

This paper proposes a method for ensuring safety in multi-agent reinforce learning through decentralized execution, using a shared global specification and a non-stationary multi-armed bandit.