Gabriele
La Malfa

ProblemWhen users query Large Language Models (LLMs), are the outputs truly safe? Content generation carries inherent risks, prompting researchers to reinforce these models using self-play loops, where a single model acts as both attacker and defender, iteratively hardening its own defenses. However, this approach raises a critical concern: How do we prevent the model from collapsing into localized self-consistency when it plays both sides?

ApproachTo answer this, we introduce Anchored Bipolicy Self-Play (ABS), a novel framework that disrupts the self-consistency assumption in LLM safety alignment. By training a pair of adversarial policies that specifically exploit the model's internal guardrails, we demonstrate that current alignment techniques remain fundamentally vulnerable to coordinated self-play attacks.

ProblemOrchestrating agents that collectively behave fairly while efficiently pursuing their goals is a critical frontier in fair reinforcement learning (Fair RL). However, a persistent challenge remains: How can we enforce and enhance fairness when certain agents possess sensitive attributes?

ApproachTo address this, we introduce Fair-PPO, an adaptive, PPO-grounded optimization method designed to dynamically re-orient policy convergence whenever a target fairness metric is violated. Unlike static approaches, Fair-PPO actively accounts for and adjusts gradient convergence by balancing two critical dimensions: historical bias derived from past trajectories, and potential future bias estimated via the value function.

We argue that large language models fundamentally lack the coordination capabilities required for effective multi-agent systems. Through systematic evaluation, we demonstrate critical gaps in strategic reasoning, theory of mind, and emergent communication when LLMs are deployed as autonomous agents.