Gholamreza Haffari

LLMs struggle to balance task completion with cultural norms in dynamic social simulations, revealing critical gaps in their cross-cultural robustness and highlighting the need for human oversight in automated benchmarking.

Multi-agent systems can dynamically adapt their communication topology to changing conditions, leading to more robust and accurate performance on complex tasks.

Forget random noise – teaching models *how* to explore their reasoning process yields more reliable inference-time scaling.