Six new papers accepted, three to ICLR and three to NAACL! DataEnvGym (ICLR) introduces a new framework for developing agents that adaptively generate data for training student models. System 1.x (ICLR): planning with LLMs that balances quick action prediction with slower/more deliberate planning through verbalizing search traces. See It from My Perspective (ICLR) quantifies the effect of language on cultural bias in large vision-language models. Persuasion-Balanced Training (NAACL): multi-agent training method teaching models to balance accepting good persuasion while resisting misinformation/bad persuasion. AdaCAD (NAACL): an adaptive method for balancing retrieved/context knowledge with a model’s parametric knowledge. MAMM-Refine (NAACL) improves generation through multi-agent multi-model discussion, focusing on refinement.