• In Search of the Ingredients of Open-Endedness: Replicating Picbreeder with Large Vision-Language Models — Research explores whether artificial agents can demonstrate open-ended discovery by replicating Picbreeder-like evolution using large vision-language models. 🔗 https://arxiv.org/abs/2605.23908

• Understanding and Mitigating Premature Confidence for Better LLM Reasoning — A study identifies "premature confidence" in long chains of thought as a key limiting factor in LLM reasoning and proposes ways to improve quality via confidence evolution monitoring. 🔗 https://arxiv.org/abs/2605.24396

• AgentFugue: Agent Scaling for Long-Horizon Tasks through Collective Reasoning — New research examines "scaling out" agentic capabilities by having multiple peer agents collaborate on long-horizon tasks without explicit role specialization. 🔗 https://arxiv.org/abs/2605.24486

• Adaptive Human-AI Coordination via Hierarchical Action Disentanglement — Researchers propose Intrinsic Action Disentanglement (IAD) to help AI agents better adapt to the diverse behaviors and skill levels of human partners. 🔗 https://arxiv.org/abs/2605.24343

• Meta-Agent: From Task Descriptions to Verified Multi-Agent Systems — A new two-phase framework, Meta-Agent, automates the construction and execution of specialized multi-agent systems from natural language task descriptions. 🔗 https://arxiv.org/abs/2605.25233

💡 Signal: The research landscape is heavily focused on solving the "long-horizon" problem in agentic AI — better coordination, hierarchical reasoning, autonomous system construction.