Significance: Directly addresses the "memory coherence problem" in long-running agents. Specifically mentions integration with the OpenClaw runtime. Achieves 33% improvement on LongMemEval-S by introducing structured episodic/semantic tiers and a PPO-based retrieval policy.
Significance: Demonstrates a framework to distill reasoning capabilities from DeepSeek-R1 into compact open-source models (Phi3, Qwen-Coder). Implements "response stabilization" to ensure reliable logic outputs in student models.
Significance: Introduces UniReasoner, a framework where the LLM performs self-critique on its own visual drafts to guide iterative generation. Closes the "understanding-generation gap" via explicit corrective signals.
DeepMind demonstrated a legitimate self-improvement loop in an open-ended 3D world. Gemini-powered agents were dropped into unseen survival environments and evolved capabilities without human intervention.
Schmidhuber highlights that modern RSI systems are finally learning to (re)define the starts and ends of their own trials, moving beyond human-defined experimental constraints.
A team of researchers from DeepMind, Meta, Liquid, and Stanford are actively building "Radical Numerics," an engine designed specifically for AI to design AI.