Agent Reasoning, Memory, and AI-for-AI Research Advances

Three papers challenge existing approaches: MemHarness proposes memory reconstruction over replay, Memory Decoder at Scale shows parametric memory scaling is more efficient, and PhiZero introduces a 'reason-then-render' world model using physical language. SpatialCLI introduces a three-stage approach (Call, Learn, Internalize) for VLMs to reason with spatial tools then internalize them, achieving Qwen3-VL-8B with tools beating GPT-5.6 Sol. Additionally, Frontis-MA1 represents a concrete step toward recursive self-improvement in ML engineering, furthering the AI4AI trend alongside AlphaEvolve and GPT-Red. These advance agent reasoning, interpretable world models, tool internalization, and autonomous AI research.