Microsoft Research turned agent skill files into trainable artifacts. SkillOpt raised GPT-5.5’s six-benchmark direct-chat average from 58.8 to 82.3 and improved all or tied for best across 52 evaluation cells without updating model weights.
Microsoft Research introduced Memora as an agent memory system that separates what is stored from how it is retrieved. Its research post says Memora outperforms Mem0, RAG, and full-context inference on LoCoMo and LongMemEval while using up to 98% fewer context tokens.
Microsoft Research, UC Berkeley, UCSF, and Columbia researchers introduced generative causal testing. The method converts black-box language-brain prediction models into short hypotheses, then tests them by having an LLM write stories meant to activate targeted cortical regions.