AI Chronicles · Phase 1 · single-user development build AI agents often fail because their instructions, or skills, are manually modified with no guarantee of improvement. Learn how SkillOpt turns skill editing into a training process, making agent behavior more reliable without changing model weights. The post SkillOpt: Agent skills as trainable parameters appeared first on Microsoft Research.
Show more Yifan Yang, Xuemei Gao, Qi Dai, Bei Liu, Kai Qiu, Dongdong Chen, Chong Luo · 30 Jun 2026 (6d ago) Showing 1 · end of Companies
✦ This week in AI· AI-summarized Models — Introduction of BamiBERT, a new BERT-based model for Vietnamese, addressing limitations of existing models. — Research on PairCoder++ explores using pair programming for generating structured artifacts with LLMs. — Development of TUDUM, a Turkish-adapted reasoning pipeline for the Qwen3.5-27B model. Research — OpenSafeIntent paper evaluates intent-calibrated safe completions in dual-use prompt sets. — Study on the migration patterns of NLP research suggests a shift away from traditional NLP conferences. — New dataset AIriskEval-edu designed for risk assessment in AI-mediated educational explanations. Tools — Launch of promptdiff for Git-style version control of LLM prompts. — Agent-replay tool introduced for recording and debugging AI agent execution traces. — Agents-control-tower allows monitoring and controlling multiple Cursor agents from a single terminal. Discussion — Japan's top court rules that AI cannot be listed as an inventor on patent applications. — Debate on the absence of LLM code in dependencies raises concerns within the community. — Discussion on the capability of LLMs to watch videos, highlighting advancements in multimodal AI.