Two recent Peking University papers, VeriSkill and SESA, demonstrate that treating agent skills as self‑evolving memory—updated from failures via responsibility attribution, lesson abstraction, and failure distillation—yields significant performance gains across verification and search tasks and transfers across models.
SkillRevise tackles the overestimation of LLM‑authored agent skills by breaking down complex search tasks, attaching evidence to verifiable sources, and introducing trace‑conditioned revisions that let engineers pinpoint and fix failures across retrieval, reasoning, and presentation layers.
The 2026 I/O keynote shows Google abandoning the race for the strongest model, unveiling the mid‑tier Gemini 3.5 Flash that outperforms its flagship on benchmarks, cuts inference cost dramatically, and launches a suite of agents—including Gemini Spark and Antigravity 2.0—to build an ecosystem that reshapes AI competition.
The paper introduces DLLM‑Searcher, which equips diffusion large language models with a two‑stage training pipeline and a P‑ReAct inference scheme, allowing the model to issue tool calls while simultaneously reasoning, yielding 14‑22% end‑to‑end speedup and matching or surpassing traditional autoregressive agents on multi‑hop QA benchmarks.
