AIBRIEF
APEx splits a deep research agent's memory into raw run logs and reusable how to skills, refined by a reward driven training loop. Authors report a 14.7 point gain over GPT 5.4 on seven benchmarks.
APEx (Agent Procedural Experience, arXiv 2609.02253), a preprint posted this week, targets the kind of AI agent that takes a hard question and spends many turns browsing, reading, and coding before answering. The framework splits the agent's memory into two layers that stay in step.
The top layer stores raw trajectories from past runs: what the agent searched, what it read, what code it tried. The bottom layer holds procedural skills, learned how-to procedures for whole categories of tasks, distilled from those trajectories. A three-module closed loop (Executor, Distiller, Planner) keeps both layers talking to each other instead of letting skills drift from real runs.
Training uses a three-stage alternating GRPO (Group Relative Policy Optimization, a reinforcement-learning method) setup with reward-guided skill distillation, so the skill library updates from the agent's own reward signal rather than from fixed prompts. At test time, the distilled skills act as priors for online adaptation, with a skill-alignment regularizer preventing policy drift.
The authors report state-of-the-art results on seven benchmarks, claiming APEx beats GPT-5.4 by 14.7 points and the strongest memory-augmented baseline by 3.0 points. The figures are self-reported on a preprint with no independent replication yet. The closed-loop training adds compute overhead, and transfer to domains outside the tested set is unproven. Code is released at github.com/J-Ding519/APEx.