工具、技能、记忆连成网，北大等提出自进化程序图PG，越用越聪明

Researchers from Google, Georgia Tech, and Peking University propose Procedural Graphs (PG), which organize tool calls, skills, memory read/writes, and task states into conditional "process-relation-process" triplets with fields for conditions, guidance, and pitfalls, enabling agents to decide when to act, stop, or recall saved context. PG operates via online local-subgraph guidance (matching the current step, reading two-hop neighbors) and offline graph evolution where candidate modifications are retained only if validated performance does not decline, with no model weight retraining. In evaluations using Gemini 3.5 Flash, PG raised MultiChallenge accuracy from a 58.93% expert-graph baseline to 92.86% via iterative evolution, improved BFCL v3 tool-calling accuracy to 67.00% versus a 58.00% baseline, and increased EnterpriseArena long-horizon survival in 50-run simulations from 6.0% to 34.0% for Gemini 3.1 Pro and from 44.0% to 58.0% for Claude Sonnet 4.6.

why now

北大等新发PG论文，Agent经验可离线演化，值得跟进。

topic

AI Tech & New Models

source

新智元