Google's ProceduralGraphs wires tools, skills and memory into evolvable graphs, lifting agent survival 6% to 34%

新智元 · wechat · 2026-09-21

Researchers from Google, Georgia Tech and Peking University propose ProceduralGraphs (PG), representing tool calls, skills, reasoning and task states as a directed graph of "process–relation–process" triples whose edges carry conditions, guidance and pitfalls. Paper: arxiv.org/pdf/2609.09153

Design: at runtime the agent locates its current node, reads the local subgraph within two outgoing hops, and generates contextual guidance; offline, a revision model proposes graph edits validated against held-out tasks before adoption — no weight updates needed. Notably, a hand-written expert graph hurt MultiChallenge success (87.5%→58.93%) while evolved graphs reached 92.86%.

Results: BFCL v3 accuracy 67% vs 58% best baseline; in EnterpriseArena's 132-month business simulation, Gemini 3.1 Pro survival rose 6%→34% and Claude Sonnet 4.6 44%→58%. Local-subgraph guidance beat full-graph guidance on ALFWorld (54.48%→81.53%) with 70.9% fewer tokens. Limitations: small gains on QA tasks and extra offline compute cost.

Related event: Google's Procedural Graphs boost LLM agent long-task performance(2 posts)→

Original post →

More from coding & agent

coding & agent channel →