Ablation of PG graph scope (full vs localized subgraph) and guidance consumption mode (raw injection vs generative) on MultiChallenge using Gemini 3.5 Flash. The localized generative subgraph configuration achieves superior performance while reducing token overhead relative to full graph generative guidance.

Ablation of PG graph scope (full vs localized subgraph) and guidance consumption mode (raw injection vs generative) on MultiChallenge using Gemini 3.5 Flash. The localized generative subgraph configuration achieves superior performance while reducing token overhead relative to full graph generative guidance. · CiteArk

测量结果

已评估 0/12 项

获得支持 0有争议或冲突 0尚未评估 12

保留同一结论下的完整结果组。已有评估排在前面；单项判断不替代整组结论。展开行查看条件、来源与历次证据。

每行是一项测量，分别显示论文报告、最近可用实测和累计证据状态。
指标与条件	论文报告	最近可用实测	证据判断
configuration: Baseline (no graph)	80.27%	—	尚未评估
configuration: Baseline (no graph)	6,629 tokens	—	尚未评估
configuration: Baseline (no graph)	3.87 steps	—	尚未评估
configuration: Full graph, raw injection	86.6%	—	尚未评估
configuration: Full graph, raw injection	10,164 tokens	—	尚未评估
configuration: Full graph, raw injection	4.54 steps	—	尚未评估
configuration: Full graph, generative	87.35%	—	尚未评估
configuration: Full graph, generative	14,434 tokens	—	尚未评估
configuration: Full graph, generative	3.08 steps	—	尚未评估
configuration: Subgraph, generative (Ours)	89.31%	—	尚未评估
configuration: Subgraph, generative (Ours)	12,295 tokens	—	尚未评估
configuration: Subgraph, generative (Ours)	4.22 steps	—	尚未评估

报告指标与观测值

Assessment（0）

关联运行（0）

测量结果