Injecting or generating guidance from the full graph can severely degrade performance on embodied tasks. On ALFWorld, full graph generative guidance drops success rate from 72.58% (baseline without graph) to 54.48% while dramatically increasing token usage from 18,055 to 96,360 tokens per sample, demonstrating the necessity of topological localization. · CiteArk
Injecting or generating guidance from the full graph can severely degrade performance on embodied tasks. On ALFWorld, full graph generative guidance drops success rate from 72.58% (baseline without graph) to 54.48% while dramatically increasing token usage from 18,055 to 96,360 tokens per sample, demonstrating the necessity of topological localization.
