Hugging Face Papers ↗ · 2026-11-07
Read Less, Solve More: Token-Efficient Sparse Reading of External Content for AI Agents
Research Hugging Face Papers Source language: Chinese
论文摘要(作者报告,未经本站复测):Web-facing AI agents repeatedly access webpages, documents, repositories, and tool responses, yet reading interfaces often expose entire objects when only sparse evidence is needed.
Full text
Web-facingAIagentsrepeatedlyaccesswebpages,documents,repositories,andtoolresponses,yetreadinginterfacesoftenexposeentireobjectswhenonlysparseevidenceisneeded.Thisover-readingincreasestokenandlatencycostsandcandiluterelevantevidence.WepresentSparseRead,areusablereadingcontrollayerthatgovernscontentadmissionwithoutmodifyingtheagentmodel.ItsReadIntentProtocoltracksevolvingneedsandsourcecoverage;AdaptiveReaderBackendsselecthowevidenceislocatedandexposed;andtheBudgetedReadGateboundscumulativerawexposureperepisodewithfaithfulfull-viewfallback.AunifiedReaderinterfacesupportsinterchangeableevidenceacquisitionacrossheterogeneousresources.Acrosssixmodels,includingClaudeOpus5,andfivescenarios,SparseReadreducestotaltokensbyupto92.9%andwalltimebyupto89.0%,preservingorimprovingtaskqualityin97%settings.OnWebdeep-searchtasks,itreducestokensby31.0--43.0%andtaskcompletiontimeby23.0--37.0%.Ablationsvalidatethethreedesigns,whileintegrationwiththreeagentframeworksdemonstratesportability.
View arXiv pageView PDF Project page GitHub 63 Add to collection