跳到正文
研踪PaperTrail
← 研踪 PaperTrail

arXiv:2605.25379 · v4

StateRAG: Typed State Contracts for Complex Retrieval-Augmented Generation

Miaohe NiuXinyu LiuRunsong ZhaoPengxiang LiJingbo ZhuTong Xiao
cs.CL
首次提交 2026-05-25 11:08
版本更新 2026-11-10 21:00
雷达观测 2026-11-11 13:08
arXiv ↗PDF ↗

摘要

Multi-step retrieval-augmented generation requires tracking evolving information needs, selected evidence, and unresolved gaps. However, free-form trajectories conflate this history, forcing each step to implicitly reconstruct the current state, while existing explicit-state methods remain vulnerable to inconsistent state overwrites. To address these limitations, we introduce StateRAG, a framework that coordinates a Planner, Navigator, Retriever, and Verifier through an explicit retrieval state with checked updates. These roles propose updates across four state fields: the plan, retrieval trace, evidence, and sufficiency assessment. A dedicated controller validates each proposal against the accepted state, rejecting inconsistent updates and guiding revisions within the query budget. A final Reader then answers strictly from the validated evidence. Ablation experiments on a shared graph backend show that StateRAG outperforms free-form trajectory and unverified explicit-state baselines on LongBench by 7.7 and 4.1 Macro F1 points, respectively. These gains consistently extend to document QA and flat passage retrieval. Furthermore, StateRAG surpasses the strongest external baselines by 5.3, 5.0, and 4.3 points on LongBench, QASPER, and DocVQA.

查看 arXiv 原文 ↗阅读 PDF ↗

AI 论文解读

基于摘要 · 非全文审阅

此版本尚未生成 AI 解读。解读由后台任务生成并按论文版本保存，登录后共享已有结果。

版本差异

仅对比实际观测的元数据，不代表 PDF 正文或实验变化。版本可能不连续；带删除线为移除，下划线为新增。

作者变化

Miaohe Niu,Xinyu Liu,Runsong Zhao,Pengxiang Li,Jingbo Zhu,Tong Xiao

摘要变化

ComplexMulti-step retrieval-augmented generation requires evidence retrieval and control over what to retrieve next, whichtracking pathsevolving toinformation exploreneeds, whetherselected evidence is sufficient, and which intermediate results tounresolved retaingaps. Existing RAG paradigms encode these decisions throughHowever, methodfree-specificform modeltrajectories contexts,conflate traversalthis procedureshistory, verification signals,forcing andeach memory.step Weto introduceimplicitly StateRAG,reconstruct whichthe representscurrent retrievalstate, controlwhile asexisting aexplicit-state typedmethods stateremain externalvulnerable to theinconsistent finalstate readeroverwrites. TheTo stateaddress recordsthese thelimitations, querywe planintroduce StateRAG, typeda traversalframework path,that candidatecoordinates evidencea Planner, verificationNavigator, verdictRetriever, and reusable artifacts,Verifier withthrough definedan fieldexplicit semanticsretrieval andstate designatedwith updatechecked sourcesupdates. Ordered roleThese operatorsroles propose field values, and theupdates controlleracross validatesfour andstate commitsfields: acceptedthe proposals.plan, Aretrieval one-timetrace, compact-evidence check, mayand selectsufficiency Bypassassessment. Otherwise,A thededicated controller combines thevalidates committedeach verdictproposal withagainst the remaining budget to select Release,accepted Revisestate, or Fallback. Therejecting finalinconsistent readerupdates isand invokedguiding onlyrevisions afterwithin athe terminalquery actionbudget. andA atfinal mostReader oncethen peranswers query.strictly Amongfrom the evaluated methods, StateRAG achievesvalidated theevidence. bestAblation scoreexperiments on 10 ofa 11shared reportedgraph qualitybackend metricsshow acrossthat LongBench,StateRAG QASPER,outperforms andfree-form DocVQA.trajectory Itsand LongBenchunverified macroexplicit-average EM andstate F1baselines exceedon ReActLongBench by 67.37 and 64.1 percentage points, and its largest task-level gain is 9.8Macro F1 points on MuSiQue. Under normalized model-size weighting, StateRAG records lowerrespectively. meanThese LLM-tokengains useconsistently perextend queryto thandocument ReActQA acrossand allflat fivepassage workloadsretrieval. In a matched LongBench control, the typedFurthermore, role-owned carrierStateRAG exceedssurpasses athe free-formstrongest sharedexternal carrierbaselines by 15.3, 2.2, and 35.0 percentage points in macro EM, F1, and accuracy, with 64.2% lower mean3 RET.points Removingon TAMLongBench, MARSQASPER, or SMP lowers every reported LongBench and QASPER ablation metricDocVQA.

观测到的版本

相邻记录可能跳过中间版本；首次发现 v2 及以上也会标为版本更新。

v4 · 2026-11-10 21:00 · 最新观测版本

StateRAG: Typed State Contracts for Complex Retrieval-Augmented Generation

Multi-step retrieval-augmented generation requires tracking evolving information needs, selected evidence, and unresolved gaps. However, free-form trajectories conflate this history, forcing each step to implicitly reconstruct the current state, while existing explicit-state methods remain vulnerable to inconsistent state overwrites. To address these limitations, we introduce StateRAG, a framework that coordinates a Planner, Navigator, Retriever, and Verifier through an explicit retrieval state with checked updates. These roles propose updates across four state fields: the plan, retrieval trace, evidence, and sufficiency assessment. A dedicated controller validates each proposal against the accepted state, rejecting inconsistent updates and guiding revisions within the query budget. A final Reader then answers strictly from the validated evidence. Ablation experiments on a shared graph backend show that StateRAG outperforms free-form trajectory and unverified explicit-state baselines on LongBench by 7.7 and 4.1 Macro F1 points, respectively. These gains consistently extend to document QA and flat passage retrieval. Furthermore, StateRAG surpasses the strongest external baselines by 5.3, 5.0, and 4.3 points on LongBench, QASPER, and DocVQA.

查看此版本解读 →查看此版本原文 ↗
v3 · 2026-11-09 10:31

StateRAG: Typed State Contracts for Complex Retrieval-Augmented Generation

Complex retrieval-augmented generation requires evidence retrieval and control over what to retrieve next, which paths to explore, whether evidence is sufficient, and which intermediate results to retain. Existing RAG paradigms encode these decisions through method-specific model contexts, traversal procedures, verification signals, and memory. We introduce StateRAG, which represents retrieval control as a typed state external to the final reader. The state records the query plan, typed traversal path, candidate evidence, verification verdict, and reusable artifacts, with defined field semantics and designated update sources. Ordered role operators propose field values, and the controller validates and commits accepted proposals. A one-time compact-evidence check may select Bypass. Otherwise, the controller combines the committed verdict with the remaining budget to select Release, Revise, or Fallback. The final reader is invoked only after a terminal action and at most once per query. Among the evaluated methods, StateRAG achieves the best score on 10 of 11 reported quality metrics across LongBench, QASPER, and DocVQA. Its LongBench macro-average EM and F1 exceed ReAct by 6.3 and 6.1 percentage points, and its largest task-level gain is 9.8 F1 points on MuSiQue. Under normalized model-size weighting, StateRAG records lower mean LLM-token use per query than ReAct across all five workloads. In a matched LongBench control, the typed, role-owned carrier exceeds a free-form shared carrier by 1.3, 2.2, and 3.0 percentage points in macro EM, F1, and accuracy, with 6.2% lower mean RET. Removing TAM, MARS, or SMP lowers every reported LongBench and QASPER ablation metric.

查看此版本解读 →查看此版本原文 ↗