PaperlayerView exact revision on arXiv

Source-linked AI summary

SearchWiki: Learning to Build and Navigate Knowledge Wikis for Active Information Seeking

Guransh Singh, Vishwajeet Kumar, Arkadeep Acharya, Adnan Qidwai, Jaydeep Sen, Sachindra Joshi

arXiv:2608.29953v1cs.AI

TL;DR

Flat retrieval-augmented generation treats corpora as bags of chunks and offers no iterative recovery when initial retrieval misses. SearchWiki compiles a typed, three-layer wiki and trains WikiResearcher-9B for multi-turn navigation; it ranks first overall on ViDoRe v3 end-to-end QA accuracy at 71.35%.

  • Problem

    Flat RAG treats corpora as undifferentiated chunks and lacks an iterative mechanism to reformulate search or explore related corpus regions when initial retrieval misses.

    Source: §1 · p. 1

  • Method

    SearchWiki compiles corpora into a typed, three-layer wiki and trains WikiResearcher-9B with online reinforcement learning for multi-turn navigation.

    Source: §6 · p. 7

  • Results

    On ViDoRe v3 and FinanceBench, WikiResearcher-9B ranks first overall on ViDoRe v3 end-to-end QA accuracy at 71.35% and approaches FinanceBench’s Oracle upper bound.

    Source: §6 · p. 7

  • Takeaways & Limitations

    Ablations indicate that hierarchy and multi-turn interaction each contribute independently beyond simply replacing BM25 with a stronger dense retriever.

    Source: §5.4 · p. 7

Abstract

from arXiv · show

Flat retrieval-augmented generation treats a corpus as a bag of chunks, discarding document hierarchy and cross document structure. We introduce SearchWiki, a harness framework that synthesizes a corpus into a hierarchical, typed, navigable wiki and trains an agent, WikiResearcher-9B, to retrieve information through multi-turn tool use. The wiki organizes knowledge into three layers - document overviews, cross- document topic pages, and page-level source records; enabling progressive refinement of retrieval when initial lookup misses. We optimize the agent's navigation policy with on-policy reinforcement learning with a multi-component reward function balancing answer correctness, retrieval quality and trajectory efficiency. Evaluation on ViDoRe-V3 (8 domains), FinanceBench, and memory benchmarks (LoCoMo, LongMemEval, PersonaMem-v2) shows that WikiResearcher- 9B which is our RL-tuned Qwen 9B model, significantly outperforms same-size untrained baselines and exceeds or matches larger external models. SearchWiki paired with WikiResearcher-9B demonstrates that learned navigation over structured corpora is a superior alternative to flat retrieval.

1 Introduction

SearchWiki addresses the limits of one-shot flat RAG by compiling corpora into a navigable, typed wiki and training an agent to search it through multi-turn tool use.

1 Introduction

  • Flat RAG retrieves independent chunks, discarding document hierarchy and cross-references, and cannot reformulate, backtrack, or explore when initial retrieval misses.The introduction argues that complex corpora require structured, multi-step information seeking.Source: §1 · p. 1
  • SearchWiki compiles corpora into document-overview, cross-document topic, and page-level source layers, enabling progressive disclosure through search, file and page reading, and grep tools.The layers provide entry-point summaries, links across documents and topics, and detailed source records for verification.Source: §1 · p. 2
  • WikiResearcher-9B learns multi-turn wiki navigation with on-policy DAPO and RTM, while a multi-component reward balances answer correctness, navigation efficiency, and backtracking.RTM restricts policy gradients to model-generated planning, tool-call, and answer tokens rather than tool observations.Source: §1 · p. 2
  • Evaluation covers eight ViDoRe v3 splits and FinanceBench, LoCoMo, LongMemEval, and PersonaMem-v2, comparing against untrained baselines and larger external models.The paper also presents a domain-agnostic wiki compilation and navigation interface and an online RL training recipe.Source: §1 · p. 2

2 Related Work

SearchWiki differs from flat retrieval and prior agentic search by having agents navigate a typed, pre-compiled wiki through domain-agnostic tools. Its related-work context spans structured memory and document indexing, retrieval-focused reinforcement learning, and RL methods for tool-using models.

Retrieval-augmented generation

  • SearchWiki replaces a single flat context pack with an interactive search-and-read loop over a pre-compiled wiki.ViDoRe (Faysse et al. 2025) emphasizes retrieval over visually complex, multi-page documents.Source: §2 · p. 2

Agentic search and reinforcement learning

  • Unlike search-query agents such as Search-R1 (Jin et al. 2025) and DeepRetrieval (Jiang et al. 2025b), SearchWiki traverses a typed wiki rather than a flat search API or raw text.The wiki links document summaries, cross-document topics, and page-level sources, using domain-agnostic tools.Source: §2 · p. 2

Structured memory versus corpus wikis

  • On the same wiki representation, SearchWiki with untrained large and ViDoRe-trained WikiResearcher-9B models outperforms NapMem on key LoCoMo metrics and leads on PersonaMem-v2.The passage presents this as evidence that structured navigation transfers beyond the training domain; it also situates SearchWiki alongside memory systems (Packer et al. 2023; Chhikara et al. 2025), AgeMem (Yu et al. 2026b), NapMem (Xu et al. 2026), PageIndex (Zhang, Tang, and Team 2025), and LLM-wiki.Source: §2 · p. 2

RL for tool-using language models

  • SearchWiki combines GRPO (Shao et al. 2024), DAPO (Yu et al. 2025), and RTM-style masking (Jin et al. 2025) with a SearchWiki-specific dense reward stack.These ingredients respectively estimate advantages without a critic, stabilize long trajectories, and avoid policy updates on environment tokens.Source: §2 · p. 2

3 SearchWiki

SearchWiki compiles an unstructured corpus into a typed, navigable wiki and trains an RL-tuned agent to seek information through multi-turn tool interaction. The agent uses intermediate observations to decide what to inspect, when to descend to page-level evidence, and when to backtrack.

3.1 Overview

  • SearchWiki combines offline wiki compilation with multi-turn agent navigation, allowing the policy to choose nodes, descend to page details, and backtrack using intermediate observations.The framework uses the resulting navigation policy at inference.Source: §3.1 · pp. 2–3

3.2 Wiki Compilation and Ingestion

  • SearchWiki wikis are built offline through a fully autonomous LLM compilation process.The compilation process transforms raw knowledge sources into the typed wiki layout.Source: §3.2 · p. 3

Knowledge Representation.

  • The wiki preserves corpus structure through document overviews, cross-document topic pages, page-level source records, and catalog indexes.Content is disclosed progressively through search and catalog reads, summary inspection, and on-demand page reads; the same structure can represent dialogue sessions and turns.Source: Knowledge Representation · p. 3

Compilation Pipeline.

  • Compilation parses page layout, uses TOCs to build document hierarchies, and jointly generates aligned summaries and registered topics to reduce information loss and duplicate concepts.Documents are compiled in parallel, while chronological sources are processed sequentially to maintain a shared topic registry.Source: Compilation Pipeline · p. 3

Tool-Based Navigation.

  • The navigation protocol supports multi-turn tool calls and requires terminal answers to include explicit document-and-page citations, with malformed trajectories down-weighted during training.The framework figure describes the harness loop as search, read, and verify.Source: Tool-Based Navigation · p. 4, Figure 1 · p. 4

Policy Optimization via Reinforcement Learning.

  • Training uses online multi-turn rollouts and excludes retrieved tool responses from policy loss, restricting updates to model-generated planning, tool calls, and answers.The policy update uses DAPO, with GDPO for group reward scaling.Source: Policy Optimization via Reinforcement… · p. 4
  • The two-stage recipe first optimizes the full search-and-QA reward stack, then updates only final-answer tokens while retaining correctness reward to prevent navigation degradation.Rewards combine correctness, reading and retrieval quality, and structural behavior; the paper reports that QA-only training without the token mask collapses navigation learning.Source: Policy Optimization via Reinforcement… · pp. 4–5

4 Experimental Setup

SearchWiki trains WikiResearcher-9B on filtered synthetic queries and evaluates it across in-distribution and out-of-distribution benchmarks under a consistent tool environment. The setup measures answer quality and retrieval quality while accounting for multi-turn context limits.

4.1 Training Setup

  • The three-stage synthetic-data pipeline retained approximately 25% of generated questions, producing 1,800 instances for training and validation.A teacher LLM generated questions from gold evidence; deterministic checks and an LLM judge filtered them. The dataset has 225 queries per domain, split into 1,600 training and 200 validation samples.Source: §4.1 · p. 5

Implementation Details.

  • WikiResearcher-9B starts from Qwen3.5-9B and trains for one epoch on 1,600 synthetic queries across eight A100 GPUs, peaking at a learning rate of 10−6.Training uses DeepSpeed ZeRO-3, sequence parallelism of 4, and a colocated vLLM generation engine (Rajbhandari et al. 2020).Source: Implementation Details · p. 5

Context Budget and Reasoning Length.

  • To avoid zero-reward truncation at the 65,536-token rollout limit, training favors concise step-by-step planning with a scratchpad over unconstrained reasoning.Long chains of thought can inflate context usage across multi-turn trajectories and destabilize training despite appearing high-quality.Source: Context Budget and Reasoning Length · p. 5

4.2 Evaluation Framework

  • Answer quality uses GPT-OSS-120B judging and strict accuracy, while retrieval quality uses Recall@5 and nDCG@5 at page and document levels.ViDoRe v3 uses its official three-way judge prompt; out-of-distribution tasks use NapMem’s binary prompt or multiple-choice accuracy (Loison et al. 2026; Xu et al. 2026). Rankings place cited evidence first, then pages or documents accessed through read_page or read_file; document-level metrics are more robust when answers come from summaries or topic pages.Source: §4.2 · p. 5
  • Evaluation covers all eight ViDoRe v3 domains, FinanceBench, and LoCoMo, LongMemEval, and PersonaMem-v2.Memory benchmarks use NapMem’s user-disjoint 60/20/20 splits, token F1, LLM judge, and multiple-choice accuracy setup (Xu et al. 2026).Source: §4.2 · p. 5
  • All models use the same SearchWiki tools and environment; comparisons include WikiResearcher-9B, untrained Qwen3.5-9B, and larger baselines.Larger baselines are Qwen3.6-27B, Qwen3.5-397B, Gemma-4-31B, and MiniMax-M2.5 (Gemma Team 2026).Source: §4.2 · p. 5

5 Results and Analysis

WikiResearcher-9B leads ViDoRe QA, performs competitively on memory benchmarks, and transfers strongly to FinanceBench. Ablations show gains from SearchWiki’s hierarchical representation and multi-turn interaction over passive retrieval.

5.1 Results on ViDoRe v3

  • 71.35±0.73 ViDoRe strict QA accuracy leads untrained Qwen3.6-27B (70.94±0.72) and Qwen3.5-397B (69.51±0.73), and is +7.4 points over same-size Qwen3.5-9B (63.95±0.76).Paired bootstrap comparisons put the 9B and 397B gains well clear of zero; the smaller edge over 27B still favors WikiResearcher-9B.Source: §5.1 · pp. 5–6, Table 2 · p. 6

5.2 Performance on Memory Benchmarks

  • 80.67 LongMemEval LLM-J for WikiResearcher-9B is competitive with NapMem’s 80.33, while SearchWiki leads NapMem on LoCoMo and the untrained 27B harness exceeds it on PersonaMem-v2.The benchmarks use NapMem’s evaluation protocol and user-level splits; WikiResearcher-9B is ViDoRe-trained, not memory-specifically trained.Source: §5.2 · p. 6, Table 3 · p. 6

5.3 Generalization to FinanceBench

  • 83.33% FinanceBench accuracy exceeds Islam et al. (2023) leaderboard baselines (all ≤19.30%) and approaches the GPT-4-Turbo Oracle (85.33%).Within the shared SearchWiki harness, WikiResearcher-9B trails untrained Qwen3.6-27B (85.33%) by 2 points and Qwen3.5-397B (84.67%) by 1.34, despite training solely on ViDoRe.Source: Table 5 · p. 7, §5.3 · p. 7

5.4 Ablations

  • 70.94 for untrained 27B SearchWiki beats flat hierarchy (65.47), dense RAG (62.61), and BM25 RAG (29.60); WikiResearcher-9B reaches 71.35.For untrained 9B, SearchWiki scores 63.95 versus 57.30 for dense RAG and 50.39 without typed hierarchy; hierarchy and multi-turn interaction each contribute independently. SearchWiki enables query revision, summary and topic-link inspection, literal search with grep, and evidence verification with read_page.Source: Table 6 · p. 7, §5.4 · p. 7

6 Conclusion

SearchWiki replaces passive flat RAG with active navigation over a typed, three-layer wiki, while WikiResearcher-9B learns multi-turn retrieval through online reinforcement learning. The reported results support this approach across document retrieval, FinanceBench, and memory benchmarks.

6 Conclusion

  • 71.35% end-to-end QA accuracy on ViDoRe v3 ranks first overall, outperforming models up to 44× its parameter scale; page-level retrieval reaches 53.22% nDCG@5.The comparison is reported for WikiResearcher-9B.Source: §6 · p. 7
  • 83.33% FinanceBench accuracy approaches the Oracle upper bound of 85.33%.This result is reported for WikiResearcher-9B.Source: §6 · p. 7
  • WikiResearcher-9B surpasses domain-specialized memory policies on LoCoMo, PersonaMem-v2, and LongMemEval without memory-specific training.Source: §6 · p. 7
  • SearchWiki compiles unstructured corpora into a typed, three-layer wiki and trains WikiResearcher-9B via online RL for multi-turn tool use, evidence verification, and dynamic backtracking.The layers comprise document overviews, cross-document topic pages, and page-level source records.Source: §6 · p. 7

Related papers

Loading 2608.29953v1…