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
TL;DR
Flat RAG treats corpora as independent chunks and lacks recovery when initial retrieval fails. SearchWiki compiles structured wikis and trains WikiResearcher-9B to navigate them through multi-turn tool use and reinforcement learning. WikiResearcher-9B achieves leading ViDoRe v3 performance and strong transfer to FinanceBench and memory benchmarks.
Problem
Flat RAG discards document hierarchy and cross-document structure and provides no recovery mechanism when initial retrieval misses the answer.
Method
SearchWiki compiles corpora into typed three-layer wikis and trains WikiResearcher-9B for multi-turn navigation with reinforcement learning.
Results
WikiResearcher-9B ranks first overall on ViDoRe v3 end-to-end QA accuracy at 71.35% and reaches 83.33% accuracy on FinanceBench.
Takeaways & Limitations
The results support learned navigation over structured corpora as an alternative to flat retrieval across document, financial, and memory benchmarks.
Abstract
from arXiv · showhide
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
Flat RAG retrieves independent chunks in a single shot, discarding corpus structure and offering no recovery when initial retrieval misses. SearchWiki instead compiles knowledge into a navigable wiki and trains an agent to search it interactively.
- Flat RAG treats a corpus as an undifferentiated bag of independently indexed chunks.
- When top-k retrieval misses the answer, flat RAG cannot reformulate searches, backtrack, or explore related corpus regions.
- SearchWiki compiles raw corpora into document overviews, cross-document topic pages, and page-level source records.
- The three-layer layout supports progressive disclosure from summaries or topics to specific source pages for verification.
- SearchWiki trains multi-turn navigation with DAPO, Retrieved Token Masking, and rewards balancing correctness, retrieval quality, and trajectory efficiency.
2 Related Work
Related work spans flat RAG, interactive search agents, structured memory systems, hierarchical document indexing, and reinforcement learning for tool-using language models. SearchWiki combines corpus wikis with learned multi-turn navigation and a task-specific reward stack.
- Retrieval-augmented generation: Canonical RAG retrieves top-k passages from a vector-indexed corpus and conditions generation on that context.
- Agentic search and reinforcement learning: Agentic search methods train models to issue queries or operate corpus tools, while SearchWiki navigates a pre-compiled wiki through interactive search and reading.
- Structured memory versus corpus wikis: Structured memory systems organize user histories as records, graphs, or hierarchies, whereas SearchWiki applies hierarchical navigation to general knowledge corpora.
- RL for tool-using language models: SearchWiki combines GRPO-style group advantages, DAPO optimization, RTM masking, and a domain-specific dense reward stack.
3 SearchWiki
SearchWiki first compiles unstructured sources into a typed, navigable wiki, then trains an agent to navigate it through multi-turn tool use. Its representation exposes hierarchical summaries, cross-document topics, and exact source records for progressive retrieval and verification.
- 3.1 Overview: SearchWiki has two stages: wiki compilation transforms an unstructured corpus into a typed three-layer wiki, and corpus navigation uses an RL-tuned agent.
- 3.1 Overview: The agent constructs multi-turn trajectories of actions, observations, and a terminal answer while dynamically inspecting nodes, descending to details, and backtracking.
- Knowledge Representation: The wiki contains document overviews, cross-document topic pages, and per-document source-page records linked through the corpus hierarchy.
- Tool-Based Navigation: Content is progressively disclosed through search and catalog reads, summary inspection, and read_page calls for exact evidence.
- Compilation Pipeline: Long documents are structured from TOCs into nested section summaries tied to source-page ranges, while short documents use a direct single-summary route.
- Policy Optimization via Reinforcement Learning: DAPO trains multi-turn tool trajectories using group rollouts, asymmetric clipping, token masking, and rewards for answer quality, retrieval quality, and trajectory discipline.
- Policy Optimization via Reinforcement Learning: Retrieved Token Masking excludes tool observations from policy loss while updating model-generated planning, tool-call, and answer tokens.
4 Experimental Setup
The evaluation tests answer and retrieval quality under the same tools and scheduler used for training, covering ViDoRe v3, FinanceBench, and dialogue-memory benchmarks. The setup also addresses rollout length and evaluates against same-size and larger baselines.
- Implementation Details: Training uses 1,600 synthetic queries for one epoch, initializing WikiResearcher-9B from Qwen3.5-9B.
- Context Budget and Reasoning Length: Long reasoning can exceed the 65,536-token rollout limit and receive zero reward, so training favors concise planning with scratchpad use.
- Evaluation: Evaluation measures answer quality with LLM judging and retrieval quality with page- and document-level ranking metrics under the training-matched harness.
- Datasets and Benchmarks: The benchmark suite includes all eight ViDoRe v3 domains, FinanceBench, LoCoMo, LongMemEval, and PersonaMem-v2.
- Comparative Baselines: Comparisons include the untrained same-size Qwen3.5-9B and larger Qwen3.6-27B, Qwen3.5-397B, Gemma-4-31B, and MiniMax-M2.5 models.
5 Results and Analysis
SearchWiki performs strongly across document, memory, and finance evaluations, while ablations show that typed hierarchy, multi-turn interaction, and reinforcement learning each contribute to retrieval and answer quality.
- ViDoRe v3: 71.35±0.73 ViDoRe QA accuracy places WikiResearcher-9B ahead of untrained Qwen3.6-27B at 70.94±0.72 and Qwen3.5-397B at 69.51±0.73.The score is reported under the official ViDoRe LLM-as-judge prompt.
- ViDoRe v3: WikiResearcher-9B surpasses Qwen3.5-397B, a same-family model with 44× more parameters, on page-level recall and nDCG.The paper reports these retrieval gains at the document level and notes that per-domain breakdowns are supplementary.
- Memory benchmarks: On memory benchmarks, WikiResearcher-9B is competitive with NapMem, reaching 80.67 LongMemEval LLM-J versus NapMem’s 80.33 without memory-specific training.The untrained 27B harness also exceeds NapMem on PersonaMem-v2, while SearchWiki leads NapMem on LoCoMo F1 and LLM-judge.
- FinanceBench: WikiResearcher-9B reaches 83.33% FinanceBench accuracy, near the 85.33% GPT-4-Turbo Oracle upper bound and above leaderboard baselines at or below 19.30%.Within the shared SearchWiki harness, it trails Qwen3.6-27B† by 2 points and Qwen3.5-397B by 1.34 points.
- Ablations: 70.94 with the untrained 27B model exceeds flat-hierarchy, dense-RAG, and BM25-RAG scores of 65.47, 62.61, and 29.60, respectively.With the untrained 9B model, SearchWiki reaches 63.95 versus 57.30 for dense RAG and 50.39 without the typed hierarchy.
- Ablations: Hierarchy and multi-turn interaction contribute independently, while reinforcement learning raises the 9B model’s score to 71.35 beyond model scaling alone.The ablation results identify gains from both structural organization and interaction, not merely from replacing BM25 with dense retrieval.
6 Conclusion
SearchWiki replaces passive flat RAG with active corpus navigation through a typed, three-layer wiki and online reinforcement learning for multi-turn tool use. WikiResearcher-9B achieves strong results across visual document, financial, and memory benchmarks.
- SearchWiki compiles unstructured corpora into document overviews, cross-document topic pages, and page-level source records.The framework supports progressive refinement, evidence verification, and dynamic backtracking during retrieval.
- WikiResearcher-9B ranks first overall on ViDoRe v3 with 71.35% end-to-end QA accuracy, outperforming models up to 44× its parameter scale.It also leads page-level retrieval ranking with 53.22% nDCG@5.
- WikiResearcher-9B achieves 83.33% accuracy on FinanceBench, approaching the 85.33% Oracle upper bound.
- Without memory-specific training, WikiResearcher-9B surpasses domain-specialized memory policies on LoCoMo, PersonaMem-v2, and LongMemEval.These results support learned wiki navigation as a generalizable alternative to chunk-based retrieval.