ICONIP 2026 Accepted paper

ClueWeaver

Reward-Guided Dual-Agent Evidence Reasoning
for Compact LLMs on Literary Long Narratives

Jihao Zhu2,*Zhiwei Yang1,*Wenxiao Zhang3,*Junqian Zhao1Qi You1 Fangqi Wang1Zheyuan Deng4Hanzhe Yang3Yu Liu1,†Jin B. Hong3,†
1 Institute of Information Engineering, CAS2 University of Aberdeen3 The University of Western Australia4 Brown University

* Equal contribution   ·   † Corresponding authors

Paper Code Model weights

Finding the clues. Connecting the evidence.

A compact Finder selects the evidence.
An Interpreter connects it into a grounded answer.

Retrieval-guided segmentation → evidence selection → evidence-grounded interpretation.
Both agents use Qwen3-4B-Instruct. Self-calibration is internal to the Interpreter.

INSIDE CLUEWEAVER

From scattered clues to a grounded answer.

01 / 07
CLAIM TO VERIFY

“Leon could unlock the attic after Mara left.”

TRUE / FALSE

The narrativeParagraph-indexed source

FinderEvidence selection

InterpreterGrounded reasoning

INPUT

One claim. Clues hundreds of paragraphs apart.

The answer depends on the lock, the key changing hands, and who still has it later.

Story paragraphs

Selected excerpts
  1. [12]

    Only the brass key could unlock the attic.

  2. [41]

    Rain drummed against the windows all evening.

  3. [86]

    Mara was leaving for the harbor that evening.

  4. [87]

    Before she left, Mara handed the brass key to Leon.

  5. [88]

    He put it inside his coat pocket.

  6. [203]

    Hours after Mara left, Leon still carried the brass key.

Answer-critical clueNearby contextDiscarded

Input to the pipeline

Narrative + claim

The source narrative and claim enter the pipeline together. Ground-truth labels are not part of the input.

Fictional, scripted illustration of inference, not a live model run or benchmark result. Candidate windows, relevance decisions and the 900-character evidence budget are illustrative. Both agents are trained separately with GRPO; self-calibration reuses the Interpreter and the same evidence packet.

THE IDEA

Abstract

Humanities and social science research requires close reading of long narrative materials such as novels, scripts, archives, and case reports, yet many users have limited access to costly proprietary long-context models. Compact, locally deployable language models are a practical alternative, but directly feeding them an entire long context remains costly, hard to inspect, and prone to missing sparse evidence.

We present ClueWeaver, an evidence-aware dual-agent framework for long-narrative question answering with compact local models. A Finder identifies passages containing answer-critical clues through retrieval-guided segmentation, while an Interpreter derives the answer from the selected evidence, produces rationales with paragraph-ID citations, and applies an internal self-calibration pass for high-risk questions.

Both agents are optimized with reward-guided reinforcement learning: Finder rewards emphasize evidence retention and faithful paragraph-ID references, and Interpreter rewards emphasize correctness, grounding, and concise explanations. This decomposition makes evidence selection and reasoning more inspectable than end-to-end prompting. Experiments across multiple long-context narrative question answering and claim verification settings show that ClueWeaver substantially improves local end-to-end language models while providing evidence coverage and paragraph-referenced reasoning traces.

METHOD

Two agents, complementary objectives.

Read the methodology ↗
01 / FINDER

Retain answer-critical clues

Retrieval guides the construction of narrative segments. The Finder makes a YES/NO evidence decision, with a short rationale and supporting paragraph IDs.

Reward focusEvidence retention · citation faithfulness

02 / INTERPRETER

Reason from surviving evidence

Selected passages are packed in narrative order. The Interpreter answers with paragraph-grounded reasoning and can re-check high-risk answers against the same evidence.

Reward focusAnswer correctness · grounding · concision

Both agents are trained separately with GRPO. The source narrative is given; retrieval operates within it.

EXPERIMENTS

Stronger compact-model reading.

Four long-context narrative benchmarks. Final answer accuracy.

59.0%

Overall accuracy

+14.5pp

vs. Qwen3-4B direct reading

+6.4pp

vs. strongest local baseline overall

Table 1 · Main results

Accuracy percentages from Table 1 of the paper. ClueWeaver is highlighted as the proposed method, not as the best API result.
Method DetectiveQA ∞Bench LongBench v2 NoCha Overall
End-to-end reader · Local
Qwen3-4B-Instruct 36.5 50.7 38.5 49.5 44.5
Qwen3-8B 45.2 56.5 26.9 55.0 49.7
Ministral-3-14B 45.2 60.9 26.9 51.4 49.4
GPT-OSS-20B 30.8 33.3 34.6 50.5 38.7
Qwen3-30B-A3B 44.2 58.0 38.5 54.1 50.3
Gemma-4-31B-it 35.6 58.0 46.2 59.5 50.0
End-to-end reader · API
Claude Haiku 4.5 62.5 73.9 38.5 64.9 63.9
GPT-5 nano 62.5 76.8 26.9 60.4 61.9
Agentic pipelines · Local
ReAct 50.0 50.7 38.5 58.6 52.3
IRCoT 53.8 44.9 34.6 60.4 52.6
Self-Ask 36.5 49.3 19.2 59.5 46.1
Chain-of-Agents 27.9 46.4 38.5 55.9 42.9
RAG-DDR 47.1 53.6 30.8 59.5 51.6
ClueWeaver Ours 55.8 63.8 50.0 61.3 59.0

310 questions: DetectiveQA (104), ∞Bench (69), LongBench v2 (26), NoCha (111). Overall accuracy is weighted by question count. Local end-to-end readers use 32K context; API readers use 128K. API readers remain stronger overall. Baseline implementations are detailed in Appendix C.

A CLOSER LOOK

Where the gains come from.

Fixed: reader wrong, ClueWeaver correct. Regression: the reverse.
Numbers are question counts; bar widths show percentages.

OPEN RELEASE

Explore ClueWeaver.

Core inference code, reward functions, and both agent checkpoints.

Code & quick start ↗Finder & Interpreter weights ↗Reproduction guide ↗

REFERENCE

BibTeX

@misc{zhu2026clueweaver,
  title={ClueWeaver: Reward-Guided Dual-Agent Evidence Reasoning for Compact LLMs on Literary Long Narratives},
  author={Jihao Zhu and Zhiwei Yang and Wenxiao Zhang and Junqian Zhao and Qi You and Fangqi Wang and Zheyuan Deng and Hanzhe Yang and Yu Liu and Jin B. Hong},
  year={2026},
  eprint={2608.25531},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2608.25531}
}

Acknowledgements

This research is supported by the National Key R&D Program of China (No. 2023YFC3303800). We thank WisPaper / QiewenPaper for Academic Agent support and GPU computational resources throughout the study.

Figure