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

Jihao Zhu; Zhiwei Yang; Wenxiao Zhang; Junqian Zhao; Qi You; Fangqi Wang; Zheyuan Deng; Hanzhe Yang

cs.CLAug 26, 2026

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

Authors: Jihao Zhu, Zhiwei Yang, Wenxiao Zhang, Junqian Zhao, Qi You, Fangqi Wang, Zheyuan Deng, Hanzhe Yang,

Organizations: University of Aberdeen, Aberdeen, United Kingdom · Institute of Information Engineering, CAS, Beijing, China · The University of Western Australia, Perth, Australia · Brown University, Providence, Rhode Island, United States

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. Code is available at https://github.com/Ameame1/ClueWeaver.

Figures & tables

Figure 1: Answer-critical clues in long narratives are often sparse and distant. While long-context models may connect them directly, compact local models operating over limited context windows can truncate evidence, separate related clues, and obscure evidence use.

Figure 2: Overview of ClueWeaver: Finder selects evidence-bearing narrative segments, and Interpreter produces a grounded answer with self-calibration.

End-to-end reader (Local)
Method	DetectiveQA	$\infty$ Bench	LongBench v2	NoCha	Overall
Qwen3-4B-Instruct [ 31 ]	36.5	50.7	38.5	49.5	44.5
Qwen3-8B [ 31 ]	45.2	56.5	26.9	55.0	49.7
Ministral-3-14B [ 17 ]	45.2	60.9	26.9	51.4	49.4
GPT-OSS-20B [ 18 ]	30.8	33.3	34.6	50.5	38.7
Qwen3-30B-A3B [ 31 ]	44.2	58.0	38.5	54.1	50.3

Table 1: Main results on four long-context narrative benchmarks (final answer accuracy, %). Among local methods, best per column is in bold and second best is underlined ; best API result per column is marked with wavy underlines .

Configuration	Acc.	$\Delta$
ClueWeaver (full)	55.8	–
w/o Interpreter ${}_{\mathrm{self\text{-}cal}}$	51.0	$-4.8$
w/o Finder	50.0	$-5.8$
w/o Finder & Interpreter	36.5	$-19.3$

Table 2: Ablation on DetectiveQA. (a) pipeline components and (b) RL training are removed from the full model (

\Delta

: accuracy change vs. the full model).

Figure 3: Accuracy–latency trade-off using Table 1 accuracies and serial single-GPU stage timings. Most additional latency comes from the Finder stage.

End-to-end reader (Local)
Setting	Overall Acc.	Access	Example GPU
Qwen3-4B [ 31 ]	44.5	16GB	RTX 5060 Ti 16GB
Qwen3-8B [ 31 ]	49.7	24GB	RTX 4090 / RTX 5090
Ministral-3-14B [ 17 ]	49.4	24GB	RTX 4090 / RTX 5090
GPT-OSS-20B [ 18 ]	38.7	24GB	RTX 4090 / RTX 5090
Qwen3-30B-A3B [ 31 ]	50.3	80GB	A100/H100

Table 3: Cost profile for local and API readers.

Figure 4: Error transitions between the direct reader and ClueWeaver across the four benchmarks. ClueWeaver yields a clear net correction gain.

Question. The reason why the robot doll was pushed downstairs is: (A) Break it for study (B) Eliminate clues (C) Hurt him (D) No particular purpose [Gold: B]
End-to-end reader selects the surface-level distractor: D $\times$
ClueWeaver keeps the decisive evidence and then self-calibrates the Interpreter output.
Finder (retained evidence from the log): YES segments include [1203], where the killer is described as wiping physical traces, and [1326], where Dr. Phil links this to pushing the robot doll downstairs.
Interpreter : <reason>Option (B) is supported by [1203], where the killer erases footprints, fingerprints, and dust traces; [1326] links this reasoning to why the robot doll was pushed downstairs.</reason> <answer>B</answer>
Interpreter ${}_{\mathrm{self\text{-}cal}}$ : <reason>The two cited passages are consistent: [1203] establishes the motive of removing physical traces, and [1326] applies that motive to the pushed doll. Thus the answer is eliminating clues. </reason> <answer>B</answer> $\checkmark$

Table 4: Case study on DetectiveQA. ClueWeaver exposes the evidence-selection and interpretation process, with self-calibration used as a final consistency check.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Component	Purpose	Score
Format validity	Output must follow the required structured format and expose a decision and rationale.	$+1.0$
Positive decision	A gold-bearing segment should be kept by predicting YES .	$+1.5$
Paragraph-ID F1	Reward overlap between predicted and gold paragraph IDs, using F1 over $P_{j}$ and $G_{j}$ .	up to $+1.5$
Compact ID bonus	Extra credit when the Finder references gold paragraphs without adding many irrelevant paragraph IDs.	$+0.25$
Correct rejection	A non-evidence segment should be rejected by predicting NO .	$+1.25$
Negative rationale quality	Reward non-trivial NO rationales with no paragraph ID and explicit irrelevance cues.	up to $+0.75$

Appendix

Table 5: Finder reward components.

Component	Purpose	Score
Format validity	Output must contain a parseable rationale and normalized answer field.	$+1.0$
Answer correctness	Reward the normalized final answer. Multiple-choice questions receive a larger gain than binary verification because their random baseline is lower.	$+2.5$ MCQ; $+1.5$ binary
Hard-case bonus	Additional reward for correctly solving examples marked as hard residual errors during training.	$+0.5$
Paragraph-ID use	Reward rationales that reference paragraph IDs from the supplied evidence. Invalid paragraph IDs are penalized when the allowed paragraph set is known.	$+0.3/-0.3$
Rationale length	Encourage concise but non-trivial rationales rather than bare answers or long summaries.	$+0.2$
Specificity	Reward concrete grounding signals such as quoted phrases, numbers, or named entities from the evidence.	$+0.3$

Appendix

Table 6: Interpreter reward components.

Figure 5: t-SNE of BGE-M3 embeddings of the training questions for the two agents, colored by source. Both pools are detective-domain dominated; the NoCha portion is mostly synthetic claim-verification, and LongBench v2 is the sparsest benchmark for the Interpreter.

Agent Parameter	Finder	Interpreter
base model	Qwen3-4B	Qwen3-4B
algorithm	GRPO	GRPO
reward	evidence selection	answer grounding
lr	$5\times 10^{-7}$	$3\times 10^{-7}$
bs	$8$	$8$
grad_accum	$8$	$2$

Appendix

Table 7: GRPO training hyperparameters for each agent.

Stage	Operation	Output
Retrieval anchors	Score paragraphs using dense retrieval and lexical matching. Select high-scoring paragraphs as anchors.	Candidate anchor paragraphs
Segment construction	Expand each anchor with nearby paragraphs and keep paragraph IDs. Add local windows to avoid missing surrounding context.	Ordered candidate segments $s_{j}$
Finder selection	Predict YES / NO , cite paragraph IDs, and emit a brief rationale for each segment.	Clue-bearing segments and paragraph IDs
Evidence packing	Merge selected evidence, remove duplicates, retain nearby supporting paragraphs, and sort by original narrative order.	Compact evidence packet $E$
Interpreter input	Combine $q$ , answer options when available, and packed evidence with paragraph IDs.	Grounded final-answer prompt

Appendix

Table 8: Evidence construction stages in ClueWeaver.

Dataset	Citation-bearing outputs	Valid cited IDs	All citations valid
DetectiveQA	98/104	235/237 (99.2%)	96/98 (98.0%)
$\infty$ Bench	65/69	155/155 (100.0%)	65/65 (100.0%)
LongBench v2	24/26	76/76 (100.0%)	24/24 (100.0%)
NoCha	88/111	219/222 (98.6%)	85/88 (96.6%)
Overall	275/310	685/690 (99.3%)	270/275 (98.2%)

Appendix

Table 9: Citation audit for final ClueWeaver traces. Citation validity is computed only for outputs with explicit paragraph citations.

Type	Question	Prediction	Observation
Fixed error	Henry’s wife Sylvia on his drug use situation.	Direct: C ; ClueWeaver: A ; Gold: A	The Finder retrieves distant paragraphs indicating that Sylvia did not know about Henry’s drug use, allowing the Interpreter to reject plausible but unsupported alternatives.
Regression	What was the cause of John Fairley’s death?	Direct: D ; ClueWeaver: A ; Gold: D	The selected evidence emphasizes a local suicide explanation and misses the option-specific detail needed to support the gold answer.
Unresolved	Tina says: “The cup is empty.” What does this mean?	Direct: C ; ClueWeaver: C ; Gold: A	The literal evidence about the empty cup is found, but both systems fail to map the symbolic statement to the intended narrative implication.

Appendix

Table 10: Additional qualitative cases.

Long-Context QA

Long-context question answering studies how models find, retain, and reason over evidence spread across lengthy documents and histories.

See all 39 papers

Narrative Knowledge Weaver: Narrative-Centric Retrieval-Augmented Reasoning for Long-Form Text Understanding
Jun 4, 2026Qiuyu Tian, Fengyi Chen, Yiding Li +9Retrieval-Augmented Generation Long-Context QA
Southeast University, Nanjing, China · Nanjing Normal University, Nanjing, China · ZhuiWen Technology Co., Ltd., Beijing, China +1
Beyond Linear Context: Graph-Guided Evidence Navigation for Long-Novel Reasoning with a Local 9B Language Model
Sep 19, 2026Wenji FuLong-Context QA LLM Reasoning with Graphs
Research Institute of Economics and Management Southwestern University of Finance and Economics
Highlight-Then-Summarize: Learning to Compress Evidence for Long-Context Understanding
Sep 25, 2026Zhaoyuan Xia, Qinghongbing Xie, Yung Xiang Hue +7Long-Context Reasoning in Language Models Long-Context Modeling
Peking University · Baidu Inc · Baidu Inc. +1

Jun 4, 2026J/K move · Enter open · S save

Narrative Knowledge Weaver: Narrative-Centric Retrieval-Augmented Reasoning for Long-Form Text Understanding

Qiuyu Tian, Fengyi Chen, Yiding Li, Youyong Kong+8

Southeast University, Nanjing, China · Nanjing Normal University, Nanjing, China · ZhuiWen Technology Co., Ltd., Beijing, China+1

Retrieval-Augmented Generation Long-Context QA Question Answering Knowledge Representation Multi-Hop Reasoning

Open paper AskPDF arXiv

Long-form narrative QA requires reasoning over evolving story worlds rather than isolated passages: answers may depend on earlier goals, changing character states, social relations, causal triggers, temporal position, and later consequences. Existing retrieval and graph-augmented generation methods improve evidence access, but their units--chunks, entities, relations, summaries, or tool actions--do not directly encode how evidence functions in a story. We introduce Narrative Knowledge Weaver(NKW), a source-grounded framework that aligns textual evidence, atomic facts, canonical graph structure, entity profiles, interactions, episodes, and storylines. At query time, NKW uses text, graph, and narrative tools with post-retrieval reading skills to assemble evidence and audit actor, scope, polarity, state, and temporal constraints. Across STAGE, FairytaleQA, and QuALITY, NKW is strongest on screenplay-level story-world QA while remaining competitive on more passage-centered benchmarks. Ablations, question-type analyses, graph-asset statistics, and case studies show complementary benefits for character, scene, temporal, causal, and narrative-progression reasoning.

Abstract

Long-Context QA

Explore similar work

Narrative Knowledge Weaver: Narrative-Centric Retrieval-Augmented Reasoning for Long-Form Text Understanding

Beyond Linear Context: Graph-Guided Evidence Navigation for Long-Novel Reasoning with a Local 9B Language Model

Highlight-Then-Summarize: Learning to Compress Evidence for Long-Context Understanding