Source-linked AI summary
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, Yu Liu, Jin B. Hong
TL;DR
Long-narrative QA challenges compact local models because relevant clues are sparse, distant, and difficult to preserve explicitly. ClueWeaver separates evidence selection from grounded interpretation using two reward-trained agents, and experiments report substantial improvements over local readers with inspectable evidence traces.
Problem
Long-narrative questions require connecting sparse, distant evidence, while direct full-context prompting can truncate, obscure, or implicitly handle supporting clues.
Method
ClueWeaver uses retrieval-aware segmentation, a Finder for answer-critical passages, and an Interpreter for paragraph-grounded answers with optional self-calibration and reward-guided training.
Results
ClueWeaver substantially improves compact locally deployable models across multiple long-context narrative settings while providing inspectable evidence traces.
Takeaways & Limitations
Explicit evidence paths support literary and humanities analysis with compact models while remaining deployable on commercial GPUs.
Takeaways & Limitations
Future work targets harder distant clues and more robust multi-hop evidence integration.
Abstract
from arXiv · showhide
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.
1 Introduction
Long-narrative question answering requires compact models to find and connect sparse, distant clues while making supporting evidence explicit. ClueWeaver addresses this with a reward-guided dual-agent pipeline that separates clue selection from grounded interpretation.
- Long narratives distribute meaning across characters, events, motives, temporal order, and causal links over hundreds of paragraphs.
- Compact local models face truncated, separated, or obscured evidence when answering questions over long narratives.
- ClueWeaver uses retrieval-aware segments, a Finder that selects answer-critical passages, and an Interpreter that connects clues into grounded answers.
- The retained passages preserve paragraph IDs and narrative order, enabling inspectable evidence use and easier localization of failures across the pipeline.
- Reward-guided reinforcement learning rewards the Finder for clue retention and faithful references, while rewarding the Interpreter for correct, grounded, compact rationales.
- Experiments across multiple long-context narrative settings show substantial improvements for compact locally deployable models with inspectable evidence traces.
2 Related Work
Prior work organizes retrieval, reasoning, agent collaboration, and post-training, but these approaches largely target open-domain retrieval, general multi-hop QA, or broad long-context aggregation. ClueWeaver instead focuses on evidence selection and paragraph-grounded reasoning within an already supplied narrative.
- Narrative benchmarks evaluate questions over books, scripts, novels, and long contexts, showing that long-narrative understanding exceeds local phrase matching or shallow salience.
- Broader long-context benchmarks indicate that larger context windows alone do not ensure robust use of information buried inside the context.
- RAG and its agentic variants retrieve external passages, while IRCoT, Self-Ask, Chain-of-Agents, and RAG-DDR structure retrieval and reasoning into staged procedures.
- Existing methods mainly address open-domain retrieval, multi-hop question answering, general long-context aggregation, or preference-oriented post-training rather than narrative evidence selection.
- ClueWeaver’s overview centers on a Finder selecting evidence-bearing narrative segments and an Interpreter producing a grounded answer with self-calibration.
3 Methodology
ClueWeaver decomposes long-narrative question answering into retrieval-aware evidence selection and evidence-grounded interpretation for compact local models. The Finder preserves answer-critical clues with paragraph references, while the Interpreter produces grounded answers and optionally self-calibrates high-risk questions; both agents use agent-specific rewards.
- Problem Definition: ClueWeaver formulates long-narrative reading as selecting an ordered compact evidence set from indexed paragraphs, then generating an answer and paragraph-traceable explanation.The task supports multiple-choice and binary verification answers, using only the supplied narrative.
- Overview: The pipeline segments the narrative with retrieval guidance, uses the Finder to retain clue-bearing passages, packs them in narrative order, and lets the Interpreter answer from the resulting evidence packet.Self-calibration rechecks the provisional answer against the same packet for binary claims and higher-risk question forms.
- Finder: The Finder predicts a clue decision, referenced paragraph IDs, and a rationale, using high-recall selection so indirect or distributed supporting clues are not discarded.YES segments enter the candidate clue pool, while paragraph indices are retained for later evidence tracing.
- Finder: Finder rewards combine valid formatting, decision accuracy, paragraph-ID overlap, compact references, and concise negative rationales, prioritizing evidence retention over aggressive filtering.A missed positive segment receives only the format reward, while unsupported IDs in NO rationales are discouraged.
- Interpreter: The Interpreter maps the ordered evidence packet to a provisional answer, rationale, and paragraph references, optionally applying an internal self-calibration pass.The packet retains paragraph indices, evidence text, and Finder rationales; self-calibration is triggered for binary verification and risky wording.
- Training Principle: Interpreter rewards favor valid output, gold-answer matching, valid paragraph citations, concise grounded rationales, and resistance to unsupported references or uncertainty.The two agents are trained with the same GRPO principle but with different reward targets.
4 Experiments
Experiments evaluate ClueWeaver across four long-context narrative benchmarks, comparing local and API readers, component and training ablations, efficiency, error correction, and qualitative grounding. ClueWeaver achieves the strongest local results while selected evidence, self-calibration, and reward-guided training each contribute to performance.
- Main Results: ClueWeaver achieves 59.0% overall accuracy, leads all local methods on every dataset, and improves over IRCoT by +6.4 points.With the same Qwen3-4B backbone, it improves over direct reading by +14.5 points; Qwen3-30B-A3B remains 8.7 points behind.
- Ablation Study: On DetectiveQA, disabling Interpreterself-cal lowers accuracy by 4.8 to 51.0%, while removing the Finder drops accuracy by 5.8 points to 50.0%.Removing both agents reduces accuracy to 36.5%.
- Ablation Study: Removing Finder RL drops accuracy by 6.8 points to 49.0%, compared with a 1.0-point loss when removing Interpreter RL.Removing both RL components returns accuracy to 50.0%, indicating Finder training contributes the larger share.
- Efficiency and Analysis: ClueWeaver answers in 8.6–9.8 seconds per question versus 2.8 seconds for direct reading, with most additional latency coming from Finder calls.The reported trade-off combines higher accuracy with practical local deployment and evidence-level inspection.
- Efficiency and Analysis: ClueWeaver recovers 84 of 172 direct-reader errors, producing a consistently positive net correction gain across the four benchmarks.Remaining difficult cases involve distant multi-hop clues or weak surface overlap.
- Case Study: Qualitative comparisons show direct reading selecting a surface-level distractor, whereas ClueWeaver grounds the answer in a previously identified relevant paragraph.Self-calibration appears as a final consistency check within the Interpreter.
5 Conclusion
ClueWeaver separates evidence selection from interpretation for compact-model long-narrative QA, using reward-guided training and self-calibration to produce grounded answers. Experiments show improved local readers, stronger comparisons with larger local models, and deployment on commercial GPUs.
- ClueWeaver separates Finder-based evidence retention from Interpreter-based grounded answer generation and high-risk self-calibration.The framework preserves sparse clue passages and uses paragraph-referenced reasoning traces.
- Experiments show that ClueWeaver improves local end-to-end readers, outperforms much larger local models, and approaches stronger API readers.
- ClueWeaver exposes evidence selection and interpretation, with self-calibration serving as a final consistency check in DetectiveQA.
- The system remains deployable on commercial GPUs while supporting literary and humanities analysis through explicit evidence paths.Future work targets harder distant clues and more robust multi-hop evidence integration.
A.1 Reward Components
The reward design gives Finder and Interpreter complementary objectives within a structured, value-free GRPO training process. Finder prioritizes faithful clue retention, while Interpreter makes correctness a prerequisite for rationale-quality rewards.
- Rule-based rewards use labels, paragraph indices, output format, and parsed XML-style responses rather than a learned reward model.Unparseable outputs receive zero reward, while valid outputs receive a small format-validity base reward.
- Finder reward favors positive evidence retention and correct paragraph IDs while preserving meaningful reward for concise negative decisions.This balances recall against an all-YES policy.
- Interpreter reward is correctness-dominant, gating rationale-quality bonuses on a correct answer to discourage fluent but wrong explanations.
- The complementary rewards push Finder to preserve sparse clues and Interpreter to convert retained evidence into correct, grounded, concise answers.
- GRPO samples complete structured outputs, normalizes rewards within each sampled group, applies a reference-model KL term, and trains without a separate critic or value model.The value-free formulation is used for both agents.
B.1 Training Data
Training uses disjoint benchmark training splits and paragraph-level supervision for long narratives. Finder training emphasizes balanced evidence decisions, while Interpreter training mixes scarce real cases with synthetic and hard examples.
- Training data come from benchmark training splits disjoint from the test set, with documents segmented at the paragraph level.Both agents are supervised on segments because source narratives often exceed 100K tokens.
- Finder training uses balanced 50/50 keep/drop decisions with hard negatives from detective-novel QA and claim-verification data.
- Interpreter training uses 1,000 evidence-packet examples whose mixture follows the test distribution.Scarce real NoCha cases are augmented with synthetic claim verification, and 27.5% are base-model failures.
- The question pools are detective-domain dominated, while LongBench v2 is the sparsest benchmark for the Interpreter.
B.2 Training Setup
Both agents use Qwen3-4B-Instruct and are trained separately with GRPO under specified decoding and optimization settings. Inference memory requirements allow sequential one-GPU or concurrent multi-GPU serving.
- Both agents use Qwen3-4B-Instruct and are trained separately with GRPO using eight sampled responses per prompt.Training uses full model weights in bfloat16 with cosine learning-rate decay and specified sampling settings.
- Training runs on a single node with 8 x NVIDIA A100 GPUs, with vLLM colocated to accelerate rollout generation.
- Each active Qwen3-4B inference instance uses about 10–12 GB of GPU memory in bfloat16.The agents can therefore be served sequentially on one GPU or concurrently on separate GPUs.
C.1 Baseline Implementations
ClueWeaver combines retrieval-based evidence construction with Finder and Interpreter agents, alongside local and retrieval-based baselines. Its implementation emphasizes paragraph-level traceability, bounded evidence packets, structured outputs, and targeted self-calibration.
- Baseline models: Local agentic baselines use Qwen3-4B-Instruct, while larger end-to-end readers use their own model weights.The direct reader receives the narrative within the configured context budget and answers in one call.
- Comparison baselines: ReAct, IRCoT, Self-Ask, Chain-of-Agents, and RAG-DDR provide retrieval, decomposition, iterative reasoning, multi-agent, and reward-based comparison baselines.In the closed-document setting, these methods retrieve or process evidence from the given narrative rather than an external corpus.
- Evidence construction: ClueWeaver scores paragraphs with BGE-M3 and complementary lexical matching, then expands top paragraphs into retrieval-anchored local-window segments preserving paragraph IDs.The Finder judges each candidate segment for useful clues and cites supporting paragraph IDs, removing many retrieval-only false positives while keeping evidence traceable.
- Evidence budgets: Evidence packets are bounded by NE, Pr, Pw, and Bc; focused settings use NE = 5, Pr = 2, Pw = 5, Bc = 13,000, while broader settings use NE = 10, Pr = 4, Pw = 6, Bc = 15,000.These configurations represent a trade-off between noise control and broader evidence coverage.
- Traceability and format audit: 99.3% of rowunique cited paragraph IDs are valid, 98.2% of citation-bearing outputs contain only valid citations, and legal answer labels are extracted in 99.4% of outputs.Across 310 questions, XML tags appear in 309 outputs (99.7%), while 275 outputs contain paragraph citations.