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

# ClueWeaver: Reward-Guided Dual-Agent Evidence Reasoning for Compact LLMs on Literary Long Narratives
Source: [https://arxiv.org/html/2608.25531](https://arxiv.org/html/2608.25531)
Jihao Zhu, Zhiwei Yang\*, Wenxiao Zhang\*, Junqian Zhao, Qi You, Fangqi††thanks:Equal contribution\.Affiliation:Institute of Information Engineering, CAS, Beijing, ChinaE\-mail[liuyu@iie\.ac\.cn](mailto:[email protected])Affiliation:University of Aberdeen, Aberdeen, United KingdomAffiliation:The University of Western Australia, Perth, AustraliaE\-mail[jin\.hong@uwa\.edu\.au](mailto:[email protected])\[\-1pt\]Wang, Zheyuan Deng, Hanzhe Yang, Yu Liu\(🖂\), Jin B\. Hong\(🖂\)Affiliation:Institute of Information Engineering, CAS, Beijing, ChinaE\-mail[liuyu@iie\.ac\.cn](mailto:[email protected])Affiliation:The University of Western Australia, Perth, AustraliaE\-mail[jin\.hong@uwa\.edu\.au](mailto:[email protected])Affiliation: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 presentClueWeaver, 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\.

###### Keywords:

Long\-context question answering Long\-narrative reasoning Evidence selection Multi\-agent reasoning

## 1Introduction

Novels, screenplays, investigation records, and long case reports require coherent reading over extended narratives\[[10](https://arxiv.org/html/2608.25531#bib.bib1),[23](https://arxiv.org/html/2608.25531#bib.bib11),[27](https://arxiv.org/html/2608.25531#bib.bib12),[9](https://arxiv.org/html/2608.25531#bib.bib13)\]\. Rather than collections of independent facts, their meaning emerges from characters, events, motives, temporal order, and causal links distributed over hundreds of paragraphs\. This setting is especially important for resource\-constrained literary and humanities research, where scholars may need to analyze novels, scripts, archives, or case materials without costly proprietary long\-context models or remote services\[[1](https://arxiv.org/html/2608.25531#bib.bib29),[26](https://arxiv.org/html/2608.25531#bib.bib30)\]\. Long\-narrative question answering therefore offers a practical testbed for strengthening compact, locally deployable models for long\-context reading\[touvron2023llama,jiang2023mistral,[28](https://arxiv.org/html/2608.25531#bib.bib23)\]\. Recent large language models, unlike earlier pretrained models with short context windows, can now accept much longer inputs in a single prompt\[[4](https://arxiv.org/html/2608.25531#bib.bib3),[30](https://arxiv.org/html/2608.25531#bib.bib14),[5](https://arxiv.org/html/2608.25531#bib.bib15)\]\. Yet window size alone does not ensure reliable narrative understanding\[[13](https://arxiv.org/html/2608.25531#bib.bib4),li2024loogle\]: narrative questions often require sparse, indirect clues spread across distant text spans\[[17](https://arxiv.org/html/2608.25531#bib.bib2),[27](https://arxiv.org/html/2608.25531#bib.bib12)\]\. Models must therefore identify salient story evidence\.

Figure[1](https://arxiv.org/html/2608.25531#S1.F1)illustrates this difficulty\. In long narratives, answer\-critical clues are often sparse and far apart, requiring models to connect them in narrative order\. Compact local models may instead rely on limited context windows, which can truncate evidence, separate related clues, and obscure evidence use\. As a result, failures may arise from missing or misusing key clues, not merely from generation errors\.

![Refer to caption](https://arxiv.org/html/2608.25531v1/moti.png)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\.This setting differs from standard retrieval\-augmented generation \(RAG\)\[[11](https://arxiv.org/html/2608.25531#bib.bib5)\]\. While typical RAG retrieves external documents from a large corpus, long\-narrative question answering assumes that the source text is already given and requires the model to find, preserve, and use sparse evidence within it\. Direct full\-context prompting is a natural baseline, but it keeps evidence selection implicit\[[13](https://arxiv.org/html/2608.25531#bib.bib4),[5](https://arxiv.org/html/2608.25531#bib.bib15)\]: failures may result from missing relevant clues, discarding them among irrelevant details, or failing to reason over available evidence\. This raises a central question: how can a compact local model answer long\-narrative questions while making its supporting clues explicit?

Existing work has made this challenge visible, but does not fully answer the question above\. Long\-narrative and long\-document benchmarks show that questions over books, scripts, and passages require more than local phrase matching or shallow salience\[[10](https://arxiv.org/html/2608.25531#bib.bib1),[17](https://arxiv.org/html/2608.25531#bib.bib2)\]\. Broader long\-context evaluations further show that simply increasing input length does not ensure robust use of relevant information, especially when evidence is buried inside the context\[[4](https://arxiv.org/html/2608.25531#bib.bib3),[13](https://arxiv.org/html/2608.25531#bib.bib4)\]\. RAG improves knowledge\-intensive generation by retrieving external passages\[[11](https://arxiv.org/html/2608.25531#bib.bib5)\]\. IRCoT interleaves retrieval with chain\-of\-thought reasoning\[[22](https://arxiv.org/html/2608.25531#bib.bib6)\], Self\-Ask decomposes questions into follow\-up queries\[[18](https://arxiv.org/html/2608.25531#bib.bib7)\], Chain\-of\-Agents distributes long\-context reading across collaborating agents\[[31](https://arxiv.org/html/2608.25531#bib.bib8)\], and RAG\-DDR optimizes RAG modules with data\-driven rewards\[[12](https://arxiv.org/html/2608.25531#bib.bib9)\]\. However, these methods mainly target open\-domain retrieval, general multi\-hop question answering, or broad long\-context aggregation\. They do not directly address the setting where the source narrative is already given, answer\-critical clues are sparse and distributed, and evidence selection itself should be explicit and inspectable\. Full\-context prompting and single\-stage readers, meanwhile, leave evidence selection implicit and couple evidence locating with answer generation\. As a result, they provide limited control over whether the model is reasoning from the right evidence\. A more suitable framework should expose evidence selection and evidence\-grounded reasoning as separate, inspectable, and trainable steps\.

To this end, we proposeClueWeaver, a dual\-agent pipeline for compact local models in literary and humanities\-oriented long\-narrative reading\. Rather than compressing a long document into a single latent state,ClueWeaverbuilds retrieval\-aware narrative segments and uses a Finder to select passages with answer\-critical clues\. These passages are packed with paragraph IDs and narrative order preserved, giving the Interpreter readable evidence for connecting clues, answering the question, and, when needed, applying Interpreterself​\-​cal\{\}\_\{\\mathrm\{self\\text\{\-\}cal\}\}as a final consistency check\. This separation makes failures easier to localize across candidate construction, clue selection, evidence packing, and final reasoning\. We further optimize both agents with reward\-guided training: the Finder is rewarded for retaining supporting clues and faithful paragraph references, while the Interpreter is rewarded for correct, grounded, and compact rationales\. In this setting, reward\-guided reinforcement learning provides a practical way to align intermediate evidence decisions with task\-level outcomes, consistent with recent progress in feedback\- and reward\-based post\-training for instruction following and reasoning\[[21](https://arxiv.org/html/2608.25531#bib.bib10)\]\. Our contributions are three\-fold:

1. 1\.We proposeClueWeaver, an evidence\-aware agentic pipeline that decomposes long\-narrative question answering into explicit evidence selection, self\-calibrated interpretation, and evidence\-grounded explanation for compact local models\.
2. 2\.We optimize both the Finder and the Interpreter with reward\-guided reinforcement learning, encouraging high\-recall evidence retention, faithful paragraph referencing, and reliable answer generation\.
3. 3\.Experiments across multiple long\-context narrative understanding settings show thatClueWeaversubstantially improves compact locally deployable language models while providing inspectable evidence traces\.

## 2Related Work

### 2\.1Agentic\-Assisted Literary and Long\-Context Reading

AI\-assisted literary reading targets long narratives whose interpretation depends on plot, characters, temporal order, and implicit causality\. NarrativeQA\[[10](https://arxiv.org/html/2608.25531#bib.bib1)\]and QuALITY\[[17](https://arxiv.org/html/2608.25531#bib.bib2)\]introduced QA over books, scripts, and long contexts; NovelQA\[[23](https://arxiv.org/html/2608.25531#bib.bib11)\], DetectiveQA\[[27](https://arxiv.org/html/2608.25531#bib.bib12)\], and NoCha\[[9](https://arxiv.org/html/2608.25531#bib.bib13)\]move further toward novel\-scale QA and claim verification\. Broader long\-context benchmarks, including LongBench\[[4](https://arxiv.org/html/2608.25531#bib.bib3)\],∞\\inftyBench\[[30](https://arxiv.org/html/2608.25531#bib.bib14)\], LooGLE\[li2024loogle\], and LongBench v2\[[5](https://arxiv.org/html/2608.25531#bib.bib15)\], show that larger context windows do not ensure robust long\-range understanding, especially when relevant information is buried in the context\[[13](https://arxiv.org/html/2608.25531#bib.bib4)\]\. However, these studies evaluate final answers and give less attention to clue discovery and evidence preservation\.

### 2\.2Agentic RAG and LLM Reasoning Pipelines

Retrieval\-Augmented Generation \(RAG\)\[[11](https://arxiv.org/html/2608.25531#bib.bib5)\]grounds generation in retrieved external passages\. Agentic variants add adaptive control: ReAct\[[29](https://arxiv.org/html/2608.25531#bib.bib16)\]combines reasoning and actions, FLARE\[[8](https://arxiv.org/html/2608.25531#bib.bib17)\]retrieves during generation, and Self\-RAG\[[3](https://arxiv.org/html/2608.25531#bib.bib18)\]adds retrieval and self\-critique\. Multi\-step pipelines such as IRCoT\[[22](https://arxiv.org/html/2608.25531#bib.bib6)\], Self\-Ask\[[18](https://arxiv.org/html/2608.25531#bib.bib7)\], Chain\-of\-Agents\[[31](https://arxiv.org/html/2608.25531#bib.bib8)\], and RAG\-DDR\[[12](https://arxiv.org/html/2608.25531#bib.bib9)\]further organize retrieval and reasoning into staged procedures\. These methods mainly target open\-domain retrieval, multi\-hop QA, or general long\-context aggregation: IRCoT and Self\-Ask build query or sub\-question chains, Chain\-of\-Agents compresses chunks through a running communication summary, and Self\-RAG/RAG\-DDR regulate general RAG behavior\. ClueWeaver instead targets closed\-document long narratives, where the source text is already given and the central issue is preserving sparse story clues\. It trains a Finder for paragraph\-level evidence decisions, keeps selected clues in narrative order with citations, and optimizes clue retention, citation fidelity, and grounded final answers rather than introducing a new retrieval primitive\.

### 2\.3Reward\-Guided Reinforcement Learning for Reasoning

Post\-training methods can improve model behavior\. SFT\[[25](https://arxiv.org/html/2608.25531#bib.bib20)\]adapts models to instruction formats, PPO\-based RLHF\[[20](https://arxiv.org/html/2608.25531#bib.bib21)\]optimizes preference rewards, and DPO\[[19](https://arxiv.org/html/2608.25531#bib.bib19)\]offers a simpler preference\-optimization path\. DeepSeekMath\[[21](https://arxiv.org/html/2608.25531#bib.bib10)\]shows that GRPO\-based RL can strengthen reasoning\. Meanwhile, answer attribution analysis\[[24](https://arxiv.org/html/2608.25531#bib.bib22)\]suggests that final answers may mix explicit reasoning with retrieval\-like memorized knowledge\. However, reward\-guided training remains less focused on evidence selection, paragraph\-level attribution, and grounded explanation in long\-narrative reading\.

## 3Methodology

![Refer to caption](https://arxiv.org/html/2608.25531v1/main.png)Figure 2:Overview of ClueWeaver: Finder selects evidence\-bearing narrative segments, and Interpreter produces a grounded answer with self\-calibration\.### 3\.1Problem Definition

We formulate agentic support for long\-narrative reading as question answering with compact language models for resource\-constrained humanities settings\. LetX=\{p1,p2,…,pm\}X=\\\{p\_\{1\},p\_\{2\},\\ldots,p\_\{m\}\\\}be a narrative, where eachpip\_\{i\}is an indexed paragraph andmmis the number of paragraphs\. Given a question or claimqq, the model predicts answeryyusing onlyXX;yymay be a multiple\-choice option or a binary verification label\. Since only a small part ofXXsupportsyy, the task is to select a compact evidence setE⊂XE\\subset X, preserve its narrative order, and generate bothyyand an explanation whose claims can be traced to paragraph\-referenced evidence inEE\.

### 3\.2Overview

Figure[2](https://arxiv.org/html/2608.25531#S3.F2)illustrates the pipeline, and Algorithm[1](https://arxiv.org/html/2608.25531#alg1)gives the inference procedure\. ClueWeaver does not ask a compact model to read the whole narrative in one pass\. It first builds retrieval\-aware segments, uses theFinderto keep clue\-bearing passages with paragraph IDs, packs the retained evidence in narrative order, and lets theInterpreterproduce a paragraph\-grounded answer\. For binary claims and high\-risk question forms, the sameInterpreteroptionally runsIθi,self​\-​calI\_\{\\theta\_\{i\},\\mathrm\{self\\text\{\-\}cal\}\}, a self\-calibration mode that re\-checks the provisional answer against the identical evidence packet\. In Algorithm[1](https://arxiv.org/html/2608.25531#alg1),𝒴\\mathcal\{Y\}is the answer space,BBis the evidence budget,θf\\theta\_\{f\}andθi\\theta\_\{i\}are the parameters of the two agents,𝒮\\mathcal\{S\}is the segment set,EEis the final evidence packet, andτ⁡\(q\)\\tau\(q\)triggers self\-calibration\. In the implementation, both agents use a compact XML interface:<reason\>contains the paragraph\-referenced explanation and<answer\>contains either a YES/NO decision or the final answer\. Complete prompts are in Appendix E\.

Algorithm 1ClueWeaver Inference1:Narrative

X=\{pi\}i=1mX=\\\{p\_\{i\}\\\}\_\{i=1\}^\{m\}, question or claim

qq, answer space

𝒴\\mathcal\{Y\}, evidence budget

BB
2:XML\-style output

zI=\(r,y,𝒞\)z^\{I\}=\(r,y,\\mathcal\{C\}\), with paragraph references extracted from

rr
3:

𝒮=\{sj\}j=1n←Segment⁡\(X,q\)\\mathcal\{S\}=\\\{s\_\{j\}\\\}\_\{j=1\}^\{n\}\\leftarrow\\mathrm\{Segment\}\(X,q\)⊳\\trianglerightnnis the number of retrieval\-aware segments

4:Initialize candidate evidence pool

C←∅C\\leftarrow\\emptyset
5:for

j=1j=1to

nndo

6:Finder:

zjF←Fθf​\(sj,q\)z\_\{j\}^\{F\}\\leftarrow F\_\{\\theta\_\{f\}\}\(s\_\{j\},q\)
7:

zjF=⟨reason⟩uj⟨/reason⟩⟨answer⟩dj⟨/answer⟩z\_\{j\}^\{F\}=\\begin\{array\}\[t\]\{l\}\\langle\\texttt\{reason\}\\rangle u\_\{j\}\\langle/\\texttt\{reason\}\\rangle\\\\ \\langle\\texttt\{answer\}\\rangle d\_\{j\}\\langle/\\texttt\{answer\}\\rangle\\end\{array\}
8:Parse

zjFz\_\{j\}^\{F\}into

\(uj,dj\)\(u\_\{j\},d\_\{j\}\)and extract referenced paragraph IDs

eje\_\{j\}
9:if

dj=YESd\_\{j\}=\\mathrm\{YES\}then

10:

C←C∪\{\(sj,ej,uj\)\}C\\leftarrow C\\cup\\\{\(s\_\{j\},e\_\{j\},u\_\{j\}\)\\\}
11:endif

12:endfor

13:Order

CCby the paragraph indices inherited from

XX
14:Build

E=Pack⁡\(C,B\)E=\\mathrm\{Pack\}\(C,B\)by keeping ordered clues within budget

BB
15:Interpreter:

z^I←Iθi,ans​\(E,q,𝒴\)\\hat\{z\}^\{I\}\\leftarrow I\_\{\\theta\_\{i\},\\mathrm\{ans\}\}\(E,q,\\mathcal\{Y\}\)
16:

z^I=⟨reason⟩r⟨/reason⟩⟨answer⟩y⟨/answer⟩\\hat\{z\}^\{I\}=\\begin\{array\}\[t\]\{l\}\\langle\\texttt\{reason\}\\rangle r\\langle/\\texttt\{reason\}\\rangle\\\\ \\langle\\texttt\{answer\}\\rangle y\\langle/\\texttt\{answer\}\\rangle\\end\{array\}
17:if

τ⁡\(q\)=1\\tau\(q\)=1then

18:

zI←Iθi,self​\-​cal​\(E,q,𝒴,z^I\)z^\{I\}\\leftarrow I\_\{\\theta\_\{i\},\\mathrm\{self\\text\{\-\}cal\}\}\(E,q,\\mathcal\{Y\},\\hat\{z\}^\{I\}\)
19:else

20:

zI←z^Iz^\{I\}\\leftarrow\\hat\{z\}^\{I\}
21:endif

22:Parse

zIz^\{I\}into

\(r,y\)\(r,y\)and extract paragraph references

𝒞\\mathcal\{C\}from

rr
23:return

zI=\(r,y,𝒞\)z^\{I\}=\(r,y,\\mathcal\{C\}\)

### 3\.3Training Principle

We train the two agents with Group Relative Policy Optimization \(GRPO\)\[[21](https://arxiv.org/html/2608.25531#bib.bib10)\]\. Leta∈\{F,I\}a\\in\\\{F,I\\\}denote the agent, whereFFis theFinderandIIis theInterpreter\. The Finder input isxF=\(sj,q\)x^\{F\}=\(s\_\{j\},q\)with outputoF=zjFo^\{F\}=z\_\{j\}^\{F\}, while the Interpreter input isxI=\(E,q,𝒴\)x^\{I\}=\(E,q,\\mathcal\{Y\}\)with outputoI=zIo^\{I\}=z^\{I\}\. For each inputxax^\{a\}, the old policyπolda\\pi^\{a\}\_\{\\mathrm\{old\}\}samplesKKcandidate outputs\{oka\}k=1K\\\{o\_\{k\}^\{a\}\\\}\_\{k=1\}^\{K\}\. Each output is scored by the agent\-specific rewardRa​\(oka,xa\)R\_\{a\}\(o\_\{k\}^\{a\};x^\{a\}\), and its group\-normalized advantage is

Aka=Ra​\(oka,xa\)−meanl​Ra​\(ola,xa\)stdl​Ra​\(ola,xa\)\+ϵ,A\_\{k\}^\{a\}=\\frac\{R\_\{a\}\(o\_\{k\}^\{a\};x^\{a\}\)\-\\mathrm\{mean\}\_\{l\}R\_\{a\}\(o\_\{l\}^\{a\};x^\{a\}\)\}\{\\mathrm\{std\}\_\{l\}R\_\{a\}\(o\_\{l\}^\{a\};x^\{a\}\)\+\\epsilon\},\(1\)whereϵ\\epsilonis a small numerical constant\. Letρka=πθaa​\(oka\|xa\)/πolda​\(oka\|xa\)\\rho\_\{k\}^\{a\}=\\pi\_\{\\theta\_\{a\}\}^\{a\}\(o\_\{k\}^\{a\}\|x^\{a\}\)/\\pi\_\{\\mathrm\{old\}\}^\{a\}\(o\_\{k\}^\{a\}\|x^\{a\}\)andρ¯ka=clip⁡\(ρka,1−δ,1\+δ\)\\bar\{\\rho\}\_\{k\}^\{a\}=\\mathrm\{clip\}\(\\rho\_\{k\}^\{a\},1\-\\delta,1\+\\delta\)\. The clipped objective for agentaais

𝒥a​\(θa\)\\displaystyle\\mathcal\{J\}\_\{a\}\(\\theta\_\{a\}\)=𝔼xa,𝐨a∼πolda​\[1K​∑k=1K\(min⁡\(ρka​Aka,ρ¯ka​Aka\)−β​Dka\)\],\\displaystyle=\\mathbb\{E\}\_\{x^\{a\},\\,\\mathbf\{o\}^\{a\}\\sim\\pi^\{a\}\_\{\\mathrm\{old\}\}\}\\Bigg\[\\frac\{1\}\{K\}\\sum\_\{k=1\}^\{K\}\\Big\(\\min\(\\rho\_\{k\}^\{a\}A\_\{k\}^\{a\},\\bar\{\\rho\}\_\{k\}^\{a\}A\_\{k\}^\{a\}\)\-\\beta D\_\{k\}^\{a\}\\Big\)\\Bigg\],\(2\)whereθa\\theta\_\{a\}is the trainable policy parameter,δ\\deltais the clipping range,β\\betacontrols KL regularization, and the sample\-level KL estimator is

Dka=πrefa​\(oka\|xa\)πθaa​\(oka\|xa\)−log⁡πrefa​\(oka\|xa\)πθaa​\(oka\|xa\)−1\.D\_\{k\}^\{a\}=\\frac\{\\pi\_\{\\mathrm\{ref\}\}^\{a\}\(o\_\{k\}^\{a\}\|x^\{a\}\)\}\{\\pi\_\{\\theta\_\{a\}\}^\{a\}\(o\_\{k\}^\{a\}\|x^\{a\}\)\}\-\\log\\frac\{\\pi\_\{\\mathrm\{ref\}\}^\{a\}\(o\_\{k\}^\{a\}\|x^\{a\}\)\}\{\\pi\_\{\\theta\_\{a\}\}^\{a\}\(o\_\{k\}^\{a\}\|x^\{a\}\)\}\-1\.\(3\)This relative objective is useful for ClueWeaver because many valid rationales can exist for the same narrative question, while their usefulness can still be judged by task\-level signals\. We therefore use the same optimization form for both agents but instantiate agent\-specific rewards\.RFR\_\{F\}favors high\-recall clue retention, faithful paragraph\-ID references, and calibrated YES/NO decisions, giving much higher reward to retaining answer\-bearing evidence than to rejecting extra candidates\.RIR\_\{I\}favors answer correctness, paragraph\-grounded support, concise explanation, and resistance to unsupported inference\. Thus, training directly aligns the pipeline stages: preserving answer\-critical evidence before reasoning and converting the retained evidence into a grounded final answer\.

### 3\.4Finder: Evidence Selection and Rationale Generation

TheFinderis responsible for converting a long narrative into a small set of answer\-relevant evidence with short rationales\. Given the paragraph sequenceXXand the questionqq, we first build retrieval\-aware segments𝒮=\{sj\}j=1n\\mathcal\{S\}=\\\{s\_\{j\}\\\}\_\{j=1\}^\{n\}\. Each segmentsjs\_\{j\}contains a contiguous paragraph spanIj⊆\{1,…,m\}I\_\{j\}\\subseteq\\\{1,\\ldots,m\\\}\. The segmentation uses lexical and dense retrieval scores to place short anchor segments around paragraphs that are likely to be relevant toqq, while the remaining text is covered by local windows\. Retrieval is therefore used to guide segmentation boundaries, not to replace reading of the given narrative\. This keeps the input to theFindershort enough for a compact model, while retaining paragraph indices needed for later evidence tracing\. For each segmentsjs\_\{j\}, theFinderpredicts

\(dj,ej,uj\)=Fθf​\(sj,q\),\(d\_\{j\},e\_\{j\},u\_\{j\}\)=F\_\{\\theta\_\{f\}\}\(s\_\{j\},q\),\(4\)wheredj∈\{YES,NO\}d\_\{j\}\\in\\\{\\mathrm\{YES\},\\mathrm\{NO\}\\\}is the clue decision,ej⊆Ije\_\{j\}\\subseteq I\_\{j\}lists referenced paragraph IDs, anduju\_\{j\}is the rationale\. In the XML output,uju\_\{j\}is written in<reason\>,djd\_\{j\}in<answer\>, andeje\_\{j\}is extracted from the paragraph\-ID references insideuju\_\{j\}\. Segments withdj=YESd\_\{j\}=\\mathrm\{YES\}are added to the candidate clue poolCC\. Since long\-narrative questions often depend on indirect or distributed clues, theFinderis designed as a high\-recall selector: it should avoid discarding answer\-supporting evidence even when the segment does not directly state the final answer\.

RL Training for Finder\.We train theFinderwith reward\-guided reinforcement learning over structured outputs\. For a training segment, letzj∈\{0,1\}z\_\{j\}\\in\\\{0,1\\\}be the gold evidence label and letGj⊆IjG\_\{j\}\\subseteq I\_\{j\}be the annotated supporting paragraphs within the segment\. TheFinderreward combines decision and evidence behavior:

RF=λfmt​Rfmt\+λdec​Rdec\+λcite​Rcite\+λcomp​Rcomp\+λneg​Rneg\.R\_\{F\}=\\lambda\_\{\\mathrm\{fmt\}\}R\_\{\\mathrm\{fmt\}\}\+\\lambda\_\{\\mathrm\{dec\}\}R\_\{\\mathrm\{dec\}\}\+\\lambda\_\{\\mathrm\{cite\}\}R\_\{\\mathrm\{cite\}\}\+\\lambda\_\{\\mathrm\{comp\}\}R\_\{\\mathrm\{comp\}\}\+\\lambda\_\{\\mathrm\{neg\}\}R\_\{\\mathrm\{neg\}\}\.\(5\)Here eachλ\\lambdais a non\-negative weight controlling the importance of its corresponding reward term\.RfmtR\_\{\\mathrm\{fmt\}\}rewards valid structured output,RdecR\_\{\\mathrm\{dec\}\}rewards the correct YES/NO decision, andRciteR\_\{\\mathrm\{cite\}\}measures overlap between predicted paragraph IDseje\_\{j\}and gold paragraphsGjG\_\{j\}\.RcompR\_\{\\mathrm\{comp\}\}rewards compact gold references, andRnegR\_\{\\mathrm\{neg\}\}rewards conciseNOrationales without unsupported IDs\. A missed positive segment receives only the format reward\. This weighted reward design reflects the role of theFinderin the pipeline: preserving answer\-critical clues is more important than aggressively filtering the narrative, because theInterpretercan only reason from surviving evidence\. See Appendix A for reward details\.

### 3\.5Interpreter: Evidence\-Grounded Interpretation

Evidence\-Grounded Interpretation\.TheInterpreterturns the selected evidence packet into the final answer\. LetE=\{\(ℓt,p~t,u~t\)\}t=1TE=\\\{\(\\ell\_\{t\},\\tilde\{p\}\_\{t\},\\tilde\{u\}\_\{t\}\)\\\}\_\{t=1\}^\{T\}denote the ordered packet after packing, whereℓt\\ell\_\{t\}is the original paragraph index,p~t\\tilde\{p\}\_\{t\}is the retained evidence text,u~t\\tilde\{u\}\_\{t\}is theFinderrationale, andTTis the number of packed evidence units\. TheInterpreterfirst predicts a provisional answer and then, when triggered, self\-calibrates it using the same evidence:

z^I\\displaystyle\\hat\{z\}^\{I\}=Iθi,ans​\(E,q,𝒴\),\\displaystyle=I\_\{\\theta\_\{i\},\\mathrm\{ans\}\}\(E,q,\\mathcal\{Y\}\),\(6\)zI\\displaystyle z^\{I\}=\{Iθi,self​\-​cal​\(E,q,z^I\),τ⁡\(q\)=1,z^I,τ⁡\(q\)=0,\\displaystyle=\\begin\{cases\}I\_\{\\theta\_\{i\},\\mathrm\{self\\text\{\-\}cal\}\}\(E,q,\\hat\{z\}^\{I\}\),&\\tau\(q\)=1,\\\\ \\hat\{z\}^\{I\},&\\tau\(q\)=0,\\end\{cases\}wherezIz^\{I\}is parsed into\(r,y\)\(r,y\)and referenced paragraphs𝒞\\mathcal\{C\}\. Hererris a concise evidence\-grounded rationale written in<reason\>,y∈𝒴y\\in\\mathcal\{Y\}is the final answer written in<answer\>, and𝒞⊆\{ℓt\}t=1T\\mathcal\{C\}\\subseteq\\\{\\ell\_\{t\}\\\}\_\{t=1\}^\{T\}is extracted from paragraph\-ID references insiderr\. The triggerτ⁡\(q\)\\tau\(q\)is active for binary claim verification and for multiple\-choice questions whose wording suggests higher risk of polarity or reasoning errors, such as negation, exception, causal, or inferential forms\. The self\-verifier is internal to theInterpreter; it re\-checks the provisional answer against the same evidence packet, uses the same compact model, and does not introduce a third agent\. The model is therefore not asked to freely summarize the whole narrative\. It must connect the selected clues, choose an answer from the allowed answer space, and make the rationale traceable to paragraph IDs\.

RL Training for Interpreter\.We train theInterpreterwith the same GRPO principle but a different target\. Given gold answery⋆y^\{\\star\}and, when available, the supplied evidence paragraph setHH, theInterpreteruses

RI=λfmt​Rfmt\+λans​Rans\+λcite​Rcite\+λground​Rground−λhall​Rhall\.R\_\{I\}=\\lambda\_\{\\mathrm\{fmt\}\}R\_\{\\mathrm\{fmt\}\}\+\\lambda\_\{\\mathrm\{ans\}\}R\_\{\\mathrm\{ans\}\}\+\\lambda\_\{\\mathrm\{cite\}\}R\_\{\\mathrm\{cite\}\}\+\\lambda\_\{\\mathrm\{ground\}\}R\_\{\\mathrm\{ground\}\}\-\\lambda\_\{\\mathrm\{hall\}\}R\_\{\\mathrm\{hall\}\}\.\(7\)HereRfmtR\_\{\\mathrm\{fmt\}\}rewards valid structured output,RansR\_\{\\mathrm\{ans\}\}rewards matching the gold answery⋆y^\{\\star\},RciteR\_\{\\mathrm\{cite\}\}rewards paragraph IDs that point toHHand penalizes invalid IDs, andRgroundR\_\{\\mathrm\{ground\}\}rewards concise rationales with concrete grounding signals\. The hallucination penaltyRhallR\_\{\\mathrm\{hall\}\}discourages unsupported uncertainty or invalid paragraph references\. This makes theInterpreterconservative with evidence: it is rewarded for correctness and traceability\.

## 4Experiments

### 4\.1Experimental Setup

Datasets\.We evaluate on four long\-context narrative settings using the same test instances for all methods: DetectiveQA\[[27](https://arxiv.org/html/2608.25531#bib.bib12)\]for sparse detective\-plot clues,∞\\inftyBench\[[30](https://arxiv.org/html/2608.25531#bib.bib14)\]and LongBench v2\[[5](https://arxiv.org/html/2608.25531#bib.bib15)\]for broader long\-context reasoning, and NoCha\[[9](https://arxiv.org/html/2608.25531#bib.bib13)\]for novel\-length claim verification\.

Baselines\.We compare with two groups\. End\-to\-end readers use the available context directly with naive head truncation, including the44B backbone and larger local models: Qwen3\-8B\[[28](https://arxiv.org/html/2608.25531#bib.bib23)\], Ministral\-3\-14B\[[14](https://arxiv.org/html/2608.25531#bib.bib24)\], GPT\-OSS\-20B\[[15](https://arxiv.org/html/2608.25531#bib.bib25)\], Qwen3\-30B\-A3B\[[28](https://arxiv.org/html/2608.25531#bib.bib23)\], and Gemma\-4\-31B\-it\[[7](https://arxiv.org/html/2608.25531#bib.bib28)\]\. We also report higher\-cost API readers, Claude Haiku 4\.5\[[2](https://arxiv.org/html/2608.25531#bib.bib27)\]and GPT\-5 nano\[[16](https://arxiv.org/html/2608.25531#bib.bib26)\], as large\-context references\. Agentic baselines include ReAct\[[29](https://arxiv.org/html/2608.25531#bib.bib16)\], IRCoT\[[22](https://arxiv.org/html/2608.25531#bib.bib6)\], Self\-Ask\[[18](https://arxiv.org/html/2608.25531#bib.bib7)\], Chain\-of\-Agents\[[31](https://arxiv.org/html/2608.25531#bib.bib8)\], and RAG\-DDR\[[12](https://arxiv.org/html/2608.25531#bib.bib9)\], all using BGE\-M3\[[6](https://arxiv.org/html/2608.25531#bib.bib31)\]\.

Metrics\.We report final answer accuracy, normalizing multiple\-choice outputs to option labels and verification outputs to binary labels\.

Implementation Details\.To control local\-model comparisons, all local end\-to\-end readers use the same3232K setting: maximum context length32,76832\{,\}768, tokenizer\-exact input budget30,59230\{,\}592, and128128output tokens\. For higher\-cost API LLMs, we report a separate large\-context setting with a128128K budget, using126,976126\{,\}976input tokens and the same128128\-token output cap\. ClueWeaver, the agentic baselines, and the44B end\-to\-end reader use Qwen3\-4B\-Instruct\[[28](https://arxiv.org/html/2608.25531#bib.bib23)\]; larger local end\-to\-end readers use their own weights\. Retrieval\-based methods share BGE\-M3\[[6](https://arxiv.org/html/2608.25531#bib.bib31)\]and the same answer parser; baseline, training, and implementation details are deferred to Appendix C\.

Table 1:Main results on four long\-context narrative benchmarks \(final answer accuracy, %\)\. Among local methods, best per column is inboldand second best isunderlined; best API result per column is marked withwavy underlines\.
### 4\.2Main Results

Table[1](https://arxiv.org/html/2608.25531#S4.T1)reports final answer accuracy on the four benchmarks\. ClueWeaver achieves the best local overall accuracy \(59\.0%59\.0\\%\), leads all local methods on every dataset, and improves over the strongest local baseline \(IRCoT,52\.6%52\.6\\%\) by\+6\.4\+6\.4points overall\. With the same Qwen3\-4B backbone, the direct end\-to\-end reader reaches only44\.5%44\.5\\%because many narratives still require truncation; ClueWeaver improves it by\+14\.5\+14\.5points\. The gain is not a scale effect: the best end\-to\-end reader up to3131B \(Qwen3\-30B\-A3B,50\.3%50\.3\\%\) remains8\.78\.7points behind ClueWeaver’s44B backbone\. These results show that retrieval\-aware evidence selection, rather than context length or model size alone, is central to compact local long\-narrative QA\. Higher\-cost API LLMs are reported as large\-context references; ClueWeaver trails the best API overall result by4\.94\.9points and surpasses it on LongBench v2 by11\.511\.5points\.

### 4\.3Ablation Study

We ablate ClueWeaver on DetectiveQA, whose sparse, distributed clues make the contribution of each part most visible\.

Component ablation\.Table[2](https://arxiv.org/html/2608.25531#S4.T2)\(a\) removes inference components from the full model\. DisablingInterpreterself​\-​cal\{\}\_\{\\mathrm\{self\\text\{\-\}cal\}\}lowers accuracy by4\.84\.8to51\.0%51\.0\\%, showing that the internal second pass helps correct fragile decisions over the same evidence\. Removing the Finder—dumping all retrieved passages to the Interpreter instead of selecting answer\-critical evidence—drops accuracy by5\.85\.8points to50\.0%50\.0\\%\. Further removing both agents, leaving a bare end\-to\-end reader that must truncate the narrative to the backbone window, falls to36\.5%36\.5\\%\. Thus, selected evidence and self\-calibration both contribute to reliable long\-narrative reading\.

Table 2:Ablation on DetectiveQA\. \(a\) pipeline components and \(b\) RL training are removed from the full model \(Δ\\Delta: accuracy change vs\. the full model\)\.\(a\) Component ablation

\(b\) Training ablation

Figure 3:Accuracy–latency trade\-off using Table[1](https://arxiv.org/html/2608.25531#S4.T1)accuracies and serial single\-GPU stage timings\. Most additional latency comes from the Finder stage\.Training ablation\.Table[2](https://arxiv.org/html/2608.25531#S4.T2)\(b\) removes reward\-guided RL from the full model with the pipeline fixed\. Removing Finder RL \(untrained Finder, RL Interpreter\) is the most damaging, dropping accuracy by6\.86\.8points to49\.0%49\.0\\%—*below*the50\.0%50\.0\\%obtained with no Finder at all \(Table[2](https://arxiv.org/html/2608.25531#S4.T2)a\): an untrained Finder discards useful evidence, so it is RL training that turns the Finder into a net gain\. Removing Interpreter RL costs1\.01\.0, and removing both returns to the untrained pipeline at50\.0%50\.0\\%\. Finder training contributes the larger share by retaining answer\-critical clues, while Interpreter training converts the selected evidence into correct answers\.

Table 3:Cost profile for local and API readers\.
### 4\.4Analysis

Efficiency\.When served on a single GPU with Qwen3\-4B, ClueWeaver answers in8\.68\.6–9\.89\.8s per question, compared with2\.82\.8s for direct reading; most extra latency comes from Finder calls\. The ClueWeaver end\-to\-end latency remains practical for local deployment, while bringing a\+14\.5\+14\.5\-point gain and evidence\-level inspection\.

Figure 4:Error transitions between the direct reader and ClueWeaver across the four benchmarks\. ClueWeaver yields a clear net correction gain\.Table 4:Case study on DetectiveQA\. ClueWeaver exposes the evidence\-selection and interpretation process, with self\-calibration used as a final consistency check\.Cost\.Table[3](https://arxiv.org/html/2608.25531#S4.T3)compares local GPU requirements with API token fees\. Strong API readers can be competitive, but long\-context calls are costly\. ClueWeaver reaches API\-level accuracy with a compact local model and can be deployed on commercial\-grade GPUs, making the cost practical for sustained use\. Its deployment footprint is also far below that of 30B\-scale local readers\.

Error analysis\.Figure[4](https://arxiv.org/html/2608.25531#S4.F4)compares direct\-reader and ClueWeaver correctness\. ClueWeaver recovers8484of172172direct\-reader errors, yielding a consistently positive net gain across the benchmarks\. Some difficult cases remain, especially when the answer depends on distant multi\-hop clues or passages with weak surface overlap, which points to stronger distant\-clue retrieval as a future direction\.

### 4\.5Case Study

Table[4](https://arxiv.org/html/2608.25531#S4.T4)contrasts direct reading with the full ClueWeaver pipeline\. The end\-to\-end reader selects a surface\-level distractor, whereas ClueWeaver first finds the relevant paragraph, then grounds the answer in that evidence\. The self\-calibration step appears as a final consistency check inside theInterpreter, not as a separate agent\. Additional qualitative cases are shown in Appendix D\.

## 5Conclusion

This paper presented ClueWeaver, a reward\-guided dual\-agent framework for long narrative QA with compact local models\. The framework separates evidence selection from interpretation: Finder retains sparse clue passages, while Interpreter connects them into grounded answers and applies self\-calibration for high\-risk questions\. Reward\-guided training aligns both stages with the end task, encouraging clue preservation, faithful paragraph references, and concise explanations\. Experiments show that ClueWeaver improves local end\-to\-end readers, outperforms much larger local models, and approaches stronger API readers while remaining deployable on commercial GPUs\. These results suggest that compact models can support literary and humanities analysis through explicit evidence paths\. Future work will focus on harder distant clues and more robust multi\-hop evidence integration\.

#### Acknowledgements

This research was supported by the National Key R&D Program of China \(No\. 2023YFC3303800\)\. We also acknowledge WisPaper and QiewenPaper for providing Academic Agent support and GPU computational resources throughout the study; seehttps://wispaper\.ai\.

## References

- \[1\]\(2025\)Literaturwissenschaftlich arbeiten mit großen sprachmodellen? zwei experimentreihen zur textgenerierung mit künstlicher intelligenz\.Zeitschrift für Literaturwissenschaft und Linguistik55,pp\. 449–474\.External Links:[Document](https://dx.doi.org/10.1007/s41244-025-00383-4)Cited by:[§1](https://arxiv.org/html/2608.25531#S1.p1.1)\.
- \[2\]Anthropic\(2025\)Introducing Claude Haiku 4\.5\.Note:Anthropic announcementCited by:[§4\.1](https://arxiv.org/html/2608.25531#S4.SS1.p2.1),[Table 1](https://arxiv.org/html/2608.25531#S4.T1.7.10.1)\.
- \[3\]A\. Asai, Z\. Wu, Y\. Wang, A\. Sil, and H\. Hajishirzi\(2024\)Self\-RAG: learning to retrieve, generate, and critique through self\-reflection\.InProceedings of the International Conference on Learning Representations,Cited by:[§2\.2](https://arxiv.org/html/2608.25531#S2.SS2.p1.1)\.
- \[4\]Y\. Bai, X\. Lv, J\. Zhang, H\. Lyu, J\. Tang, Z\. Huang, Z\. Du, X\. Liu, A\. Zeng, L\. Hou, Y\. Dong, J\. Tang, and J\. Li\(2024\)LongBench: a bilingual, multitask benchmark for long context understanding\.InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),Bangkok, Thailand,pp\. 3119–3137\.External Links:[Document](https://dx.doi.org/10.18653/v1/2024.acl-long.172)Cited by:[§1](https://arxiv.org/html/2608.25531#S1.p1.1),[§1](https://arxiv.org/html/2608.25531#S1.p4.1),[§2\.1](https://arxiv.org/html/2608.25531#S2.SS1.p1.1)\.
- \[5\]Y\. Bai, S\. Tu, J\. Zhang, H\. Peng, X\. Wang, X\. Lv, S\. Cao, J\. Xu, L\. Hou, Y\. Dong, J\. Tang, and J\. Li\(2025\)LongBench v2: towards deeper understanding and reasoning on realistic long\-context multitasks\.InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),Vienna, Austria,pp\. 3639–3664\.External Links:[Document](https://dx.doi.org/10.18653/v1/2025.acl-long.183)Cited by:[§1](https://arxiv.org/html/2608.25531#S1.p1.1),[§1](https://arxiv.org/html/2608.25531#S1.p3.1),[§2\.1](https://arxiv.org/html/2608.25531#S2.SS1.p1.1),[§4\.1](https://arxiv.org/html/2608.25531#S4.SS1.p1.1)\.
- \[6\]J\. Chen, S\. Xiao, P\. Zhang, K\. Luo, D\. Lian, and Z\. Liu\(2024\)M3\-embedding: multi\-linguality, multi\-functionality, multi\-granularity text embeddings through self\-knowledge distillation\.InFindings of the Association for Computational Linguistics: ACL 2024,Bangkok, Thailand,pp\. 2318–2335\.External Links:[Document](https://dx.doi.org/10.18653/v1/2024.findings-acl.137)Cited by:[§4\.1](https://arxiv.org/html/2608.25531#S4.SS1.p2.1),[§4\.1](https://arxiv.org/html/2608.25531#S4.SS1.p4.1)\.
- \[7\]Google DeepMind\(2026\)Gemma 4 31B IT model card\.Note:Hugging Face model cardCited by:[§4\.1](https://arxiv.org/html/2608.25531#S4.SS1.p2.1),[Table 1](https://arxiv.org/html/2608.25531#S4.T1.7.8.1),[Table 3](https://arxiv.org/html/2608.25531#S4.T3.1.8.1.1.1)\.
- \[8\]Z\. Jiang, F\. Xu, L\. Gao, Z\. Sun, Q\. Liu, J\. Dwivedi\-Yu, Y\. Yang, J\. Callan, and G\. Neubig\(2023\)Active retrieval augmented generation\.InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing,Singapore,pp\. 7969–7992\.External Links:[Document](https://dx.doi.org/10.18653/v1/2023.emnlp-main.495)Cited by:[§2\.2](https://arxiv.org/html/2608.25531#S2.SS2.p1.1)\.
- \[9\]M\. Karpinska, K\. Thai, K\. Lo, T\. Goyal, and M\. Iyyer\(2024\)One thousand and one pairs: a “novel” challenge for long\-context language models\.InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing,Miami, Florida, USA,pp\. 17048–17085\.External Links:[Document](https://dx.doi.org/10.18653/v1/2024.emnlp-main.948)Cited by:[§1](https://arxiv.org/html/2608.25531#S1.p1.1),[§2\.1](https://arxiv.org/html/2608.25531#S2.SS1.p1.1),[§4\.1](https://arxiv.org/html/2608.25531#S4.SS1.p1.1)\.
- \[10\]T\. Kočiský, J\. Schwarz, P\. Blunsom, C\. Dyer, K\. M\. Hermann, G\. Melis, and E\. Grefenstette\(2018\)The NarrativeQA reading comprehension challenge\.Transactions of the Association for Computational Linguistics6,pp\. 317–328\.External Links:[Document](https://dx.doi.org/10.1162/tacl%5Fa%5F00023)Cited by:[§1](https://arxiv.org/html/2608.25531#S1.p1.1),[§1](https://arxiv.org/html/2608.25531#S1.p4.1),[§2\.1](https://arxiv.org/html/2608.25531#S2.SS1.p1.1)\.
- \[11\]P\. Lewis, E\. Perez, A\. Piktus, F\. Petroni, V\. Karpukhin, N\. Goyal, H\. Küttler, M\. Lewis, W\. Yih, T\. Rocktäschel, S\. Riedel, and D\. Kiela\(2020\)Retrieval\-augmented generation for knowledge\-intensive NLP tasks\.InAdvances in Neural Information Processing Systems,Vol\.33,pp\. 9459–9474\.Cited by:[§1](https://arxiv.org/html/2608.25531#S1.p3.1),[§1](https://arxiv.org/html/2608.25531#S1.p4.1),[§2\.2](https://arxiv.org/html/2608.25531#S2.SS2.p1.1)\.
- \[12\]X\. Li, S\. Mei, Z\. Liu, Y\. Yan, S\. Wang, S\. Yu, Z\. Zeng, H\. Chen, G\. Yu, Z\. Liu, M\. Sun, and C\. Xiong\(2025\)RAG\-DDR: optimizing retrieval\-augmented generation using differentiable data rewards\.InProceedings of the International Conference on Learning Representations,Cited by:[§0\.C\.1](https://arxiv.org/html/2608.25531#Pt0.A3.SS1.p1.1),[§1](https://arxiv.org/html/2608.25531#S1.p4.1),[§2\.2](https://arxiv.org/html/2608.25531#S2.SS2.p1.1),[§4\.1](https://arxiv.org/html/2608.25531#S4.SS1.p2.1),[Table 1](https://arxiv.org/html/2608.25531#S4.T1.7.17.1)\.
- \[13\]N\. F\. Liu, K\. Lin, J\. Hewitt, A\. Paranjape, M\. Bevilacqua, F\. Petroni, and P\. Liang\(2024\)Lost in the middle: how language models use long contexts\.Transactions of the Association for Computational Linguistics12,pp\. 157–173\.External Links:[Document](https://dx.doi.org/10.1162/tacl%5Fa%5F00638)Cited by:[§1](https://arxiv.org/html/2608.25531#S1.p1.1),[§1](https://arxiv.org/html/2608.25531#S1.p3.1),[§1](https://arxiv.org/html/2608.25531#S1.p4.1),[§2\.1](https://arxiv.org/html/2608.25531#S2.SS1.p1.1)\.
- \[14\]Mistral AI\(2025\)Introducing Mistral 3\.Note:Mistral AI announcement and model cardsCited by:[§4\.1](https://arxiv.org/html/2608.25531#S4.SS1.p2.1),[Table 1](https://arxiv.org/html/2608.25531#S4.T1.7.5.1),[Table 3](https://arxiv.org/html/2608.25531#S4.T3.1.5.1.1.1)\.
- \[15\]OpenAI\(2025\)gpt\-oss\-120b and gpt\-oss\-20b model card\.arXiv preprint arXiv:2508\.10925\.Cited by:[§4\.1](https://arxiv.org/html/2608.25531#S4.SS1.p2.1),[Table 1](https://arxiv.org/html/2608.25531#S4.T1.7.6.1),[Table 3](https://arxiv.org/html/2608.25531#S4.T3.1.6.1.1.1)\.
- \[16\]OpenAI\(2025\)Introducing GPT\-5 for developers\.Note:OpenAI announcementCited by:[§4\.1](https://arxiv.org/html/2608.25531#S4.SS1.p2.1),[Table 1](https://arxiv.org/html/2608.25531#S4.T1.7.11.1)\.
- \[17\]R\. Y\. Pang, A\. Parrish, N\. Joshi, N\. Nangia, J\. Phang, A\. Chen, V\. Padmakumar, J\. Ma, J\. Thompson, H\. He, and S\. R\. Bowman\(2022\)QuALITY: question answering with long input texts, yes\!\.InProceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies,Seattle, United States,pp\. 5336–5358\.External Links:[Document](https://dx.doi.org/10.18653/v1/2022.naacl-main.391)Cited by:[§1](https://arxiv.org/html/2608.25531#S1.p1.1),[§1](https://arxiv.org/html/2608.25531#S1.p4.1),[§2\.1](https://arxiv.org/html/2608.25531#S2.SS1.p1.1)\.
- \[18\]O\. Press, M\. Zhang, S\. Min, L\. Schmidt, N\. Smith, and M\. Lewis\(2023\)Measuring and narrowing the compositionality gap in language models\.InFindings of the Association for Computational Linguistics: EMNLP 2023,External Links:[Document](https://dx.doi.org/10.18653/v1/2023.findings-emnlp.378)Cited by:[§0\.C\.1](https://arxiv.org/html/2608.25531#Pt0.A3.SS1.p1.1),[§1](https://arxiv.org/html/2608.25531#S1.p4.1),[§2\.2](https://arxiv.org/html/2608.25531#S2.SS2.p1.1),[§4\.1](https://arxiv.org/html/2608.25531#S4.SS1.p2.1),[Table 1](https://arxiv.org/html/2608.25531#S4.T1.7.15.1)\.
- \[19\]R\. Rafailov, A\. Sharma, E\. Mitchell, C\. D\. Manning, S\. Ermon, and C\. Finn\(2023\)Direct preference optimization: your language model is secretly a reward model\.InAdvances in Neural Information Processing Systems,Vol\.36\.Cited by:[§2\.3](https://arxiv.org/html/2608.25531#S2.SS3.p1.1)\.
- \[20\]J\. Schulman, F\. Wolski, P\. Dhariwal, A\. Radford, and O\. Klimov\(2017\)Proximal policy optimization algorithms\.arXiv preprint arXiv:1707\.06347\.Cited by:[§2\.3](https://arxiv.org/html/2608.25531#S2.SS3.p1.1)\.
- \[21\]Z\. Shao, P\. Wang, Q\. Zhu, R\. Xu, J\. Song, X\. Bi, H\. Zhang, M\. Zhang, Y\. K\. Li, Y\. Wu, and D\. Guo\(2024\)DeepSeekMath: pushing the limits of mathematical reasoning in open language models\.arXiv preprint arXiv:2402\.03300\.Cited by:[§1](https://arxiv.org/html/2608.25531#S1.p5.1),[§2\.3](https://arxiv.org/html/2608.25531#S2.SS3.p1.1),[§3\.3](https://arxiv.org/html/2608.25531#S3.SS3.p1.2)\.
- \[22\]H\. Trivedi, N\. Balasubramanian, T\. Khot, and A\. Sabharwal\(2023\)Interleaving retrieval with chain\-of\-thought reasoning for knowledge\-intensive multi\-step questions\.InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),Toronto, Canada,pp\. 10014–10037\.External Links:[Document](https://dx.doi.org/10.18653/v1/2023.acl-long.557)Cited by:[§0\.C\.1](https://arxiv.org/html/2608.25531#Pt0.A3.SS1.p1.1),[§1](https://arxiv.org/html/2608.25531#S1.p4.1),[§2\.2](https://arxiv.org/html/2608.25531#S2.SS2.p1.1),[§4\.1](https://arxiv.org/html/2608.25531#S4.SS1.p2.1),[Table 1](https://arxiv.org/html/2608.25531#S4.T1.7.14.1)\.
- \[23\]C\. Wang, R\. Ning, B\. Pan, T\. Wu, Q\. Guo, C\. Deng, G\. Bao, X\. Hu, Z\. Zhang, Q\. Wang, and Y\. Zhang\(2025\)NovelQA: benchmarking question answering on documents exceeding 200k tokens\.InProceedings of the International Conference on Learning Representations,Cited by:[§1](https://arxiv.org/html/2608.25531#S1.p1.1),[§2\.1](https://arxiv.org/html/2608.25531#S2.SS1.p1.1)\.
- \[24\]Y\. Wang, C\. Li, G\. Chen, J\. Liang, and T\. Wang\(2026\)Reasoning or retrieval? a study of answer attribution on large reasoning models\.InProceedings of the International Conference on Learning Representations,Cited by:[§2\.3](https://arxiv.org/html/2608.25531#S2.SS3.p1.1)\.
- \[25\]J\. Wei, M\. Bosma, V\. Y\. Zhao, K\. Guu, A\. W\. Yu, B\. Lester, N\. Du, A\. M\. Dai, and Q\. V\. Le\(2022\)Finetuned language models are zero\-shot learners\.InProceedings of the International Conference on Learning Representations,Cited by:[§2\.3](https://arxiv.org/html/2608.25531#S2.SS3.p1.1)\.
- \[26\]D\. G\. Widder, M\. Whittaker, and S\. M\. West\(2024\)Why ‘open’ AI systems are actually closed, and why this matters\.Nature635,pp\. 827–833\.External Links:[Document](https://dx.doi.org/10.1038/s41586-024-08141-1)Cited by:[§1](https://arxiv.org/html/2608.25531#S1.p1.1)\.
- \[27\]Z\. Xu, J\. Ye, X\. Liu, X\. Liu, T\. Sun, Z\. Liu, Q\. Guo, L\. Li, Q\. Liu, X\. Huang, and X\. Qiu\(2025\)DetectiveQA: evaluating long\-context reasoning on detective novels\.InICLR 2025 Workshop on Reasoning and Planning for Large Language Models,Cited by:[§1](https://arxiv.org/html/2608.25531#S1.p1.1),[§2\.1](https://arxiv.org/html/2608.25531#S2.SS1.p1.1),[§4\.1](https://arxiv.org/html/2608.25531#S4.SS1.p1.1)\.
- \[28\]A\. Yang, A\. Li, B\. Yang, B\. Zhang, B\. Hui, B\. Zheng, B\. Yu, C\. Gao, C\. Huang, C\. Lv,et al\.\(2025\)Qwen3 technical report\.arXiv preprint arXiv:2505\.09388\.Cited by:[§1](https://arxiv.org/html/2608.25531#S1.p1.1),[§4\.1](https://arxiv.org/html/2608.25531#S4.SS1.p2.1),[§4\.1](https://arxiv.org/html/2608.25531#S4.SS1.p4.1),[Table 1](https://arxiv.org/html/2608.25531#S4.T1.7.3.1),[Table 1](https://arxiv.org/html/2608.25531#S4.T1.7.4.1),[Table 1](https://arxiv.org/html/2608.25531#S4.T1.7.7.1),[Table 3](https://arxiv.org/html/2608.25531#S4.T3.1.3.1.1.1),[Table 3](https://arxiv.org/html/2608.25531#S4.T3.1.4.1.1.1),[Table 3](https://arxiv.org/html/2608.25531#S4.T3.1.7.1.1.1)\.
- \[29\]S\. Yao, J\. Zhao, D\. Yu, N\. Du, I\. Shafran, K\. Narasimhan, and Y\. Cao\(2023\)ReAct: synergizing reasoning and acting in language models\.InProceedings of the International Conference on Learning Representations,Cited by:[§0\.C\.1](https://arxiv.org/html/2608.25531#Pt0.A3.SS1.p1.1),[§2\.2](https://arxiv.org/html/2608.25531#S2.SS2.p1.1),[§4\.1](https://arxiv.org/html/2608.25531#S4.SS1.p2.1),[Table 1](https://arxiv.org/html/2608.25531#S4.T1.7.13.1)\.
- \[30\]X\. Zhang, Y\. Chen, S\. Hu, Z\. Xu, J\. Chen, M\. K\. Hao, X\. Han, Z\. L\. Thai, S\. Wang, Z\. Liu, and M\. Sun\(2024\)∞\\inftyBench: Extending long context evaluation beyond 100k tokens\.InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),Bangkok, Thailand,pp\. 15262–15277\.External Links:[Document](https://dx.doi.org/10.18653/v1/2024.acl-long.814)Cited by:[§1](https://arxiv.org/html/2608.25531#S1.p1.1),[§2\.1](https://arxiv.org/html/2608.25531#S2.SS1.p1.1),[§4\.1](https://arxiv.org/html/2608.25531#S4.SS1.p1.1)\.
- \[31\]Y\. Zhang, R\. Sun, Y\. Chen, T\. Pfister, R\. Zhang, and S\. Ö\. Arık\(2024\)Chain of agents: large language models collaborating on long\-context tasks\.InAdvances in Neural Information Processing Systems,Vol\.37\.External Links:[Document](https://dx.doi.org/10.52202/079017-4202)Cited by:[§0\.C\.1](https://arxiv.org/html/2608.25531#Pt0.A3.SS1.p1.1),[§1](https://arxiv.org/html/2608.25531#S1.p4.1),[§2\.2](https://arxiv.org/html/2608.25531#S2.SS2.p1.1),[§4\.1](https://arxiv.org/html/2608.25531#S4.SS1.p2.1),[Table 1](https://arxiv.org/html/2608.25531#S4.T1.7.16.1)\.

## Appendix 0\.AReward Design Details

### 0\.A\.1Reward Components

We use rule\-based rewards rather than a learned reward model, because the two agents have explicit structured roles and the training signals can be defined from labels, paragraph indices, and output format\. The final trained agents use the balanced Finder reward and the test\-aligned Interpreter reward described below\. All rewards are computed after parsing the generated XML\-style output\. If the output cannot be parsed, the reward is set to zero\. Otherwise, format validity provides a small base reward, while task\-specific correctness and evidence behavior determine the remaining score\. The constants below are the unnormalized rewards used before GRPO advantage normalization\.

##### Finder reward\.

For a segmentsjs\_\{j\}, letzj∈\{0,1\}z\_\{j\}\\in\\\{0,1\\\}indicate whether it contains supporting evidence, letGjG\_\{j\}be the annotated gold paragraphs inside the segment, and letPjP\_\{j\}be the paragraph IDs referenced in the Finder’s<reason\>field\. The final Finder reward is balanced: it makes positive evidence retention more valuable than a short negative response, while still giving enough reward to correctNOdecisions to prevent an all\-YESpolicy\.

Table 5:Finder reward components\.This design intentionally favors recall without collapsing into a trivial all\-YESpolicy\. Positive segments can receive the highest score only when the decision and paragraph IDs are both correct; negative segments still receive meaningful reward when they are rejected with a concise explanation\.

##### Interpreter reward\.

The Interpreter receives the packed evidence and predicts the final answer\. Its reward is correctness\-dominant, with all rationale\-quality bonuses gated by a correct answer\. This prevents the model from receiving high reward for fluent but wrong explanations\.

Table 6:Interpreter reward components\.The two rewards therefore optimize complementary abilities\. The Finder is pushed to preserve sparse, answer\-relevant clues with faithful paragraph\-ID references, whereas the Interpreter is pushed to convert the retained evidence into a correct, grounded, and concise answer\.

### 0\.A\.2GRPO Objective Details

For each agenta∈\{F,I\}a\\in\\\{F,I\\\}and training inputxax^\{a\}, GRPO samplesKKcomplete structured outputs from the previous policy and compares them within the same group\. Letoka=\(wk,1,…,wk,Tk\)o\_\{k\}^\{a\}=\(w\_\{k,1\},\\ldots,w\_\{k,T\_\{k\}\}\)be thekk\-th sampled token sequence\. Its sequence log\-probability under a policyπ\\piis

log⁡π⁡\(oka\|xa\)=∑t=1Tklog⁡π⁡\(wk,t\|xa,wk,<t\)\.\\log\\pi\(o\_\{k\}^\{a\}\|x^\{a\}\)=\\sum\_\{t=1\}^\{T\_\{k\}\}\\log\\pi\(w\_\{k,t\}\|x^\{a\},w\_\{k,<t\}\)\.\(8\)The sequence\-level policy ratio used in Eq\.[2](https://arxiv.org/html/2608.25531#S3.E2)is therefore

ρka=exp⁡\(log⁡πθaa​\(oka\|xa\)−log⁡πolda​\(oka\|xa\)\)\.\\rho\_\{k\}^\{a\}=\\exp\\\!\\left\(\\log\\pi\_\{\\theta\_\{a\}\}^\{a\}\(o\_\{k\}^\{a\}\|x^\{a\}\)\-\\log\\pi\_\{\\mathrm\{old\}\}^\{a\}\(o\_\{k\}^\{a\}\|x^\{a\}\)\\right\)\.\(9\)After parsing the full XML\-style response, we compute the agent\-specific rewardRa​\(oka,xa\)R\_\{a\}\(o\_\{k\}^\{a\};x^\{a\}\)and normalize it within the sampled group:

μa=1K​∑l=1KRa​\(ola,xa\),σa=1K​∑l=1K\(Ra​\(ola,xa\)−μa\)2,\\mu\_\{a\}=\\frac\{1\}\{K\}\\sum\_\{l=1\}^\{K\}R\_\{a\}\(o\_\{l\}^\{a\};x^\{a\}\),\\quad\\sigma\_\{a\}=\\sqrt\{\\frac\{1\}\{K\}\\sum\_\{l=1\}^\{K\}\(R\_\{a\}\(o\_\{l\}^\{a\};x^\{a\}\)\-\\mu\_\{a\}\)^\{2\}\},\(10\)Aka=Ra​\(oka,xa\)−μaσa\+ϵ\.A\_\{k\}^\{a\}=\\frac\{R\_\{a\}\(o\_\{k\}^\{a\};x^\{a\}\)\-\\mu\_\{a\}\}\{\\sigma\_\{a\}\+\\epsilon\}\.\(11\)The KL term is evaluated on the same sampled sequence against the reference model\. Withrka=πrefa​\(oka\|xa\)/πθaa​\(oka\|xa\)r\_\{k\}^\{a\}=\\pi\_\{\\mathrm\{ref\}\}^\{a\}\(o\_\{k\}^\{a\}\|x^\{a\}\)/\\pi\_\{\\theta\_\{a\}\}^\{a\}\(o\_\{k\}^\{a\}\|x^\{a\}\), we use

Dka=rka−log⁡rka−1\.D\_\{k\}^\{a\}=r\_\{k\}^\{a\}\-\\log r\_\{k\}^\{a\}\-1\.\(12\)This value\-free formulation is used for both agents, so no separate critic or value model is trained\.

## Appendix 0\.BTraining Details

### 0\.B\.1Training Data

All training data are drawn from the benchmarks’ training splits, disjoint from the test set\. As the source narratives are long, often exceeding100100K tokens, we segment each document at the paragraph level and supervise both agents on segments rather than whole texts \(Fig\.[5](https://arxiv.org/html/2608.25531#Pt0.A2.F5)\)\. The Finder learns per\-segment keep/drop decisions, balanced50/5050/50with hard negatives, from a question\-answering and a claim\-verification split of detective novels\. The Interpreter learns from1,0001\{,\}000examples, each pairing a question with an evidence packet assembled from selected segments; the mixture follows the test distribution, scarce real NoCha cases are augmented with synthetic claim verification, and27\.5%27\.5\\%are items the base model fails, concentrating the reward on hard cases\. A t\-SNE projection of BGE\-M3 question embeddings shows both pools are dominated by the detective domain; LongBench v2 is the most under\-represented benchmark for the Interpreter, consistent with the smaller end\-to\-end gains seen there\.

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\.
### 0\.B\.2Training Setup

All experiments use Qwen3\-4B\-Instruct as the base model for both agents\. The Finder and Interpreter are trained separately with GRPO, using eight sampled responses per prompt to estimate group\-relative advantages\. We train full model weights in bfloat16 with cosine learning\-rate decay, a warmup ratio of0\.10\.1, temperature1\.01\.0, top\-p=0\.9p=0\.9, top\-k=20k=20, and GRPO KL coefficientβ=0\.04\\beta=0\.04\. We disable model\-internal thinking during both training and inference so that the emitted traces follow the required structured format\.

Table 7:GRPO training hyperparameters for each agent\.Training is performed on a single node with 8 x NVIDIA A100 GPUs\. During GRPO, vLLM is colocated with training workers to accelerate rollout generation\. At inference time, each active Qwen3\-4B model instance uses about 10–12 GB of GPU memory with bfloat16 weights\. The Finder and Interpreter can therefore be served sequentially on one GPU or concurrently on separate GPUs\. Dense retrieval uses BGE\-M3 on GPU, and all methods share the same answer normalization and output parser to avoid evaluation differences from formatting alone\. Local end\-to\-end readers are evaluated with a 32K context budget \(30,592 input tokens plus 128 output tokens\), while API end\-to\-end readers use a 128K context budget \(126,976 input tokens plus the same 128\-token output cap\)\.

## Appendix 0\.CImplementation Details

### 0\.C\.1Baseline Implementations

All baselines use the same normalized question format and answer parser as ClueWeaver\. Local agentic baselines use Qwen3\-4B\-Instruct as the backbone; larger end\-to\-end readers use their own model weights\. The direct reader receives the narrative under the configured context budget and answers in one call\. BM25, dense, and hybrid RAG retrieve top paragraphs and answer from the packed evidence\. Our ReAct baseline\[[29](https://arxiv.org/html/2608.25531#bib.bib16)\]is a ReAct\-style iterative RAG adaptation: in this closed\-document setting, the action is retrieval over the given narrative, the observation is the retrieved paragraph evidence, and the model alternates reasoning and retrieval before emitting the final answer\. IRCoT\[[22](https://arxiv.org/html/2608.25531#bib.bib6)\]interleaves retrieval with one generated reasoning sentence per step\. Self\-Ask\[[18](https://arxiv.org/html/2608.25531#bib.bib7)\]first generates follow\-up sub\-questions, retrieves evidence for them, and answers from the resulting trace\. Chain\-of\-Agents\[[31](https://arxiv.org/html/2608.25531#bib.bib8)\]splits the narrative into chunks, lets worker agents update a communication summary, and uses a manager agent for the final answer\. RAG\-DDR\[[12](https://arxiv.org/html/2608.25531#bib.bib9)\]retrieves candidate passages, applies a prompt\-only knowledge\-refinement YES/NO filter, and answers from the retained passages without DDR training\.

### 0\.C\.2Retrieval, Segmentation, and Evidence Packing

ClueWeaver uses retrieval as a front\-end for candidate construction, not as a replacement for narrative reading\. Given a questionqqand a long narrativeD=\{pi\}i=1mD=\\\{p\_\{i\}\\\}\_\{i=1\}^\{m\}, we first score paragraphs with dense retrieval and lexical matching\. Dense retrieval is implemented with BGE\-M3, while lexical matching is used only as a complementary signal for robust candidate coverage\. The top paragraphs are used as anchors for local windows, producing candidate segments that preserve paragraph IDs and nearby context\.

Each candidate segment is then judged by the Finder\. Unlike standard RAG, which passes top\-ranked chunks directly to the answer model, ClueWeaver asks the Finder to decide whether a segment contains useful clues and to cite the supporting paragraph IDs\. This step removes many retrieval\-only false positives and keeps evidence traceable at paragraph level\.

Table 8:Evidence construction stages in ClueWeaver\.The evidence packet is controlled by four implementation parameters\.NEN\_\{E\}is the maximum number of selected evidence segments\.PrP\_\{r\}andPwP\_\{w\}are the paragraph budgets for retrieval\-anchored and local\-window segments, respectively\.BcB\_\{c\}is the total character budget for the packed evidence\. For example, a focused configuration usesNE=5N\_\{E\}=5,Pr=2P\_\{r\}=2,Pw=5P\_\{w\}=5, andBc=13,000B\_\{c\}=13\{,\}000, whereas a broader configuration usesNE=10N\_\{E\}=10,Pr=4P\_\{r\}=4,Pw=6P\_\{w\}=6, andBc=15,000B\_\{c\}=15\{,\}000\. The former illustrates noise control, while the latter illustrates recall\-oriented packing for sparse and distributed clues\. Selected evidence is sorted by its original paragraph index before being passed to the Interpreter, so the final model receives clues in narrative order rather than retrieval\-score order\. For the main results, the task\-level budgets are DetectiveQA\(10,4,6,15000\)\(10,4,6,15000\),∞\\inftyBench\(7,3,6,14000\)\(7,3,6,14000\), LongBench v2\(8,4,6,15000\)\(8,4,6,15000\), and NoCha\(10,6,8,16000\)\(10,6,8,16000\)\.

### 0\.C\.3Traceability and Format Audit

We further audit the final ClueWeaver traces used in Table[1](https://arxiv.org/html/2608.25531#S4.T1)\. A citation is valid if the paragraph ID cited by the Interpreter appears in the Finder\-provided evidence packet\. Missing citations are not counted as invalid; the audit measures the faithfulness of explicit paragraph references\. Across 310 questions, 275 outputs contain paragraph citations\. Among citation\-bearing outputs, 685 of 690 row\-unique cited paragraph IDs are valid \(99\.3%\), and 270 of 275 outputs contain only valid citations \(98\.2%\)\. The structured output is also stable:<reason\>and<answer\>tags are present in 309 of 310 outputs \(99\.7%\), and the final parser extracts a legal answer label in 308 of 310 outputs \(99\.4%\)\.

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

## Appendix 0\.DAdditional Case Studies

Table[10](https://arxiv.org/html/2608.25531#Pt0.A4.T10)gives representative qualitative examples from DetectiveQA\. We include one fixed case, one regression, and one unresolved case to show where the proposed pipeline helps and where it still fails\.

Table 10:Additional qualitative cases\.
## Appendix 0\.EPrompt Templates and Structured Outputs

We use three prompt families:Finder,Interpreter, andInterpreterself​\-​cal\{\}\_\{\\mathrm\{self\\text\{\-\}cal\}\}\. In implementation, each family has a multiple\-choice instantiation and a binary claim\-verification instantiation withTRUE/FALSEanswers\. Self\-calibration is invoked only inside theInterpreter; there is no self\-calibration prompt for the Finder\.

### 0\.E\.1Prompt Templates

YouaretheevidenceFinderagentinalong\-narrativeQApipeline\.

YoureadONEsegmentfromalongstory\.Paragraphsarenumberedlike\[N\]\.

DecidewhetherthesegmentcontainsconcreteevidencethatshouldbeshowntoadownstreamInterpreter\.

SayYESonlywhenthesegmentcontainsaconcretefactthathelpsanswerthequestion,chooseorruleoutoneoption,orconfirmorrefuteaclaim\.Strongevidenceincludesarelevantaction,dialogueline,motive,relationship,causalexplanation,time/placeclue,object,identity,orexplicitcontradiction\.

SayNOwhen:

\-Thesegmentisscene\-setting,scenery,weather,ortransitionnarrative\.

\-Thesegmentmentionscharactersbutsaysnothingaboutwhatthequestionisasking\.

\-Theoverlapisonlyacommonword,optionword,orpassingnamewithnorelevantfact\.

\-Thesegmentmerelyraisessuspicionbutgivesnofactthatdistinguishesoptions\.

\-Youcannotnameaconcretecluefromthesegment\.

Bias:preserveanswer\-criticalevidence\.PreferNOforpurebackground,butchooseYESforanyconcretefactthatcouldhelpanswer,ruleoutanoption,confirm,orrefutetheclaim\.Judgethissegmentindependently;partialevidenceisstillevidence\.

Questionorclaim:\{question\}

Answerspace:

\(A\)\{opt\_a\}

\(B\)\{opt\_b\}

\(C\)\{opt\_c\}

\(D\)\{opt\_d\}

orTRUE/FALSEforclaimverification

Segment\(paragraphs\{start\_para\}\-\{end\_para\}\):

\{segment\_text\}

OUTPUTFORMAT–exactlytwoXMLfieldsandnothingelse:

<reason\>onesentencenamingtheconcreteclue,orsayingnoconcreteclueispresent;cite\[N\]whenapplicable</reason\>

<answer\>YES</answer\>or<answer\>NO</answer\>

YouarethefinalInterpreterinadual\-agentlong\-narrativeQApipeline\.

AnswerthequestionorverifytheclaimusingONLYthesuppliedevidence\.Evidenceparagraphsarenumberedlike\[N\]\.

Rules:

\-Firstidentifytheexactfactthequestionasksfor;donotdrifttoarelatedevent\.

\-Matchbymeaning,notwording\.

\-Checkquestionpolarity,includingfalse,except,ornottrue\.

\-Compareeveryoptionagainstdirectevidence\.

\-Missingevidenceforanoptiondoesnotproveitwrong;onlyexplicitcontradictionrulesitout\.

\-Prefertheoptionwiththestrongestpositivesupportintheevidence\.

\-Forclaimverification,checkeachessentialelementoftheclaimandanswerTRUEorFALSE\.

\-Citeparagraphnumberslike\[478\]foreverydecisivefact\.

\-Keep<reason\>under120words\.

Questionorclaim:\{question\}

Answerspace:

\(A\)\{opt\_a\}

\(B\)\{opt\_b\}

\(C\)\{opt\_c\}

\(D\)\{opt\_d\}

orTRUE/FALSEforclaimverification

Evidence\(keptsegments,inorder\):

\{evidence\_block\}

OUTPUT–onlythesetwoXMLfields,nothingelse:

<reason\>conciseevidence\-groundedrationalewith\[N\]citations,under120words</reason\>

<answer\>A</answer\>or<answer\>TRUE</answer\>

YouarethesameInterpreterperformingaself\-calibrationstep\.

Thepreviousanswermaybewrong\.Re\-checkthequestionorclaim,answerspace,

evidence,andpreviousanswerusingONLYthesuppliedevidence\.

Rules:

\-Ifthepreviousrationalesupportsoneoptionbuttheanswertagnamesanother,correcttheanswer\.

\-Matchtheexactquestionintent,includingwhy,false,except,nottrue,deduce,andinfer\.

\-Forclaimverification,donotaddrequirementsthatarenotstatedintheclaim\.

\-Preferconcreteevidencelinksoverbroadorisolatedwordoverlap\.

\-Ifanotheroptionisbettersupportedbytheevidence,changetheanswer\.

\-Keep<reason\>under100wordsandcitedecisiveparagraphnumbers\.

Questionorclaim:\{question\}

Answerspace:

\(A\)\{opt\_a\}

\(B\)\{opt\_b\}

\(C\)\{opt\_c\}

\(D\)\{opt\_d\}

orTRUE/FALSEforclaimverification

Evidence:

\{evidence\_block\}

Previousanswer:\{previous\_answer\}

Previousrationale:\{previous\_reason\}

OUTPUT–onlythesetwoXMLfields,nothingelse:

<reason\>self\-calibratedrationalewith\[N\]citations</reason\>

<answer\>A</answer\>or<answer\>TRUE</answer\>