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

Wenji Fu

cs.AISep 19, 2026

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

Authors: Wenji Fu

Organizations: Research Institute of Economics and Management Southwestern University of Finance and Economics

Abstract

Long-context models read a novel the way a person reads a printout: one token after another, in narrative order, with the whole history competing for a fixed budget of attention. A detective does not work that way. They sort what happened when, and they keep a map of who relates to whom, so a clue from chapter one can meet a question asked at the end of the book. We test whether a frozen knowledge graph can give a small local model that same freedom. Thirty detective novels and 234 multiple-choice questions are answered by one fixed qwen3.5:9b reader under nine conditions: five graph routes, a recent-window baseline, whole-book compression, ordinary vector retrieval, and a question-only control. The strongest graph route reaches 53.85% (126/234) against 46.15% for the recent window, 51.28% for compression, 51.71% for vector retrieval and 40.17% for question-only. On the subset that no model can answer without the book, the graph route reaches 42.86%. None of the fifteen graph-baseline contrasts survives Holm correction, so we present the result as exploratory evidence about a design. Two structural findings survive scrutiny better than the headline number: annotated evidence concentrates in the topological core of these graphs (2.35x enrichment, pooled), and the two graph-building pipelines differ so much in annotation coverage (16% versus 73% of clue paragraphs) that pooled accuracy alone would hide which bottleneck is being measured.

Figures & tables

Panel of Figure 1 — Figure 1: Provenance chain for one graph edge. A relation is admitted only if its evidence string occurs verbatim in a specific passage, which is itself an offset-bounded slice of the frozen source file whose SHA-256 is in the manifest. The audit walks the chain in both directions: from a scored answer back to the characters it quotes, and from any source character to the nodes and edges that claim it.

Figure 2: The build. Two passes over verbatim text separate plot-bearing spans from literary padding, then extract typed relations from the survivors. Gates are explicit, rejections are counted, and the complete novel is retained as a lossless sidecar.

Figure 3: The four route families behind the five graph conditions. Fusion and reranking select a dossier before the reader sees it; refereeing compares two independent readers and arbitrates only on disagreement; walking hands the choice of the next edge to the reader and validates every move against the adjacency list. No route has access to a baseline prediction, a gold answer or a gold paragraph.

Figure 4: What the protocol freezes, what no method may see, what must be reported apart, and what remains sealed. The pooled thirty-novel value is descriptive because the two cohorts come from different build pipelines; the confirmatory run follows a separate protocol whose answers have not been opened.

Condition	Correct	Accuracy	95% Wilson	Novel macro	Q0-hard	vs. Q0
G1 order-1	118/234	50.43	[44.1, 56.8]	51.03	52/140	+10.26
G2 order-2	112/234	47.86	[41.5, 54.2]	48.02	52/140	+7.69
G3 order-3	118/234	50.43	[44.1, 56.8]	50.71	55/140	+10.26
G4 graph majority	122/234	52.14	[45.8, 58.5]	52.81	53/140	+11.97
G5 tight expansion	126/234	53.85	[47.4, 60.1]	53.34	60/140	+13.68
B1 tail window	108/234	46.15	[39.9, 52.6]	45.27	47/140	+5.98

Table 1: Accuracy on the frozen thirty-novel evaluation (234 questions, one local qwen3.5:9b reader, reasoning disabled). Pooled values are descriptive because the two cohorts were produced by different build pipelines. Wilson intervals are per condition; the

\mathrm{Q0\text{-}hard}

column restricts to the 140 questions the question-only control answers incorrectly. Every value is regenerated from the archived per-question table.

Figure 5: Accuracy by condition. (a) All 234 questions with Wilson intervals and correct/total counts. (b) The same conditions split by build cohort. (c) The

\mathrm{Q0\text{-}hard}

subset, where no model can answer from the question alone. (d) Novel-macro accuracy against micro accuracy; the diagonal is the identity. Panel (b) shows the largest single source of variance in the study. It belongs to the graphs themselves.

Figure 6: Per-novel accuracy for every condition. (a) Rows are novels in cohort order, with the cohort boundary marked; columns are conditions. (b) Questions per novel. Cells are computed from the released per-question table, so the heatmap and the pooled table cannot disagree. The old20 block is visibly weaker for the graph routes, and its items are also unevenly distributed across novels.

Figure 7: Paired differences between the strongest graph route and each baseline. Points are paired accuracy differences in percentage points; bars are novel-clustered 95% bootstrap intervals; annotations give wins, losses, exact McNemar

p

and the Holm-adjusted value across the fifteen planned graph–baseline contrasts. Nothing survives correction, so the contrast with the recent window points at where to look next.

Figure 8: Difficulty census across the nine conditions. (a) How many conditions answer each question correctly. The spike at zero counts real failures. (b) Questions where a graph condition is the only correct answer, where a baseline is the only correct answer, and where nothing is correct.

Figure 9: Evidence audit. (a) Share of nodes inside the graph 2-core, split by whether the node overlaps an annotated paragraph; multipliers give the enrichment ratio. (b) Share of annotated clue and answer positions that a graph represents at all, by cohort. (c) Graph size against annotation mapping rate. The larger graphs are not the better-mapped ones. Panel (a) is an association between annotation overlap and topology; it is not evidence about model attention or about answer accuracy.

Figure 10: The frozen graphs as inputs. (a) Nodes against edges with the identity line; points below the diagonal are the sparse legacy graphs. (b) Edge density per novel, with the 0.50 builder gate marked; novels are sorted by density and colours give the cohort. (c) Isolated nodes in the same order. Two cohorts built by different pipelines differ in density by a factor of four, which is a property of the builders and a candidate explanation for the cohort gap in Figure 9 (b).

Figure 11: Option-position effects. (a) Oracle accuracy by gold option letter for an options-first oracle and a position-neutral one. (b) Which options the graph reader chose, against where the gold answers actually sit. Any single oracle number is therefore a weak ceiling: it can be inflated by the option layout, and the position-neutral variant is the one worth quoting.

Condition	old20 (164)	new10 (70)	pooled (234)	preserved on Q0-correct
G1 order-1	50.0	51.4	50.4	70.2
G2 order-2	49.4	44.3	47.9	63.8
G3 order-3	51.2	48.6	50.4	67.0
G4 graph majority	52.4	51.4	52.1	73.4
G5 tight expansion	54.3	52.9	53.8	70.2
B1 tail window	48.2	41.4	46.2	64.9

Table 2: Accuracy by cohort and the preservation of questions the question-only control already answers. The pooled column is descriptive, because the two cohorts differ in both builder and mapping recall.

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

ClueWeaver: Reward-Guided Dual-Agent Evidence Reasoning for Compact LLMs on Literary Long Narratives
Aug 26, 2026Jihao Zhu, Zhiwei Yang, Wenxiao Zhang +7Long-Context QA Multi-Agent LLM Systems
University of Aberdeen, Aberdeen, United Kingdom · Institute of Information Engineering, CAS, Beijing, China · The University of Western Australia, Perth, Australia +1
Beyond Context Length: Characterising the Difficulty of Long-Context Reasoning
Oct 15, 2026Shiwen Qin, Cameron Barker, Henry Gouk +4Long-Context Reasoning in Language Models Long-Context Language Model Evaluation
University of Edinburgh, Edinburgh, United Kingdom
ROZA Graphs: Self-Improving Near-Deterministic RAG through Evidence-Centric Feedback
Apr 8, 2026Matthew PenarozaRetrieval-Augmented Generation Graph Retrieval-Augmented Generation

Aug 26, 2026J/K move · Enter open · S save

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

Jihao Zhu, Zhiwei Yang, Wenxiao Zhang, Junqian Zhao+6

University of Aberdeen, Aberdeen, United Kingdom · Institute of Information Engineering, CAS, Beijing, China · The University of Western Australia, Perth, Australia+1

Long-Context QA Multi-Agent LLM Systems LLM Grounding Evidence Retrieval RL for Language Model Reasoning

Open paper AskPDF arXiv

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.

Abstract

Long-Context QA

Explore similar work

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

Beyond Context Length: Characterising the Difficulty of Long-Context Reasoning

ROZA Graphs: Self-Improving Near-Deterministic RAG through Evidence-Centric Feedback