ClueWeaver, a compact two-agent AI system, recorded 59.0% overall accuracy across four long-context narrative benchmarks and led the local methods on each one, according to an arXiv preprint. Its observed overall accuracy was 6.4 percentage points higher than IRCoT and 14.5 points higher than a direct reader using the same Qwen3-4B model. It remained 4.9 percentage points below the best API result overall.
The score describes performance on evaluated tasks, not outcomes for human literary analysis. The comparisons show that the systems' observed results differed, but they do not establish why. The report provides no confidence intervals or other estimates of uncertainty.
Separating evidence from interpretation
ClueWeaver divides the task between evidence selection and interpretation. A Finder selects clue passages identified by paragraph, then the system packs those passages in narrative order for an Interpreter. For high-risk answers, the same Interpreter can run a self-calibration step, checking the answer against that evidence packet. The design puts evidence selection before interpretation.
The two agents have separate reward objectives. The Finder is trained to retain clues with high recall and cite paragraphs faithfully. The Interpreter is trained for answer correctness, concise explanations grounded in the selected evidence, and resistance to unsupported inference. Both use Qwen3-4B-Instruct and were trained separately with GRPO, using eight sampled responses per prompt on a single node with eight NVIDIA A100 GPUs.
The training examples came from benchmark training splits. Finder keep/drop decisions were balanced 50/50 and included hard negatives. Interpreter training used 1,000 examples, 27.5% of which were base-model failures.
Results varied by benchmark
Performance varied across the tests. Accuracy was 55.8% on DetectiveQA, 63.8% on ∞Bench, 50.0% on LongBench v2 and 61.3% on NoCha. ∞Bench was the high point, while LongBench v2 was the low point.
On DetectiveQA, the full model scored 55.8%. The reported component variants scored 51.0% without Interpreter self-calibration, 50.0% without the Finder and 36.5% without both agents. A separate training ablation reported 49.0% without Finder RL, 54.8% without Interpreter RL and 50.0% without both reinforcement-learning components.
Those gaps are informative comparisons, but they are not causal evidence. The configurations do not establish that a particular component or training step caused the change.
Across the four benchmarks, ClueWeaver recovered 84 of the 172 errors made by the direct reader. The error analysis still identifies distant multi-hop clues and passages with weak surface overlap as residual trouble spots.
Traceability came at a cost
The trace audit covered 310 questions: 104 from DetectiveQA, 69 from ∞Bench, 26 from LongBench v2 and 111 from NoCha. Of the 310 outputs, 275 contained citations. Among 690 cited paragraph IDs, 685 were valid, a 99.3% rate, and 270 of the 275 citation-bearing outputs had every citation valid, a 98.2% rate. Structured tags appeared in 309 of 310 outputs, or 99.7%, and legal answer labels in 308 of 310, or 99.4%.
The audit's validity measure is narrow. It checks whether explicit paragraph references point to valid IDs, not whether those passages fully support the substance of an answer.
The pipeline also took longer than direct reading. In the reported single-GPU setup with Qwen3-4B, ClueWeaver required 8.6 to 9.8 seconds per question, compared with 2.8 seconds for direct reading, and most of the added latency was attributed to Finder calls. The reported GPU access was 24GB for ClueWeaver, versus 80GB for the listed Qwen3-30B-A3B and Gemma-4-31B-it local readers.
Where the evidence stops
The evaluation covered four long-context narrative settings, with the same test instances used for all methods. Its evidence is therefore limited to those benchmark tasks and does not by itself show how the system would perform in other settings.
Training data came from benchmark training splits and included synthetic claim-verification augmentation. The reported mixture is not a test of generalization beyond those data.
The authors interpret the benchmark and ablation results as support for explicit evidence selection, self-calibrated interpretation and reward-guided training as a practical approach to making compact local models more effective and inspectable on these tasks. They identify distant-clue retrieval and multi-hop evidence integration as priorities for future work.
The document identifies itself as arXiv:2608.25531v2 and reports that code is available at the cited GitHub repository. It acknowledges support from the National Key R&D Program of China, with WisPaper and QiewenPaper providing academic-agent support and GPU computational resources.
Paper data and sources
Original title: ClueWeaver: Reward-Guided Dual-Agent Evidence Reasoning for Compact LLMs on Literary Long Narratives
Authors: Jihao Zhu, Zhiwei Yang, Wenxiao Zhang et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text