ClueWeaver

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

Official Finder and Interpreter checkpoints for ClueWeaver: Reward-Guided Dual-Agent Evidence Reasoning for Compact LLMs on Literary Long Narratives, accepted at ICONIP 2026.

ClueWeaver is an evidence-aware dual-agent framework for long-narrative question answering with compact local models. The Finder selects clue-bearing passages from retrieval-guided segments. The Interpreter answers from the selected evidence, provides paragraph-ID citations, and applies internal self-calibration for high-risk questions. Self-calibration uses the same Interpreter and evidence packet; it is not a separate agent.

Both agents use Qwen3-4B-Instruct and are trained separately with GRPO. Finder rewards emphasize evidence retention and faithful paragraph references; Interpreter rewards emphasize answer correctness, grounding, and concise explanations. See Sections 3.4-3.5 and Appendices A-B of the paper.

Checkpoints

Model	Role
Finder	Evidence selection and rationale generation
Interpreter	Evidence-grounded interpretation

Each directory contains full model weights, configuration, and tokenizer files, not LoRA adapters. Optimizer states and datasets are not included.

Loading

Select the required agent using subfolder:

from transformers import AutoModelForCausalLM, AutoTokenizer

repo_id = "Ameame1002/ClueWeaver"
agent = "Finder"  # or "Interpreter"
tokenizer = AutoTokenizer.from_pretrained(repo_id, subfolder=agent)
model = AutoModelForCausalLM.from_pretrained(
    repo_id, subfolder=agent, torch_dtype="auto", device_map="auto"
)

For the complete pipeline, download the repository with huggingface_hub.snapshot_download, then serve the Finder/ and Interpreter/ directories with a Qwen3-compatible inference engine. Use the prompts, retrieval, paragraph identifiers, answer parser, and self-calibration logic from the GitHub repository. Loading the weights alone does not reproduce the full pipeline.

As specified in Appendix B, disable model-internal thinking during inference (enable_thinking=False when applying the Qwen3 chat template). The structured responses use <reason> and <answer> fields; prompt templates are provided in Appendix E.

Main-result evidence packing (N_E, P_r, P_w, B_c) follows Appendix C. Here, N_E limits selected segments, P_r and P_w are paragraph budgets for retrieval-anchored and local-window segments, and B_c is the packed-evidence character budget.

Dataset	Packing
DetectiveQA	(10, 4, 6, 15000)
InfiniteBench	(7, 3, 6, 14000)
LongBench v2	(8, 4, 6, 15000)
NoCha	(10, 6, 8, 16000)

Limitations

The Finder can omit important clues, and the Interpreter can produce incorrect answers or citations. Check generated reasoning against the source narrative. The released checkpoints target the paper's evidence-selection and reasoning tasks; generic chat prompting may behave differently.

Citation

@misc{zhu2026clueweaver,
  title={ClueWeaver: Reward-Guided Dual-Agent Evidence Reasoning for Compact LLMs on Literary Long Narratives},
  author={Jihao Zhu and Zhiwei Yang and Wenxiao Zhang and Junqian Zhao and Qi You and Fangqi Wang and Zheyuan Deng and Hanzhe Yang and Yu Liu and Jin B. Hong},
  year={2026},
  eprint={2608.25531},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2608.25531}
}