Elamur Research⌕
☰
home / papers / 2026 / Targeting the Needle, Ignoring the Haystack: Anchoring Crucial Cues for Evolving Scam Call Detection via an LLM-Assisted Classifier

Targeting the Needle, Ignoring the Haystack: Anchoring Crucial Cues for Evolving Scam Call Detection via an LLM-Assisted Classifier

Paper · 2026 · Tong Wu, Qinliang Su, Jianxing Yu et al.

Summary

The paper targets automatic detection of scam phone calls on online service platforms such as recruitment services. Labeled data in this setting is scarce, and fraudsters keep changing the cover story they use, for example moving from software jobs to livestreaming. Classifiers trained directly on labels tend to latch onto these variable narratives instead of the actual deceptive tactic. LLMs handle this better, but they are too costly to run on millions of daily calls.

The authors propose a lightweight two-stage framework built on Semantic Primitives, which are expert-defined, interpretable cues such as "claim of abundant job openings."

  • Stage 1: A compact detector, trained on LLM-curated and LLM-augmented data, predicts which primitives appear in a transcript.
  • Stage 2: Boolean rules combine the detected primitives into a scam or normal verdict.

The framework was evaluated on a real Chinese recruitment-call dataset and a public synthetic one. It stays robust where conventional baselines degrade sharply under shifting contexts, and it roughly matches an LLM-based detector at far lower inference cost. This makes it a practical option for large-scale or real-time deployment.

Key points

  • Datasets: RealCall, a de-identified collection of real recruitment-platform calls in Chinese, and SynthScamCall, a public synthetic dataset from Ma et al., 2025.

  • Detection logic: Each call is mapped to a binary Primitive Status Vector. Scam patterns are then written as logic formulas over that vector, such as "no specific job titles AND abundant openings claimed" (¬p1 ∧ p2).

  • Component ablation: The scores below come from the paper's Table 4, which builds on a basic two-stage setup.

    ConfigurationSynthScamCallRealCall
    Basic two-stage0.570.48
    + data augmentation, token-level evidence task, global–local representation alignment0.710.57

  • Span extraction ablation: Removing primitive-relevant span extraction, which supplies background-negative and local-evidence training instances, lowers the scores. SynthScamCall drops from 0.63 to 0.45, and RealCall drops from 0.54 to 0.49.

  • Further findings and limitations: Gains from synthetic data level off as the generated set grows, and the alignment weight is set to λ = 0.3. The method may inherit label noise from the LLM-generated supervision, and it has only been tested on Chinese recruitment calls.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Sun Yat-sen UniversityGuangzhou · China72 works · 9,502 citations🏢China MobileBeijing · China10 works · 21 citations

 Fields, topics and concepts

📚 Shelves 3The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksInformation Systems · 1,730 works

Cite

Cite as Tong Wu, Qinliang Su, Jianxing Yu, Bo Liang and Minhua Huang. 2026. Targeting the Needle, Ignoring the Haystack: Anchoring Crucial Cues for Evolving Scam Call Detection via an LLM-Assisted Classifier. Findings of the Association for Computational Linguistics: ACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.findings-acl.1025

@inproceedings{wu2026targeting,
  title = {{Targeting the Needle, Ignoring the Haystack: Anchoring Crucial Cues for Evolving Scam Call Detection via an LLM-Assisted Classifier}},
  author = {Tong Wu, Qinliang Su, Jianxing Yu, Bo Liang and Minhua Huang},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.1025},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.1025}
}

References

31 references, 6 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Mohammed Rasol Al Saidat, Suleiman Y Yerima, and Khaled Shaalan. 2024. Advancements of sms spam detection: A comprehensive survey of nlp and ml techniques. Procedia Computer Science, 244:248– 259. find
  2. Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, and 1 others. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877–1901.
  3. Wenbin Chen and Changqing Chen. 2024. Deep learning-based model for detecting fraudulent sms messages. In Proceedings of the 2024 2nd International Conference on Information Education and Artificial Intelligence, pages 346–350. find
  4. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies,…
  5. Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario March, and Victor Lempitsky. 2016. Domain-adversarial training of neural networks. Journal of machine learning research, 17(59):1–35. find
  6. Abdallah Ghourabi, Mahmood A Mahmood, and Qusay M Alzubi. 2020. A hybrid cnn-lstm model for sms spam detection in arabic and english messages. Future Internet, 12(9):156. find
  7. Xinxin Hu, Haotian Chen, Hongchang Chen, Shuxin Liu, Xing Li, Shibo Zhang, Yahui Wang, and Xiangyang Xue. 2023. Cost-sensitive gnn-based imbalanced learning for mobile social network fraud detection. IEEE Transactions on Computational Social Systems, 11(2):2675–2690. find
  8. Maximilian Ilse, Jakub Tomczak, and Max Welling. 2018. Attention-based deep multiple instance learning. In International conference on machine learning, pages 2127–2136. PMLR. find
  9. Ankit Kumar Jain, Kamaljeet Kaur, Naveen Kumar Gupta, and Ankit Khare. 2025. Detecting smishing messages using bert and advanced nlp techniques. SN Computer Science, 6(2):109. find
  10. Liming Jiang. 2024. Detecting scams using large language models. arXiv preprint arXiv:2402.03147. — [arXiv:2402.03147](https://arxiv.org/abs/2402.03147) find
  11. Yufeng Kou, Chang-Tien Lu, Sirirat Sirwongwattana, and Yo-Ping Huang. 2004. Survey of fraud detection techniques. In IEEE international conference on networking, sensing and control, 2004, volume 2, pages 749–754. IEEE. find
  12. Tao Lei, Regina Barzilay, and Tommi Jaakkola. 2016. Rationalizing neural predictions. arXiv preprint arXiv:1606.04155.
  13. Jun Li, Cheng Zhang, and Lanlan Jiang. 2024. Innovative telecom fraud detection: A new dataset and an advanced model with roberta and dual loss functions. Applied Sciences, 14(24):11628. find
  14. Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, and 1 others. 2024. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437. find
  15. Yang Liu, Xiang Ao, Zidi Qin, Jianfeng Chi, Jinghua Feng, Hao Yang, and Qing He. 2021. Pick and choose: a gnn-based imbalanced learning approach for fraud detection. In Proceedings of the web conference 2021, pages 3168–3177. find
  16. Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692.
  17. Ilya Loshchilov and Frank Hutter. 2017. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101. find
  18. Haoyu Ma, Qinliang Su, Minhua Huang, and Wu Kai. 2025. Detecting continuously evolving scam calls under limited annotation: A llm-augmented expert rule framework. In Findings of the Association for Computational Linguistics: EMNLP 2025, pages 5047– 5068. find
  19. Oded Maron and Tomás Lozano-Pérez. 1997. A framework for multiple-instance learning. Advances in neural information processing systems, 10. find
  20. Dare Azeez Oyeyemi and Adebola K Ojo. 2024. Sms spam detection and classification to combat abuse in telephone networks using natural language processing. arXiv preprint arXiv:2406.06578. — [arXiv:2406.06578](https://arxiv.org/abs/2406.06578) find
  21. ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 2157–2166. Hong-Kui Xu, Tong-Tong Jiang, Xin Li, and 1 others. 2022. Bilstm network fraud phone recognition based on attention mechanism. Computer Systems and Applications, 31(3):326–332. find
  22. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, and 1 others. 2025. Qwen3 technical report. arXiv preprint arXiv:2505.09388.
  23. Jiacheng Ye, Jiahui Gao, Qintong Li, Hang Xu, Jiangtao Feng, Zhiyong Wu, Tao Yu, and Lingpeng Kong. 2022a. Zerogen: Efficient zero-shot learning via dataset generation. arXiv preprint arXiv:2202.07922.
  24. Jiacheng Ye, Jiahui Gao, Zhiyong Wu, Jiangtao Feng, Tao Yu, and Lingpeng Kong. 2022b. Progen: Progressive zero-shot dataset generation via in-context feedback. In Findings of the association for computational linguistics: EMNLP 2022, pages 3671–3683. find
  25. Josh Jia-Ching Ying, Ji Zhang, Che-Wei Huang, Kuan-Ta Chen, and Vincent S Tseng. 2018. Fraudetec-tor+ an incremental graph-mining approach for efficient fraudulent phone call detection. ACM Transactions on Knowledge Discovery from Data (TKDD), 12(6):1–35. find
  26. YiFan Zhang, Xue Wang, Jian Liang, Zhang Zhang, Liang Wang, Rong Jin, and Tieniu Tan. 2023. Free lunch for domain adversarial training: Environment label smoothing. arXiv preprint arXiv:2302.00194. — [arXiv:2302.00194](https://arxiv.org/abs/2302.00194) find
  27. Mohammad Sadegh Rasooli and Joel R. Tetreault. 2015. Yara parser: A fast and accurate dependency parser. Computing Research Repository, arXiv:1503.06733. Version 2. — [arXiv:1503.06733](https://arxiv.org/abs/1503.06733) find
  28. Zitong Shen, Kangzhong Wang, Youqian Zhang, Grace Ngai, and Eugene Yujun Fu. 2025a. Combating phone scams with llm-based detection: Where do we stand?(student abstract). In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 29487–29489. find
  29. Zitong Shen, Sineng Yan, Youqian Zhang, Xiapu Luo, Grace Ngai, and Eugene Yujun Fu. 2025b. " it warned me just at the right moment": Exploring llm-based real-time detection of phone scams. In Proceedings of the Extended Abstracts of the CHI Conference on Human Factors in Computing Systems, pages 1–7… find
  30. Gurjot Singh, Prabhjot Singh, and Maninder Singh. 2025. Advanced real-time fraud detection using rag-based llms. arXiv preprint arXiv:2501.15290. — [arXiv:2501.15290](https://arxiv.org/abs/2501.15290) find
  31. Vincent S Tseng, Jia-Ching Ying, Che-Wei Huang, Yimin Kao, and Kuan-Ta Chen. 2015. Fraudetec-tor: A graph-mining-based framework for fraudulent phone call detection. In Proceedings of the 21th find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/wu-2026-targeting-needle-ignoring-haystack-anchoring/