Elamur Research⌕
☰
home / papers / 2026 / KDW at SemEval-2026 Task 12: Logic-Driven Distillation with Knowledge Graphs for Efficient Abductive Reasoning

KDW at SemEval-2026 Task 12: Logic-Driven Distillation with Knowledge Graphs for Efficient Abductive Reasoning

Paper · 2026 · Sihan Zhu, Hongjie Wu and Xinyan Xu

Summary

The paper describes the KDW team's entry to SemEval-2026 Task 12, which targets abductive event reasoning. Given a target event and long, noisy related documents, a system must pick the direct causes from several candidate options or decide that "None of the others" (NOTA) applies. Running large LLMs on this is costly, so the authors aim for a small, cheap model that can still reason well.

For each option, they build a sentence-level evidence graph from retrieved documents. A shortest path through this graph links the hypothesis to the event and serves as an evidence chain. Using these chains, DeepSeek-R1 writes structured chain-of-thought rationales. Hidden states from Qwen2.5-14B, as it reads those rationales, are then distilled into a DeBERTa-v3-Large student. Training adds entity masking, contrastive alignment and curriculum-style loss weighting. At inference, the student scores each option and applies a threshold together with a NOTA rule.

The system scores 0.88 on the official test set and ranks 8th, beating zero-shot DeepSeek-V3.2 (0.72). Inference runs on a single T4 GPU, which shows that compact distilled models can compete with frontier LLMs here.

Key points

  • Evidence graph:
    • Nodes are spaCy-split sentences; edges are built with asymmetric KNN (top-6 neighbours for endpoint nodes, top-4 for evidence nodes).
    • Edge weights combine all-MiniLM-L6-v2 similarity, named-entity overlap and an adjacency bonus within the same document.
    • Paths are found with Dijkstra, and chains longer than three hops are penalised.
  • Distillation setup:
    • DeepSeek-R1 produces four-step rationales covering the event, temporal order, causal sufficiency and directness.
    • Qwen2.5-14B hidden states from layers 24 and 48 (5120-dim) are aligned with the 1024-dim DeBERTa-v3-Large student through MLP projection heads and a cosine loss.
  • Scoring and test results:
    • Each instance earns 1.0 for an exact match and 0.5 for a partial match.
    • On the test set, the full system scores 0.88, against 0.72 for zero-shot DeepSeek-V3.2, 0.71 for BGE-Distilled DeBERTa and 0.36 for Vanilla DeBERTa.
  • Ablations: removing a component lowers the score to:
    • 0.67 without CoT distillation;
    • 0.73 without intermediate-layer alignment;
    • 0.79 without KG evidence;
    • 0.81 without contrastive learning;
    • 0.83 without entity masking.
  • Resources:
    • Training used one A100 GPU, and inference needs only one T4.
    • Code is on GitHub, and the official task data come from the sooo66/semeval2026-task12-dataset repository.

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

🎓National University of SingaporeSingapore · Singapore180 works · 34,304 citations🎓University of TübingenTübingen · Germany75 works · 41,081 citations

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksComputer Vision and Pattern Recognition · 2,869 works

🧭 Topics 3Research subjects and languages the work is about.Reasoning · 1,426 worksKnowledge graphs & ontologies · 192 worksEfficient models & compression · 242 works

Cite

Cite as Sihan Zhu, Hongjie Wu and Xinyan Xu. 2026. KDW at SemEval-2026 Task 12: Logic-Driven Distillation with Knowledge Graphs for Efficient Abductive Reasoning. Proceedings of the 20th International Workshop on Semantic Evaluation (2026).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.semeval-1.124

@inproceedings{zhu2026kdw,
  title = {{KDW at SemEval-2026 Task 12: Logic-Driven Distillation with Knowledge Graphs for Efficient Abductive Reasoning}},
  author = {Sihan Zhu, Hongjie Wu and Xinyan Xu},
  booktitle = {Proceedings of the 20th International Workshop on Semantic Evaluation (2026)},
  year = {2026},
  doi = {10.18653/v1/2026.semeval-1.124},
  url = {https://doi.org/10.18653/v1/2026.semeval-1.124}
}

References

16 references, 10 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Chandra Bhagavatula, Ronan Le Bras, Chaitanya Malaviya, Keisuke Sakaguchi, Ari Holtzman, Hannah Rashkin, Doug Downey, Scott Yih, and Yejin Choi. 2020. Abductive commonsense reasoning. In International Conference on Learning Representations (ICLR).
  2. Jianlv Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. 2024. M3-embedding: Multilinguality, multi-functionality, multi-granularity text embeddings through self-knowledge distillation. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Vol… find
  3. Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. 2018. Think you have solved question answering? try ARC, the AI2 reasoning challenge. arXiv preprint arXiv:1803.05457.
  4. DeepSeek-AI and others. 2025. DeepSeek-V3.2: Pushing the frontier of open large language models. arXiv preprint arXiv:2512.02556.
  5. Ming Ding, Chang Zhou, Qibin Chen, Hongxia Yang, and Jie Tang. 2019. Cognitive graph for multi-hop reading comprehension at scale. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2694–2703. find
  6. Guo, Daya and Yang, Dejian and Zhang, Haowei and Song, Junxiao and Wang, Peiyi and Zhu, Qihao and others. 2025. DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning. Nature, 645(8081):633–638. find
  7. Pengcheng He, Jianfeng Gao, and Weizhu Chen. 2023. DeBERTaV3: Improving DeBERTa using ELECTRA-style pre-training with gradient-disentangled embedding sharing. In International Conference on Learning Representations (ICLR).
  8. Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. 2021. DeBERTa: Decoding-enhanced BERT with disentangled attention. In International Conference on Learning Representations (ICLR).
  9. Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. 2015. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531.
  10. Matthew Honnibal, Ines Montani, Sofie Van Lan-deghem, and Adriane Boyd. 2020. spaCy: Industrial-strength natural language processing in python. Zenodo. find
  11. Xiaobo Liang, Lijun Wu, Juntao Li, Yue Wang, Qi Meng, Tao Qin, Wei Chen, Min Zhang, and Tie-Yan Liu. 2021. R-Drop: Regularized Dropout for Neural Networks. In Advances in Neural Information Processing Systems (NeurIPS), volume 34, pages 10871–10882. find
  12. Bill Yuchen Lin, Xinyue Chen, Jamin Chen, and Xiang Ren. 2019. KagNet: Knowledge-aware graph networks for commonsense reasoning. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-… find
  13. Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. SQuAD: 100,000+ Questions for Machine Comprehension of Text. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2383–2392, Austin, Texas. Association for Computational Linguistics.
  14. Nils Reimers and Iryna Gurevych. 2019. Sentence-BERT: Sentence embeddings using siamese BERT-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3982–3992.
  15. Srivastava, Aarohi and Rastogi, Abhinav and Rao, Abhishek and others. 2023. Beyond the imitation game: Quantifying and extrapolating the capabilities of language models. Transactions on Machine Learning Research (TMLR).
  16. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. 2022. Chain-of-Thought prompting elicits reasoning in large language models. In Advances in Neural Information Processing Systems (NeurIPS), volume 35, pages 24824–24837.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhu-2026-kdw-semeval-2026-task-12/