Elamur Research⌕
☰
home / papers / 2026 / SRCB at SemEval-2026 Task 5 A Multi-Target Finetuning Framework for Large Language Models with Joint Regression and Text Generation

SRCB at SemEval-2026 Task 5 A Multi-Target Finetuning Framework for Large Language Models with Joint Regression and Text Generation

Paper · 2026 · Yuming Zhang, Junyu Zhou, Hongyu Li et al.

Summary

The paper describes the first-ranked system for SemEval-2026 Task 5. In this task, models do not pick one correct sense of a homonym. They instead predict how plausible humans find a given sense in a short story, on a 1–5 scale. Evaluation uses Spearman correlation together with an accuracy measure that counts a prediction as correct if it lies no more than one standard deviation from the human ratings. The authors fine-tune decoder-only LLMs, mainly Qwen3-14B and Qwen3-30B, with two objectives that share one backbone:

  • a regression head that outputs a continuous score;
  • a generation branch that writes a rationale, a coarse verbal option and a numeric score.

The final prediction averages the two outputs. Annotators often disagree, so the team adds several robustness measures:

  • auxiliary training data produced by Qwen3-Next-80B-A3B, explaining why different endings change plausibility;
  • trimming or smoothing of high-variance labels;
  • separate heads that model individual annotators;
  • a batch-wise differentiable Spearman loss.

An ensemble of these variants placed first overall and on both metrics. The work shows that combining regression with text generation helps when the target is graded and subjective rather than a single correct label.

Key points

  • Data: the AmbiStory dataset (English, five-sentence stories) has 2280 training, 588 development and 930 test instances.
  • Results: the system reached 93.33% accuracy within SD and 85.57% Spearman on the test set, an average of 89.45%. It ranked 1st on both metrics.
  • Module combination: weighting the regression and generation outputs equally gave the best cross-validation average of 86.88%. This beat either module used alone.
  • Ensemble: combining variants added roughly 3 points over single models. Even weak members helped, and 5-fold cross-validation scores tracked test scores closely.
  • Baselines: BERT-style baselines (DeBERTa-v3-large) and prompting DeepSeek-v3 performed clearly worse than the fine-tuned LLMs. The main stated limitation is slow inference, because the generation step is costly.

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

 Fields, topics and concepts

📚 Shelves 4The 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 works

🧭 Topics 2Research subjects and languages the work is about.Language models & pretraining · 3,320 worksText generation · 202 works

Cite

Cite as Yuming Zhang, Junyu Zhou, Hongyu Li, Yongwei Zhang, Shanshan Jiang and Bin Dong. 2026. SRCB at SemEval-2026 Task 5 A Multi-Target Finetuning Framework for Large Language Models with Joint Regression and Text Generation. 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-1.246

@inproceedings{zhang2026srcb,
  title = {{SRCB at SemEval-2026 Task 5 A Multi-Target Finetuning Framework for Large Language Models with Joint Regression and Text Generation}},
  author = {Yuming Zhang, Junyu Zhou, Hongyu Li, Yongwei Zhang, Shanshan Jiang and Bin Dong},
  booktitle = {Proceedings of the 20th International Workshop on Semantic Evaluation (2026)},
  year = {2026},
  doi = {10.18653/v1/2026.semeval-1.246},
  url = {https://doi.org/10.18653/v1/2026.semeval-1.246}
}

References

9 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Mathieu Blondel, Olivier Teboul, Quentin Berthet, and Josip Djolonga. 2020a. Fast differentiable sorting and ranking. In Proceedings of the 37th International Conference on Machine Learning, ICML’20. JMLR.org. find
  2. Mathieu Blondel, Olivier Teboul, Quentin Berthet, and Josip Djolonga. 2020b. Fast differentiable sorting and ranking. In International conference on machine learning, pages 950–959. PMLR. find
  3. Janosch Gehring, Selina Meyer, and Michael Roth. 2026. SemEval-2026 task 5: Rating plausibility of word senses in ambiguous stories through narrative understanding. In Proceedings of the 20th International Workshop on Semantic Evaluation, San Diego, California. Association for Computational Linguist… find
  4. Pengcheng He, Jianfeng Gao, and Weizhu Chen. 2021. Debertav3: Improving deberta using electra-style pre-training with gradient-disentangled embedding sharing. arXiv preprint arXiv:2111.09543.
  5. Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Liang Wang, Weizhu Chen, and 1 others. 2022. Lora: Low-rank adaptation of large language models. Iclr, 1(2):3.
  6. Wenhang Shi, Shuqing Bian, Yiren Chen, Xinyi Zhang, Zhe Zhao, Pengfei Hu, Wei Lu, and Xiaoyong Du. 2025. Investigating the impact of rationales for llms on natural language understanding. Preprint, arXiv:2510.16686. — [arXiv:2510.16686](https://arxiv.org/abs/2510.16686) find
  7. Qwen Team. 2025a. Qwen3-next: Towards ultimate training & inference efficiency. find
  8. Qwen Team. 2025b. Qwen3 technical report. Preprint, arXiv:2505.09388.
  9. Linhan Xia, Mingzhan Yang, Guohui Yuan, Shengnan Tao, Yujing Qiu, Guo Yu, and Kai Lei. 2025. Polybert: Fine-tuned poly encoder bert-based model for word sense disambiguation. Preprint, arXiv:2506.00968. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhang-2026-srcb-semeval-2026-task-5/