Elamur Research⌕
☰
home / papers / 2026 / Perspicere at SemEval-2026 Task 2: Modeling Longitudinal Valence and Arousal via Dense Embeddings and Agentic Reasoning

Perspicere at SemEval-2026 Task 2: Modeling Longitudinal Valence and Arousal via Dense Embeddings and Agentic Reasoning

Paper · 2026 · Kamyar Moradian Zehab, Mohammad Sadegh Poulaei and Nasser Mozayani

Summary

The paper describes a system for SemEval-2026 Task 2 (Subtask 1), where continuous Valence and Arousal scores must be predicted from time-ordered series of first-person "ecological essays" and feeling-word lists. The authors compare three parallel approaches, each using more context than the last:

  • Static regression: Ridge, XGBoost and SVR trained on frozen, L2-normalized embeddings from BERT, RoBERTa, XLM-RoBERTa, E5-Large, Qwen3 and Jasper.
  • Temporal modeling: an LSTM run over each user's history.
  • TG-CRA agent: an open-weights LLM given few-shot reasoning examples written by a frontier teacher model.

Feeling-word lists are first rewritten as sentences starting with "I am feeling" to bring them closer to the essay format.

Results:

  • Per-entry regression beats sequence modeling. Jasper embeddings with XGBoost, trained on the full training data, give the best overall Valence score.
  • The LSTM falls behind, likely because entries arrive at sparse, uneven intervals.
  • The agent is weaker at tracking moment-to-moment changes but best at capturing stable differences between users.

Arousal remains much harder to predict than Valence from text alone. The work suggests that efficient distilled embeddings and interpretable LLM reasoning serve complementary roles in longitudinal affect modeling.

Key points

  • Best system: Jasper (1.9B, Matryoshka-distilled) with XGBoost reached Valence composite r = 0.654, 17.4% above the baseline. This was a refinement made after the official submission, which used Jasper-SVR tuned on data subsets. On those subsets, SVR (0.654) outperformed an overfitting XGBoost (0.572).
  • LSTM vs. static models: the LSTM underperformed static heads. Qwen3-4B dropped from 0.623 (SVR) to 0.564 (LSTM), and Jasper from 0.654 to 0.622.
  • TG-CRA agent: it achieved the top Between-User Valence correlation (r = 0.725) but a lower composite score (0.637). Qualitatively, it handled elongated spellings such as "goooooood" and treated spiritual language as a softener of negative mood.
  • Arousal gap: Arousal peaked at roughly r composite 0.53, versus about 0.65 for Valence.
  • Speed (NVIDIA T4, batch size 32):
    • Jasper needs 1.41s per batch, about 7× slower than BERT-Base.
    • Qwen3-4B needs 5.37s per batch, about 26.8× slower than BERT-Base.

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 5The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksPsychology · 1,649 worksExperimental and Cognitive Psychology · 930 worksSocial Psychology · 330 works

🧭 Topics 1Research subjects and languages the work is about.Reasoning · 1,685 works

Cite

Cite as Kamyar Moradian Zehab, Mohammad Sadegh Poulaei and Nasser Mozayani. 2026. Perspicere at SemEval-2026 Task 2: Modeling Longitudinal Valence and Arousal via Dense Embeddings and Agentic 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-1.97

@inproceedings{zehab2026perspicere,
  title = {{Perspicere at SemEval-2026 Task 2: Modeling Longitudinal Valence and Arousal via Dense Embeddings and Agentic Reasoning}},
  author = {Kamyar Moradian Zehab, Mohammad Sadegh Poulaei and Nasser Mozayani},
  booktitle = {Proceedings of the 20th International Workshop on Semantic Evaluation (2026)},
  year = {2026},
  doi = {10.18653/v1/2026.semeval-1.97},
  url = {https://doi.org/10.18653/v1/2026.semeval-1.97}
}

References

35 references, 12 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Bagus Tris Atmaja and Masato Akagi. 2021. Two-stage dimensional emotion recognition by fusing predictions of acoustic and text networks using svm. Speech Communication, 126:9–21. 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. In NeurIPS.
  3. Bing Liu. 2012. Sentiment analysis and opinion mining. Morgan and Claypool Publishers. find
  4. 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.
  5. Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In International Conference on Learning Representations. find
  6. Tianqi Chen and Carlos Guestrin. 2016. XGBoost: A scalable tree boosting system. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 785–794. Association for Computing Machinery. find
  7. Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2020. Unsupervised cross-lingual representation learning at scale. In Proceedings of the 58th Annual Meeting of the Association fo…
  8. 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…
  9. Harris Drucker, Chris J.C. Burges, Linda Kaufman, Alex Smola, and Vladimir Vapnik. 1997. Support vector regression machines. Advances in Neural Information Processing Systems, 9. find
  10. Panagiotis Giannos and Orestis Delardas. 2024. Applications of large language models in psychiatry: A systematic review. Frontiers in Psychiatry, 15:1422807. find
  11. Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long short-term memory. Neural computation, 9(8):1735– 1780. find
  12. Yoon Kim. 2014. Convolutional neural networks for sentence classification. arXiv preprint arXiv:1408.5882.
  13. Peter Kuppens, Nicholas B Allen, and Lisa B Sheeber. 2010. Emotional inertia and psychological malad-justment. Psychological science, 21(7):984–991. find
  14. Peter Kuppens, Zita Oravecz, and Francis Tuerlinckx. 2012. The nature of emotional inertia: A matter of temporal structure rather than intensity or variability. Psychological Review, 119(3):1161–1174. find
  15. Aditya Kusupati, Gantavya Bhatt, Aniket Rber, Matthew Wallingford, Aditya Sinha, Vivek Prasanna, Ali Farhadi, Sham Kakade, and Prateek Jain. 2024. Matryoshka representation learning. In Advances in Neural Information Processing Systems (NeurIPS). find
  16. Navonil Majumder, Soujanya Poria, Devamanyu Hazarika, Rada Mihalcea, Alexander Gelbukh, and Erik Cambria. 2019. Dialoguernn: An attentive rnn for emotion detection in conversations. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 6818–6825. find
  17. OpenAI, Sandhini Agarwal, Lama Ahmad, Jason Ai, Sam Altman, Andy Applebaum, Edwin Arbus, Rahul K. Arora, Yu Bai, Bowen Baker, and 1 others. 2025. GPT-OSS-120B & GPT-OSS-20B model card. arXiv preprint arXiv:2508.10925.
  18. Bo Pang, Lillian Lee, and 1 others. 2008. Opinion mining and sentiment analysis. Foundations and Trends® in Information Retrieval, 2(1–2):1–135. find
  19. Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, and 2 others. 2019.… — PyTorch: An Imperative Style, High-Performance Deep Learning Library
  20. Fabian Pedregosa, Gaël Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, Jake Vanderplas, Alexandre Passos, David Cournapeau, Matthieu Brucher, Matthieu Perrot, and Édouard Duchesnay. 2011. Scikit-learn:… find
  21. James A. Russell. 1980. A circumplex model of affect. Journal of Personality and Social Psychology, 39(6):1161–1178. find
  22. Björn W Schuller. 2018. Speech emotion recognition: Two decades in a nutshell, benchmarks, and ongoing trends. Communications of the ACM, 61(5):90–99. find
  23. Nikita Soni, H. Andrew Schwartz, Ryan L. Boyd, Phi Long Bui, Syeda Mahwish, August Håkan Nilsson, Adithya V Ganesan, Lyle Ungar, Niranjan Balasubramanian, and Saif M. Mohammad. 2026. SemEval-2026 task 2: Predicting variation in emotional valence and arousal over time from ecological essays. In Proce…
  24. Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. 2024. Multilingual e5 text embeddings: A technical report. arXiv preprint arXiv:2402.05672. – mE5-Large (Wang et al., 2024): Alpha: Standard range (0.0–1.0) plus extended range: 1.1 to 2.0 (step 0.1).
  25. Distilled Multimodal Embeddings (Jasper (Zhang et al., 2024)): 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 Sys…
  26. Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pier-ric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, and 3 others. 2020. Transf…
  27. Generative LLM Embeddings (Qwen3 (Zhang et al., 2025)): – Alpha: Standard range (0.0–1.0) plus extended range: 1.1 to 2.0 (step 0.1). \#\# 2. find
  28. XGBoost (Chen and Guestrin, 2016): • Learning Rate: 0.01, 0.05 find
  29. Max Depth: 3, 4, 6 • Min Child Weight: 4, 6, 8 • Subsample: Fixed at 0.8 3. Support Vector Regression (SVR) (Drucker et al., 1997): Xuhai Xu, Bingshen Yao, Yuanzhe Dong, Saadia Gabriel, Hong Yu, James Hendler, Marzyeh Ghassemi, Anind K. Dey, and Dakuo Wang. 2024. Mental-LLM: Leveraging large languag… find
  30. Bidirectional Encoder Embeddings: – BERT (Devlin et al., 2019), RoBERTa (Liu et al., 2019), XLM-RoBERTa (Conneau et al., 2020): Kernel: Linear, RBF; C (Regularization): 0.01, 0.1, 1.0, 5.0, 10.0; Epsilon (Margin): 0.01, 0.1, 0.2, 0.3. – mE5-Large (Wang et al., 2024): Kernel: Linear, RBF; C: 0.1, 1.0… find
  31. Kailai Yang, Shaoxiong Ji, Tianlin Zhang, Qianqian Xie, and Sophia Ananiadou. 2023. Mentallama: Interpretable mental health analysis on social media with large language models. arXiv preprint arXiv:2309.13567. — [arXiv:2309.13567](https://arxiv.org/abs/2309.13567) find
  32. Distilled Multimodal Embeddings (Jasper (Zhang et al., 2024)): – Kernel: Linear, RBF – C (Extended): Added 2.0, 2.5, 3.0, 4.0, 15.0, 20.0 – Epsilon (Extended): Added 0.4 to 0.9 (step 0.1) find
  33. Dun Zhang, Jiacheng Li, Ziyang Zeng, and Fulong Wang. 2024. Jasper and stella: Distillation of sota embedding models. arXiv preprint arXiv:2412.19048. — [arXiv:2412.19048](https://arxiv.org/abs/2412.19048) find
  34. Yanzhao Zhang, Mingxin Li, Dingkun Long, Xin Zhang, Huan Lin, Baosong Yang, Pengjun Xie, An Yang, Dayiheng Liu, Junyang Lin, Fei Huang, and Jingren Zhou. 2025. Qwen3 embedding: Advancing text embedding and reranking through foundation models. arXiv preprint arXiv:2506.05176.
  35. A • Generative LLM Embeddings (Qwen3 (Zhang et al., 2025)): – Kernel: Linear, RBF – C: 0.1, 1.0, 2.0 – Epsilon: 0.1 to 0.7 (step 0.1) 4. Long Short-Term Memory (LSTM) (Hochreiter and Schmidhuber, 1997): Hyperparameter Specifications • Hidden Dimension: 64, 128, 256 find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zehab-2026-perspicere-semeval-2026-task-2/