Elamur Research⌕
☰
home / papers / 2026 / Dynamic Model Switching to Mitigate Outdated Knowledge in Large Language Models

Dynamic Model Switching to Mitigate Outdated Knowledge in Large Language Models

Summary

Large language models trained on fixed data snapshots gradually become outdated, which makes them less reliable. The authors address this by routing each input to one of two generators instead of repeatedly retraining a single model. One is a large model with broad but older knowledge (Phi-3-medium-128k-instruct). The other is a small model fine-tuned on recent material (Phi-3-mini-4k-instruct). A lightweight switch, trained in a multitask setup, makes the choice using contextual and temporal cues. Its training combines a model-selection loss, a semantic cross-entropy loss and a temporal-validity loss, and an adaptive threshold weighs aggregated reward signals to estimate whether a text has become obsolete.

For evaluation, the authors built TaDD, a dataset of revision pairs from Wikipedia and Guardian articles. Each pair is labelled as either a temporal (factual) update or a purely semantic edit. The switch reaches a balanced precision–recall profile, although several fine-tuned 7B–14B models score higher in absolute terms. Its appeal is efficiency: only the router is trained. Possible uses include flagging stale content and supporting fact-checking.

Key points

  • TaDD dataset: 2,500 sentence pairs (1,081 temporal updates, 1,419 semantic updates). Together with Edit-Intentions, Wiki-TIDE and One Document, Many Revisions, the benchmark pool totals 9,504 samples.
  • Annotation: GPT-4.1 Mini produced draft labels, and human annotators made every final decision. Inter-annotator agreement was Cohen's kappa of 0.664.
  • Efficiency: the switch module has about 120M trainable parameters. At inference it uses both generators (14B + 3.8B), and neither generator is retrained.
  • Results: on the Combined Temporal Validity data, Dynamic-Switch scores about 0.65–0.66 on accuracy, precision, recall and F1. For comparison, Mistral-7B Fine-Tuned reaches 0.75 accuracy and 0.75 F1.
  • Availability: the code and dataset are released on GitHub.

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: Linguistics & languages

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.63317/2n9poorzn8s3

@inproceedings{pinninti2026dynamic,
  title = {{Dynamic Model Switching to Mitigate Outdated Knowledge in Large Language Models}},
  author = {Pinninti, Ramakrishna and Kamila, Sabyasachi and Mazumder, Ayan and Hasanuzzaman, Mohammed},
  booktitle = {Proceedings of the Fifteenth Language Resources and Evaluation Conference},
  year = {2026},
  doi = {10.63317/2n9poorzn8s3},
  url = {https://doi.org/10.63317/2n9poorzn8s3}
}

References

24 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Axel Almquist and Adam Jatowt. 2019. Towards content expiry date determination: Predicting validity periods of sentences. In Advances in Information Retrieval: 41st European Conference on IR Research, ECIR 2019, Cologne, Germany, April 14–18, 2019, Proceedings, Part I, volume 11437 of Lecture Notes… find
  2. Hsuvas Borkakoty and Luis Espinosa-Anke. 2023. Wikitide: A wikipedia-based timestamped definition pairs dataset. find
  3. Shuyang Cao and Lu Wang. 2022. Time-aware prompting for text generation. In Findings of the Association for Computational Linguistics: EMNLP 2022, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics. find
  4. Taishi Hosokawa, Adam Jatowt, and Kazunari Sugiyama. 2024. Temporal validity reassessment: Commonsense reasoning about information obsoleteness. Discovery Computing, 27:4. find
  5. Raghav Jain, Daivik Sojitra, Arkadeep Acharya, Sriparna Saha, Adam Jatowt, and Sandipan Dandapat. 2023. Do language models have a common sense regarding time? revisiting temporal commonsense reasoning in the era of large language models. In Proceedings of the 2023 Conference on Empirical Methods in… find
  6. Adam Jatowt, Ching-Man Au Yeung, and Katsumi Tanaka. 2013. Estimating document focus time. In Proceedings of the 22nd ACM International Conference on Information & Knowledge Management, CIKM ’13, pages 2273–2278, New York, NY, USA. Association for Computing Machinery. find
  7. Sunbowen Lee, Junting Zhou, Chang Ao, Kaige Li, Xinrun Du, Sirui He, Haihong Wu, Tianci Liu, Jiaheng Liu, Hamid Alinejad-Rokny, Min Yang, Yitao Liang, Zhoufutu Wen, and Shiwen Ni. 2025. Quantification of large language model distillation. find
  8. Xiang Lisa Li, Ari Holtzman, Daniel Fried, Percy Liang, Jason Eisner, Tatsunori Hashimoto, Luke Zettlemoyer, and Mike Lewis. 2023. Contrastive decoding: Open-ended text generation as optimization. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: L… find our card
  9. Eric Mitchell, Charles Lin, Antoine Bosselut, Christopher D. Manning, and Chelsea Finn. 2022. Memory-based model editing at scale. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (EMNLP), Online. Association for Computational Linguistics. find
  10. OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2024. Gpt-4 technical report. ArXiv preprint arXiv:2303.08774. find our card
  11. Bhawna Piryani, Abdelrahman Abdallah, Jamshid Mozafari, Avishek Anand, and Adam Jatowt. 2025. It’s high time: A survey of temporal information retrieval and question answering. find
  12. Dheeraj Rajagopal, Xuchao Zhang, Michael Gamon, Sujay Kumar Jauhar, Diyi Yang, and Eduard Hovy. 2022. One document, many revisions: A dataset for classification and description of edit intents. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 5517–5524, Marseille,… find
  13. Rishav Ranaut, Sriparna Saha, Adam Jatowt, and Manish Gupta. 2025. Text obsoleteness detection using large language models. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2827–2831, Padua, Italy. Association for Computing Mac… find
  14. Gerald Rau and Yu-Shan Shih. 2021. Evaluation of cohen’s kappa and other measures of inter-rater agreement for genre analysis and other nominal data. Journal of English for Academic Purposes, 53:101026. find
  15. Daivik Sojitra, Raghav Jain, Sriparna Saha, Adam Jatowt, and Manish Gupta. 2024. Timeline summarization in the era of llms. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2657–2661, Washington DC, USA. Association for Computi… find
  16. Alexander Spangher, Kung-Hsiang Huang, Hyundong Cho, and Jonathan May. 2024. Newsedits 2.0: Learning the intentions behind updating news. find
  17. Jonas Wallat, Abdelrahman Abdallah, Adam Jatowt, and Avishek Anand. 2025. A study into investigating temporal robustness of LLMs. In Findings of the Association for Computational Linguistics: ACL 2025, pages 15685–15705, Vienna, Austria. Association for Computational Linguistics. find
  18. Jonas Wallat, Adam Jatowt, and Avishek Anand. 2024. Temporal blind spots in large language models. In Proceedings of the 17th ACM International Conference on Web Search and Data Mining, WSDM ’24, pages 683–692, Merida, Mexico. Association for Computing Machinery. find
  19. Jiapu Wang, Kai Sun, Linhao Luo, Wei Wei, Yongli Hu, Alan Wee-Chung Liew, Shirui Pan, and Baocai Yin. 2025. Large language models-guided dynamic adaptation for temporal knowledge graph reasoning. In Proceedings of the 38th International Conference on Neural Information Processing Systems, NIPS ’24,… find
  20. Georg Wenzel and Adam Jatowt. 2023. An overview of temporal commonsense reasoning and acquisition. find
  21. Georg Wenzel and Adam Jatowt. 2024. Temporal validity change prediction. In Findings of the Association for Computational Linguistics: ACL 2024, ACL Findings, pages 1424–1446, Bangkok, Thailand. Association for Computational Linguistics. find
  22. Diyi Yang, Aaron Halfaker, Robert Kraut, and Eduard Hovy. 2017. Identifying semantic edit intentions from revisions in wikipedia. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2000– 2010, Copenhagen, Denmark. Association for Computational Linguistic… find
  23. Xiaoying Zhang, Baolin Peng, Ye Tian, Jingyan Zhou, Yipeng Zhang, Haitao Mi, and Helen Meng. 2025. Self-tuning: Instructing llms to effectively acquire new knowledge through self-teaching. find
  24. Ben Zhou, Daniel Khashabi, Qiang Ning, and Dan Roth. 2019. “going on a vacation” takes longer than “going for a walk”: A study of temporal commonsense understanding. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference o… find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/pinninti-2026-dynamic-model-switching-mitigate-outdated/