Elamur Research⌕
☰
home / papers / 2026 / Probing and Steering Uncertainty in Biomedical Language Models: Representational Structure and Behavioral Limits

Probing and Steering Uncertainty in Biomedical Language Models: Representational Structure and Behavioral Limits

Paper · 2026 · Debmalya Pal

Summary

The paper asks whether biomedical language models internally represent linguistic uncertainty, meaning epistemic hedging such as speculative or modal phrasing in clinical text, and whether that representation can be adjusted at inference time without retraining. The author builds a balanced 400-sentence contrast set from the BioScope corpus by collapsing speculation-cue annotations into sentence-level labels. Linear probes are then trained on mean-pooled hidden states of six models: two decoders (BioGPT, BioMedLM) and four BERT-style encoders (BioBERT, ClinicalBERT, BlueBERT, SciBERT). A steering layer is chosen with a 2/3-depth heuristic plus a small local search. The probe direction, optionally orthogonalized against length and hedge-count confounds, is added to activations through a forward hook. Uncertainty turns out to be linearly decodable in every model. Steering increases hedged output in the decoders and shifts encoder representations consistently across pretraining corpora. However, decodability does not guarantee clean steerability, and safe control depends on careful norm calibration and multi-metric evaluation. This matters for building more calibrated clinical text generation.

Key points

  • Probing accuracy. Probes reach 80–87% 5-fold CV accuracy across all six models (110M–2.7B parameters, five pretraining corpora).
  • Pooling choice. Mean pooling separates the two classes far better than last-token pooling in BioGPT (Cohen's d 1.33 vs. 0.09).
  • Layer selection. The ±2 window around the 2/3-depth anchor gains between +0.25pp and +2.50pp. The largest gain is for BioBERT, where L10 beats the L8 anchor.
  • Decoder steering. Dynamic hidden-norm calibration gives about 205 for BioGPT L16, versus a static estimate of 616. The static estimate caused perplexity above 400 at α_frac = 0.25, while calibration avoided collapse across the whole tested range. The two decoders behave differently:
    • BioGPT shows a non-monotonic hedging peak at α_frac = 0.125.
    • BioMedLM responds monotonically and without collapse.
  • Encoder steering and limitations. Encoder shifts are assessed with five metrics: movement of the projection along the probe axis, change in the probe's predicted probability, the rate at which predicted labels flip, the change in cosine similarity to class centroids, and the share of displacement lying on the axis. Flip rates are about 50%, and the response curves are nearly identical across encoders. Stated limitations include:
    • only radiology prompts were used for the decoders;
    • the hedge score relies on a 27-term lexicon;
    • there is no downstream task or expert evaluation.

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

🎓University of California San DiegoSan Diego · United States73 works · 51,948 citations

 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 worksInterpretability · 130 works

Cite

Cite as Debmalya Pal. 2026. Probing and Steering Uncertainty in Biomedical Language Models: Representational Structure and Behavioral Limits. BioNLP 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.bionlp-1.86

@inproceedings{pal2026probing,
  title = {{Probing and Steering Uncertainty in Biomedical Language Models: Representational Structure and Behavioral Limits}},
  author = {Debmalya Pal},
  booktitle = {BioNLP 2026},
  year = {2026},
  doi = {10.18653/v1/2026.bionlp-1.86},
  url = {https://doi.org/10.18653/v1/2026.bionlp-1.86}
}

References

27 references, 19 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Heike Adel and Hinrich Schütze. 2015. Comparing approaches to the coreference resolution of clinical speculation and negation. In Proceedings of the BioNLP 2015 Workshop. find
  2. Shashank Agarwal and Hongfang Yu. 2010. Detecting hedging in biomedical text using a semantic-based approach. In Proceedings of the BioNLP 2010 Workshop. find
  3. Emily Alsentzer, John R. Murphy, William Boag, Wei-Hung Weng, Di Jin, Tristan Naumann, and Matthew B. A. McDermott. 2019. Publicly available clinical BERT embeddings. In Proceedings of the 2nd Clinical Natural Language Processing Workshop, pages 72–78.
  4. Yonatan Belinkov. 2022. Probing classifiers: Promises, shortcomings, and advances. Computational Linguistics, 48(1):207–219.
  5. Iz Beltagy, Kyle Lo, and Arman Cohan. 2019. SciBERT: A pretrained language model for scientific text. 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 3615–3620.
  6. Elliot Bolton, Abhinav Venigalla, Michihiro Yasunaga, David Hall, Betty Xiong, Tony Lee, Roxana Daneshjou, Jonathan Frankle, Percy Liang, Michael Carbin, and Christopher D. Manning. 2024. BioMedLM: A 2.7b parameter language model trained on biomedical text. arXiv preprint arXiv:2403.18421.
  7. Yanai Elazar, Shauli Ravfogel, Alon Jacovi, and Yoav Goldberg. 2021. Amnesic probing: Behavioral explanation with amnesic counterfactuals. Transactions of the Association for Computational Linguistics, 9:160– 175.
  8. Jiayang Geng, Fei Cai, Yida Wang, Heinz Koeppl, Preslav Nakov, and Iryna Gurevych. 2024. A survey of confidence estimation and calibration in large language models. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Te… find
  9. John Hewitt and Christopher D. Manning. 2019. A structural probe for finding syntax in word representations. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4129–4138.
  10. Jinhyuk Lee, Wonjin Yoon, Sungdong Kim, Donghyeon Kim, Sunkyu Kim, Chan Ho So, and Jaewoo Kang. 2020. BioBERT: A pre-trained biomedical language representation model for biomedical text mining. Bioinformatics, 36(4):1234–1240.
  11. Yao Lu, Max Bartolo, Alastair Moore, Sebastian Riedel, and Pontus Stenetorp. 2022. Fantastically ordered prompts and where to find them: Overcoming few-shot prompt order sensitivity. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics, pages 8086–8098.
  12. Renqian Luo, Liai Sun, Yingce Xia, Tao Qin, Sheng Zhang, Hoifung Poon, and Tie-Yan Liu. 2022. BioGPT: Generative pre-trained transformer for biomedical text generation and mining. Briefings in Bioinformatics, 23(6). find
  13. David M. Panicek and Hedvig Hricak. 2016. How the language of radiology reports can foster (or foil) clinical communication. Journal of the American College of Radiology, 13(12):1394–1395. Editorial. find
  14. Arjun Panickssery, Nick Gabrieli, Julia Bowman, Meg Tong, Evan Hubinger, and Alexander Mallen. 2024. Steering llama 2 via contrastive activation addition. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 15504–15522. find
  15. Kiho Park, Yo Joong Choe, and Victor Veitch. 2024. The linear representation hypothesis and the geometry of large language models. In Proceedings of the 41st International Conference on Machine Learning.
  16. Yifan Peng, Shankai Yan, and Zhiyong Lu. 2019. Transfer learning in biomedical natural language processing: An evaluation of BERT and ELMo on ten benchmarking datasets. In Proceedings of the 18th BioNLP Workshop and Shared Task, pages 58–65.
  17. Abhilasha Ravichander, Yonatan Belinkov, and Eduard Hovy. 2021. Probing the probing paradigm: Does probing accuracy entail task relevance? In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics, pages 3363–3377.
  18. Karan Singhal, Shekoofeh Azizi, Tao Tu, S. Sara Mahdavi, Jason Wei, Hyung Won Chung, Nathan Scales, Ajay Tanwani, Heather Cole-Lewis, Stephen Pfohl, Perry Payne, Martin Seneviratne, Paul Gamble, Chris Kelly, Abubakr Babiker, Nathanael Schärli, Aakanksha Chowdhery, Philip Mansfield, Dina Demner-Fushm… — Large Language Models Encode Clinical Knowledge
  19. Mujeen Sung, Jinhyuk Lee, Sean Yi, Minji Jeon, Sungdong Kim, and Jaewoo Kang. 2021. Can language models be biomedical knowledge bases? In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 4723–4734.
  20. Ian Tenney, Dipanjan Das, and Ellie Pavlick. 2019. BERT rediscovers the classical nlp pipeline. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4593– 4601.
  21. Alex Turner, Lisa Thiergart, Gavin Leech, David Udell, Juan J. Vazquez, Ulisse Mini, and Monte MacDiarmid. 2023. Activation addition: Steering language models without optimization. arXiv preprint arXiv:2308.10248.
  22. Logesh Kumar Umapathi, Ankit Pal, and Malaikannan Sankarasubbu. 2023. Med-HALT: Medical hallucination test bed for large language models. arXiv preprint arXiv:2307.15343.
  23. Veronika Vincze, György Szarvas, Richárd Farkas, György Móra, and János Csirik. 2008. The Bio-Scope corpus: Biomedical texts annotated for uncertainty, negation and their scopes. BMC Bioinformatics, 9(Suppl 11):S9. find
  24. Ivan Vulić, Edoardo Maria Ponti, Robert Litschko, Goran Glavaš, and Anna Korhonen. 2020. Probing pretrained language models for lexical semantics. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 7222–7240.
  25. Miao Xiong, Zhiyuan Hu, Xinyang Lu, Yifei Li, Jie Fu, Junxian He, and Bryan Hooi. 2024. Can llms express their uncertainty? an empirical evaluation of confidence elicitation in large language models. In Proceedings of the 12th International Conference on Learning Representations. find
  26. Zihao Zhao, Eric Wallace, Shi Feng, Dan Klein, and Sameer Singh. 2021. Calibrate before use: Improving few-shot performance of language models. In Proceedings of the 38th International Conference on Machine Learning, pages 12697–12706.
  27. Andy Zou, Long Phan, Sarah Chen, James Campbell, Phillip Guo, Richard Ren, Alexander Pan, Xuwang Yin, Mantas Mazeika, Ann-Kathrin Dombrowski, Shashwat Goel, Nathaniel Li, Michael J. Byun, Zifan Wang, Alex Mallen, Steven Basart, Sanmi Koyejo, Dawn Song, Matt Fredrikson, and 2 others. 2023. Representa…

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/pal-2026-probing-steering-uncertainty-biomedical-language/