Elamur Research⌕
☰
home / papers / 2026 / Linear Probes Detect Task Format, Not Reasoning Mode in Language Model Hidden States

Linear Probes Detect Task Format, Not Reasoning Mode in Language Model Hidden States

Paper · 2026 · Subramanyam Sahoo, Vinija Jain, Aman Chadha et al.

Summary

Interpretability research often takes high linear-probe accuracy on hidden states as a sign that LLMs develop distinct internal encodings for each of the three reasoning types (deduction, induction and abduction). This paper tests whether that conclusion holds once surface format is controlled. The authors build a balanced set of 750 items (250 each) from LogiQA 2.0, ARC-Challenge and αNLI, labelling reasoning mode by benchmark of origin. They run Qwen3-14B with thinking disabled and extract last-input-token hidden states at every layer. Their analysis combines:

  • logistic-regression probes and manifold-geometry measures;
  • a four-step confound pipeline: source probe, option-count probe, 4-choice-only comparison, and Ridge residualisation of format features;
  • matching of reasoning traces to mode "anchor" descriptions via embeddings;
  • activation steering compared against random directions.

Probes separate the three classes perfectly at layer 32. Yet once source, option count and response length are regressed out, accuracy falls to chance. Traces show little mode-specific behaviour, and targeted steering performs no better than random perturbation. The work shows that multi-benchmark probing confuses dataset format with computation, and argues that deconfounding and random-direction controls should become routine.

Key points

  • Setup: Qwen3-14B (40 layers, hidden size 5120, bfloat16) evaluated on 750 tasks. The 2-option αNLI format, versus 4 options for LogiQA 2.0 and ARC-Challenge, is a built-in confound.
  • At layer 32, probes reach 100% 5-fold CV accuracy with clean geometry: TwoNN intrinsic dimensionalities of 20.6, 28.5 and 33.6, and convex hull contamination of at most 1.5%.
  • Regressing format features (source one-hot, option count, response length) out of hidden states drops probe accuracy to chance level.
  • The model solves 86% of tasks, but trace-mode agreement is only 42.5% against a 33.3% chance baseline, suggesting a largely uniform strategy across reasoning types.
  • Steering along centroid-difference directions is not distinguishable from 20 random directions (Laplace-corrected p = 0.286). Acknowledged limitations include a single model and a small steering evaluation (at most 15 tasks); code is 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.

 Institutions

🏢GoogleMountain View · United States430 works · 424,363 citations🏢Google DeepMindLondon · United Kingdom56 works · 78,880 citations🎓Northeastern UniversityBoston · United States57 works · 6,533 citations

Also on the byline: Horizon Research.

 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 3Research subjects and languages the work is about.Language models & pretraining · 3,320 worksReasoning · 1,426 worksInterpretability · 130 works

Cite

Cite as Subramanyam Sahoo, Vinija Jain, Aman Chadha and Divya Chaudhary. 2026. Linear Probes Detect Task Format, Not Reasoning Mode in Language Model Hidden States. Proceedings of the 6th Workshop on Trustworthy NLP (TrustNLP 2026).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.trustnlp-main.12

@inproceedings{sahoo2026linear,
  title = {{Linear Probes Detect Task Format, Not Reasoning Mode in Language Model Hidden States}},
  author = {Subramanyam Sahoo, Vinija Jain, Aman Chadha and Divya Chaudhary},
  booktitle = {Proceedings of the 6th Workshop on Trustworthy NLP (TrustNLP 2026)},
  year = {2026},
  doi = {10.18653/v1/2026.trustnlp-main.12},
  url = {https://doi.org/10.18653/v1/2026.trustnlp-main.12}
}

References

24 references, 14 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Janice Ahn, Rishu Verma, Renze Lou, Di Liu, Rui Zhang, and Wenpeng Yin. 2024. Large language models for mathematical reasoning: Progresses and challenges.
  2. Mubashara Akhtar, Anka Reuel, Prajna Soni, Sanchit Ahuja, Pawan Sasanka Ammanamanchi, Ruchit Rawal, Vilém Zouhar, Srishti Yadav, Chenxi Whitehouse, Dayeon Ki, Jennifer Mickel, Leshem Choshen, Marek Šuppa, Jan Batzner, Jenny Chim, Jeba Sania, Yanan Long, Hossein A. Rahmani, Christina Knight, and 18 o… — [arXiv:2602.16763](https://arxiv.org/abs/2602.16763) find
  3. Elena Facco, Maria d’Errico, Alex Rodriguez, and Alessandro Laio. 2017. Estimating the intrinsic dimension of datasets by a minimal neighborhood information. Scientific Reports, 7(1):12140. find
  4. Guillaume Alain and Yoshua Bengio. 2018. Understanding intermediate layers using linear classifier probes. find
  5. Yonatan Belinkov, Nadir Durrani, Fahim Dalvi, Hassan Sajjad, and James Glass. 2017. What do neural machine translation models learn about morphology? In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 861–872, Vancouver, Canada.…
  6. Luciana Benotti and Patrick Blackburn. 2021. Grounding as a collaborative process. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 515–531, Online. Association for Computational Linguistics. find
  7. Chandra Bhagavatula, Ronan Le Bras, Chaitanya Malaviya, Keisuke Sakaguchi, Ari Holtzman, Hannah Rashkin, Doug Downey, Scott Wen tau Yih, and Yejin Choi. 2020. Abductive commonsense reasoning. Preprint, arXiv:1908.05739.
  8. Simeng Han, Hailey Schoelkopf, Yilun Zhao, Zhenting Qi, Martin Riddell, Wenfei Zhou, James Coady, David Peng, Yujie Qiao, Luke Benson, Lucy Sun, Alex Wardle-Solano, Hannah Szabo, Ekaterina Zubova, Matthew Burtell, Jonathan Fan, Yixin Liu, Brian Wong, Malcolm Sailor, and 16 others. 2024. Folio: Natur…
  9. John Hewitt and Percy Liang. 2019. Designing and interpreting probes with control tasks. 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 2733–2743, Hong Kong, Chi… find
  10. 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, Volume 1 (Long and Short Papers), pages 4129…
  11. Jie Huang and Kevin Chen-Chuan Chang. 2023. Towards reasoning in large language models: A survey.
  12. Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter,… — Language Models are Few-Shot Learners
  13. 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. Preprint, arXiv:1803.05457.
  14. Alexis Conneau, German Kruszewski, Guillaume Lample, Loïc Barrault, and Marco Baroni. 2018. What you can cram into a single $&!\#* vector: Probing sentence embeddings for linguistic properties. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long…
  15. Romain Cosentino and Sarath Shekkizhar. 2024. Reasoning in large language models: A geometric perspective. Preprint, arXiv:2407.02678. — [arXiv:2407.02678](https://arxiv.org/abs/2407.02678) find
  16. Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix,… find
  17. Kenneth Li, Oam Patel, Fernanda Viégas, Hanspeter Pfister, and Martin Wattenberg. 2024. Inference-time intervention: Eliciting truthful answers from a language model.
  18. Hanmeng Liu, Jian Liu, Leyang Cui, Zhiyang Teng, Nan Duan, Ming Zhou, and Yue Zhang. 2023. Logiqa 2.0—an improved dataset for logical reasoning in natural language understanding. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 31:2947–2962. find
  19. Samuel Marks and Max Tegmark. 2023. The geometry of truth: Emergent linear structure in large language model representations of true/false datasets. arXiv preprint arXiv:2310.06824.
  20. Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. 2023. Locating and editing factual associations in gpt. find
  21. Neel Nanda, Lawrence Chan, Tom Lieberum, Jess Smith, and Jacob Steinhardt. 2023. Progress measures for grokking via mechanistic interpretability.
  22. Nathan Young, Qiming Bao, Joshua Bensemann, and Michael Witbrock. 2022. AbductionRules: Training transformers to explain unexpected inputs. pages 218–227. find
  23. Andy Zou, Long Phan, Sarah Chen, James Campbell, Catherine Olsson, Nelson Elhage, Neel Nanda, Phillip Guo, Richard Ren, Alexander Pan, Xuwang Nicholas Joseph, Nova DasSarma, Tom Henighan, Yin, Mantas Mazeika, Ann-Kathrin Dombrowski, Ben Mann, Amanda Askell, Yuntao Bai, Anna Chen, Shashwat Goel, Nath… — In-context Learning and Induction Heads
  24. OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Balcom, Paul Baltescu, Haiming Bao, Mohammad Bavarian, Jeff Belgum, and 262 others.… — GPT-4 Technical Report

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/sahoo-2026-linear-probes-detect-task-format/