---
title: "Measuring and Mitigating Shortcut Reliance in Language Models"
description: "The proposed metric, Correlation-Dependent Representational Entanglement (CDRE), is the absolute cosine between the two… 24 references."
authors: ["Divyajot Singh"]
year: 2026
venue: "Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop)"
source: "https://aclanthology.org/2026.acl-srw.59/"
doi: "10.18653/v1/2026.acl-srw.59"
licence: "CC BY 4.0 (ACL Anthology)"
licence_url: "https://creativecommons.org/licenses/by/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/singh-2026-measuring-mitigating-shortcut-reliance-language/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/singh-2026-measuring-mitigating-shortcut-reliance-language/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2026](https://md.research.elamur.ai/free/en/papers/2026/page/36/) / Measuring and Mitigating Shortcut Reliance in Language Models with Probe-Based Representation Entanglement

# Measuring and Mitigating Shortcut Reliance in Language Models with Probe-Based Representation Entanglement

**Authors:** Divyajot Singh · **Year:** 2026 · **Venue:** [Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop)](https://md.research.elamur.ai/free/en/venues/acl/) ([ACL 2026](https://md.research.elamur.ai/free/en/venues/acl/2026/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://aclanthology.org/2026.acl-srw.59/) · [PDF](https://aclanthology.org/2026.acl-srw.59.pdf) · [DOI 10.18653/v1/2026.acl-srw.59](https://doi.org/10.18653/v1/2026.acl-srw.59) · [OpenAlex](https://openalex.org/W7166835376)

**Citations:** 0 citations (OpenAlex, 2026-10-14)

## Summary
The paper asks how to tell when a fine-tuned language model has truly come to depend on a spurious cue, and whether a mitigation step is worth applying. The author fine-tunes two small models, meta-llama/Llama-3.2-1B and google/gemma-3-1b-it, as binary classifiers in three biased settings. Two are artificial cues injected into SST-2: capitalization and speaker prefixes. The third is the naturally occurring lexical-overlap heuristic in MNLI, recast as entailment vs. non-entailment. Shortcut ratios are varied across runs.

Throughout training, layer-wise logistic-regression probes are fit for the task label and for the shortcut attribute. The proposed metric, Correlation-Dependent Representational Entanglement (CDRE), is the absolute cosine between the two probe weight vectors. As shortcut prevalence grows, the gap between shortcut-aligned and shortcut-free accuracy widens, and deep-layer CDRE climbs alongside it.

A training penalty based on CDRE restores much of the lost robustness for capitalization and overlap. It does little for speaker prefixes, where the two directions stay close to orthogonal even though the shortcut is easy to decode. The practical upshot is a cheap diagnostic that signals in advance whether debiasing is likely to pay off.

## Key points
- **Effect of a 99% capitalization bias on clean SST-2 accuracy:**
  - Llama falls from 93.2% to 49.0%.
  - Gemma falls from 91.8% to 60.2%.
- **CDRE regularization on capitalization (99% prevalence):**
  - OOD accuracy rises by 10 points for both models (Llama 69.6→79.6, Gemma 61.2→71.2).
  - JTT gives no comparable gain.
- **CDRE regularization on MNLI lexical overlap:**
  - The robustness gap shrinks from 15.0 to 4.8 (Llama) and from 13.0 to 2.6 (Gemma), removing roughly two-thirds to four-fifths of it.
  - OOD accuracy reaches about 90–91%.
- **Speaker-prefix control:** OOD accuracy is essentially unchanged under CDRE regularization (about 86–88%). This shows that a shortcut being linearly recoverable does not, by itself, predict failure.
- **Experimental setup:**
  - SST-2 probing runs use 2,000 training examples and a balanced 100-example probe set.
  - MNLI runs use 10,000 training examples and a 500-example probe set.
  - Probes are refit every 10% of training on mean-pooled hidden states.

*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

- [Birla Institute of Technology and Science, Pilani](https://md.research.elamur.ai/free/en/universities/birla-institute-of-technology-and-science-pilani/) — University, Pilāni, India · 16 works · 41,879 citations

📚 Shelves (4): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 16,667 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 16,194 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 16,039 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 15,891 works

🧭 Topics (1): [🧠 Language models & pretraining](https://md.research.elamur.ai/free/en/topics/language-models/) · 3,320 works

## Cite

Cite as: Divyajot Singh. 2026. Measuring and Mitigating Shortcut Reliance in Language Models with Probe-Based Representation Entanglement. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop).

The paper itself, not this page. BibTeX from the library's record of the work.

```bibtex
@inproceedings{singh2026measuring,
  title = {{Measuring and Mitigating Shortcut Reliance in Language Models with Probe-Based Representation Entanglement}},
  author = {Divyajot Singh},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-srw.59},
  url = {https://doi.org/10.18653/v1/2026.acl-srw.59}
}
```

- [← previous Measure Twice, Click Once: Co-evolving Proposer and Visual Critic via Reinforcement Learning for GUI Grounding](https://md.research.elamur.ai/free/en/papers/wang-2026-measure-twice-click-once-co/)
- [next → Measuring Digital Labour Market Transitions with a Digital Semantic Score: An AI-Based Methodology Applied to the Dutch Labour Market](https://md.research.elamur.ai/free/en/papers/shahmohammadi-2026-measuring-digital-labour-market-transitions/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/36/)

## References

24 references, 16 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Guillaume Alain and Yoshua Bengio. 2017. Understanding intermediate layers using linear classifier probes. arXiv preprint arXiv:1610.01644. [find](https://scholar.google.com/scholar?q=Guillaume+Alain+and+Yoshua+Bengio%2E+2017%2E+Understanding+intermediate+layers+using+linear+classifier+probes%2E+arXiv+preprint+arXiv%3A1610%2E01644%2E)
2. Yonatan Belinkov. 2022. [Probing classifiers: Promises, shortcomings, and advances](https://md.research.elamur.ai/free/en/papers/belinkov-2022-probing-classifiers-promises-shortcomings-advances/). Computational Linguistics, 48(1):207–219.
3. Samuel R. Bowman, Gabor Angeli, Christopher Potts, and Christopher D. Manning. 2015. [A large annotated corpus for learning natural language inference](https://md.research.elamur.ai/free/en/papers/bowman-2015-large-annotated-corpus-learning-natural/). In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 632–642.
4. Christopher Clark, Mark Yatskar, and Luke Zettlemoyer. 2019. Don’t take the easy way out: Ensemble based methods for avoiding known dataset biases. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language… [find](https://scholar.google.com/scholar?q=Christopher+Clark%2C+Mark+Yatskar%2C+and+Luke+Zettlemoyer%2E+2019%2E+Don%E2%80%99t+take+the+easy+way+out%3A+Ensemble+based+methods+for+avoiding+known+dataset+biases%2E+In+Proceedings+of+the+2019+Conference+on+Empirical+M)
5. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. [Bert: Pre-training of deep bidirectional transformers for language understanding](https://md.research.elamur.ai/free/en/papers/devlin-2019-bert-pre-training-deep-bidirectional/). In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies,…
6. Yanai Elazar and Yoav Goldberg. 2018. [Adversarial removal of demographic attributes from text data](https://md.research.elamur.ai/free/en/papers/elazar-2018-adversarial-removal-demographic-attributes-text/). In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 11– 21.
7. Allyson Ettinger. 2020. [What bert is not: Lessons from a new suite of psycholinguistic diagnostics for language models](https://md.research.elamur.ai/free/en/papers/ettinger-2020-what-bert-is-not-lessons/). Transactions of the Association for Computational Linguistics, 8:34–48.
8. Matt Gardner, Yoav Artzi, Victoria Basmov, Jonathan Berant, Ben Bogin, Sihao Chen, Pradeep Dasigi, Dheeru Dua, Yanai Elazar, Ananth Gottumukkala, Nitish Gupta, Hannaneh Hajishirzi, Gabriel Ilharco, Daniel Khashabi, Kevin Lin, Jiangming Liu, Nelson F. Liu, Phoebe Mulcaire, Qiang Ning, and 7 others. 2… — [Evaluating Models' Local Decision Boundaries via Contrast Sets](https://md.research.elamur.ai/free/en/papers/gardner-2020-evaluating-models-local-decision-boundaries/)
9. Robert Geirhos, J"orn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A. Wichmann. 2020. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2(11):665–673. [find](https://scholar.google.com/scholar?q=Robert+Geirhos%2C+J%22orn%2DHenrik+Jacobsen%2C+Claudio+Michaelis%2C+Richard+Zemel%2C+Wieland+Brendel%2C+Matthias+Bethge%2C+and+Felix+A%2E+Wichmann%2E+2020%2E+Shortcut+learning+in+deep+neural+networks%2E+Nature+Machine+Intell)
10. Suchin Gururangan, Swabha Swayamdipta, Omer Levy, Roy Schwartz, Samuel Bowman, and Noah A. Smith. 2018. [Annotation artifacts in natural language inference data](https://md.research.elamur.ai/free/en/papers/gururangan-2018-annotation-artifacts-natural-language-inference/). In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Techno…
11. 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, pages 2733–2743. [find](https://scholar.google.com/scholar?q=John+Hewitt+and+Percy+Liang%2E+2019%2E+Designing+and+interpreting+probes+with+control+tasks%2E+In+Proceedings+of+the+2019+Conference+on+Empirical+Methods+in+Natural+Language+Processing+and+the+9th+Internati)
12. Divyansh Kaushik, Eduard Hovy, and Zachary C. Lipton. 2020. Learning the difference that makes a difference with counterfactually-augmented data. In International Conference on Learning Representations. [find](https://scholar.google.com/scholar?q=Divyansh+Kaushik%2C+Eduard+Hovy%2C+and+Zachary+C%2E+Lipton%2E+2020%2E+Learning+the+difference+that+makes+a+difference+with+counterfactually%2Daugmented+data%2E+In+International+Conference+on+Learning+Representation)
13. Evan Z. Liu, Behzad Haghgoo, Annie S. Chen, Aditi Raghunathan, Pang Wei Koh, Shiori Sagawa, Percy Liang, and Chelsea Finn. 2021. Just train twice: Improving group robustness without training group information. Proceedings of the 38th International Conference on Machine Learning. [find](https://scholar.google.com/scholar?q=Evan+Z%2E+Liu%2C+Behzad+Haghgoo%2C+Annie+S%2E+Chen%2C+Aditi+Raghunathan%2C+Pang+Wei+Koh%2C+Shiori+Sagawa%2C+Percy+Liang%2C+and+Chelsea+Finn%2E+2021%2E+Just+train+twice%3A+Improving+group+robustness+without+training+group+inf)
14. Tianyu Liu, Zheng Xin, Xiaoan Ding, Baobao Chang, and Zhifang Sui. 2020. [An empirical study on model-agnostic debiasing strategies for robust natural language inference](https://md.research.elamur.ai/free/en/papers/liu-2020-empirical-study-model-agnostic-debiasing/). In Proceedings of the 24th Conference on Computational Natural Language Learning, pages 596–608.
15. R. Thomas McCoy, Ellie Pavlick, and Tal Linzen. 2019. [Right for the wrong reasons: Diagnosing syntactic heuristics in natural language inference](https://md.research.elamur.ai/free/en/papers/mccoy-2019-right-wrong-reasons-diagnosing-syntactic/). In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3428–3448.
16. Tal Schuster, Darsh Shah, Yun Jie Serene Yeo, Daniel Roberto Filizzola Ortiz, Enrico Santus, and Regina Barzilay. 2019. [Towards debiasing fact verification models](https://md.research.elamur.ai/free/en/papers/schuster-2019-debiasing-fact-verification-models/). In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on…
17. Prasetya Ajie Utama, Nafise Sadat Moosavi, and Iryna Gurevych. 2020. Towards debiasing nlu models from unknown biases. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 7597–7610. [find](https://scholar.google.com/scholar?q=Prasetya+Ajie+Utama%2C+Nafise+Sadat+Moosavi%2C+and+Iryna+Gurevych%2E+2020%2E+Towards+debiasing+nlu+models+from+unknown+biases%2E+In+Proceedings+of+the+2020+Conference+on+Empirical+Methods+in+Natural+Language+Pr)
18. Elena Voita and Ivan Titov. 2020. [Information-theoretic probing with minimum description length](https://md.research.elamur.ai/free/en/papers/voita-2020-information-theoretic-probing-minimum-description/). In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 183–196.
19. Adina Williams, Nikita Nangia, and Samuel R. Bowman. 2018. [A broad-coverage challenge corpus for sentence understanding through inference](https://md.research.elamur.ai/free/en/papers/williams-2018-broad-coverage-challenge-corpus-sentence/). In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long…
20. Xiang Zhou and Mohit Bansal. 2020. [Towards robustifying nli models against lexical dataset biases](https://md.research.elamur.ai/free/en/papers/zhou-2020-robustifying-nli-models-against-lexical/). In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8759– 8771.
21. Yixin Nie, Adina Williams, Emily Dinan, Mohit Bansal, Jason Weston, and Douwe Kiela. 2020. [Adversarial nli: A new benchmark for natural language understanding](https://md.research.elamur.ai/free/en/papers/nie-2020-adversarial-nli-new-benchmark-natural/). In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4885–4901.
22. Matthew E. Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. 2018. [Deep contextualized word representations](https://md.research.elamur.ai/free/en/papers/peters-2018-deep-contextualized-word-representations/). In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Techno…
23. Adam Poliak, Jason Naradowsky, Aparajita Haldar, Rachel Rudinger, and Benjamin Van Durme. 2018. [Hypothesis only baselines in natural language inference](https://md.research.elamur.ai/free/en/papers/poliak-2018-hypothesis-only-baselines-natural-language/). In Proceedings of the Seventh Joint Conference on Lexical and Computational Semantics, pages 180– 191.
24. Shiori Sagawa, Pang Wei Koh, Tatsunori B. Hashimoto, and Percy Liang. 2020. Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. arXiv preprint arXiv:1911.08731. [find](https://scholar.google.com/scholar?q=Shiori+Sagawa%2C+Pang+Wei+Koh%2C+Tatsunori+B%2E+Hashimoto%2C+and+Percy+Liang%2E+2020%2E+Distributionally+robust+neural+networks+for+group+shifts%3A+On+the+importance+of+regularization+for+worst%2Dcase+generalization%2E)

*Updated 1 November 2026*