---
title: "Disentangling the Effects of Unlearning in Measuring Parametric"
description: "Tutek et al. (2025) proposed parametric faithfulness, which tests whether a Chain-of-Thought (CoT) reflects a model's internal… 14 references."
authors: ["Ryo Mitsuhashi", "Gaku Morio", "Ayana Niwa", "Masahiro Kaneko", "Kentaro Inui", "Terufumi Morishita", "Yuta Koreeda", "Yasuhiro Sogawa"]
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.36/"
doi: "10.18653/v1/2026.acl-srw.36"
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/mitsuhashi-2026-disentangling-effects-unlearning-measuring-parametric/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/mitsuhashi-2026-disentangling-effects-unlearning-measuring-parametric/)

[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/27/) / Disentangling the Effects of Unlearning in Measuring Parametric Faithfulness of Chain-of-Thought

# Disentangling the Effects of Unlearning in Measuring Parametric Faithfulness of Chain-of-Thought

**Authors:** Ryo Mitsuhashi, [Gaku Morio](https://md.research.elamur.ai/free/en/authors/gaku-morio/), [Ayana Niwa](https://md.research.elamur.ai/free/en/authors/ayana-niwa/), [Masahiro Kaneko](https://md.research.elamur.ai/free/en/authors/masahiro-kaneko/), [Kentaro Inui](https://md.research.elamur.ai/free/en/authors/kentaro-inui/), [Terufumi Morishita](https://md.research.elamur.ai/free/en/authors/terufumi-morishita/), [Yuta Koreeda](https://md.research.elamur.ai/free/en/authors/yuta-koreeda/) and [Yasuhiro Sogawa](https://md.research.elamur.ai/free/en/authors/yasuhiro-sogawa/) · **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.36/) · [PDF](https://aclanthology.org/2026.acl-srw.36.pdf) · [DOI 10.18653/v1/2026.acl-srw.36](https://doi.org/10.18653/v1/2026.acl-srw.36) · [OpenAlex](https://openalex.org/W7166842407)

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

## Summary
Tutek et al. (2025) proposed parametric faithfulness, which tests whether a Chain-of-Thought (CoT) reflects a model's internal reasoning. The method unlearns individual CoT steps and checks whether the model's No-CoT answer changes. That design attributes every answer change to the removed reasoning step. This paper argues the assumption is too strong, because unlearning can disturb a model in ways unrelated to the target knowledge.

To separate these side effects, the authors add two control tasks:
- **Irrelevant Step (IS):** the target step is swapped for an unrelated CoT sentence of similar length before unlearning.
- **Word Shuffle (WS):** the target step's words are scrambled, so the vocabulary stays but the meaning is lost.

Subtracting each control score from the original FF gives two corrected metrics, ∆IS and ∆WS.

The experiments use LLaMA-3 models (3B, 8B, 70B) on four QA benchmarks. The control tasks alone trigger many answer changes, so the original FF overstates faithfulness. Much of the reported growth in FF with longer CoTs and with the smaller model also turns out to be an unlearning artifact. The work therefore matters for anyone using parameter interventions to judge whether CoT explanations can be trusted.

## Key points
- **Setup:** the study uses LLaMA-3 8B and 70B plus LLaMA-3.2 3B. The benchmarks are OpenBookQA, ARC-Challenge, StrategyQA, and the BigBench-Hard sports subtask. It reuses the original codebase, where unlearning uses an NPO loss with a KL retain term and updates only the second MLP matrix (FF2).
- **Main result (8B, averaged over datasets):** the original FF is 36.3%, FF_IS is 22.1% and FF_WS is 24.7%. This gives ∆IS = 14.2 and ∆WS = 11.6.
- **Dataset differences:** the corrected scores are larger on OpenBookQA and smaller on sports and StrategyQA, so domain differences remain even after correction.
- **Sanity checks:** the efficacy of IS is only about 11% of the original method's, and specificity is mostly above 95%. The authors conclude the controls neither erase the target knowledge nor damage unrelated samples.
- **Trends:**
  - FF rises with the number of CoT steps, but the controls rise in a similar way.
  - The 3B model's high FF is largely an unlearning effect.
  - For the largest (70B) model, both corrected metrics, ∆IS and ∆WS, shrink considerably. The authors tentatively explain this by larger models having alternative reasoning paths, and leave verification to future work.

*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

- [Princeton University](https://md.research.elamur.ai/free/en/universities/princeton-university/) — University, Princeton, United States · 90 works · 39,109 citations
- [Hitachi](https://md.research.elamur.ai/free/en/companies/hitachi/) — Company, Tokyo, Japan · 4 works · 69 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): [🧮 Reasoning](https://md.research.elamur.ai/free/en/topics/reasoning/) · 1,426 works

## Cite

Cite as: Ryo Mitsuhashi, Gaku Morio, Ayana Niwa, Masahiro Kaneko, Kentaro Inui, Terufumi Morishita, Yuta Koreeda and Yasuhiro Sogawa. 2026. Disentangling the Effects of Unlearning in Measuring Parametric Faithfulness of Chain-of-Thought. 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{mitsuhashi2026disentangling,
  title = {{Disentangling the Effects of Unlearning in Measuring Parametric Faithfulness of Chain-of-Thought}},
  author = {Ryo Mitsuhashi, Gaku Morio, Ayana Niwa, Masahiro Kaneko, Kentaro Inui, Terufumi Morishita, Yuta Koreeda and Yasuhiro Sogawa},
  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.36},
  url = {https://doi.org/10.18653/v1/2026.acl-srw.36}
}
```

- [← previous Disentangling Linguistic Relatedness from Task Alignment in Cross-Lingual Transfer](https://md.research.elamur.ai/free/en/papers/ahmed-2026-disentangling-linguistic-relatedness-task-alignment/)
- [next → Disinformation between Knowledge and Ignorance. an Epistemological Comparison](https://md.research.elamur.ai/free/en/papers/lizzadri-2026-disinformation-between-knowledge-ignorance-epistemological/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/27/)

## References

14 references, 8 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. James Chua and Owain Evans. 2025. Are DeepSeek R1 and other reasoning models more faithful? In ICLR 2025 Workshop on Foundation Models in the Wild. [find](https://scholar.google.com/scholar?q=James+Chua+and+Owain+Evans%2E+2025%2E+Are+DeepSeek+R1+and+other+reasoning+models+more+faithful%3F+In+ICLR+2025+Workshop+on+Foundation+Models+in+the+Wild%2E)
2. 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](https://md.research.elamur.ai/free/en/papers/clark-2018-think-have-solved-question-answering/). Preprint, arXiv:1803.05457.
3. Mor Geva, Daniel Khashabi, Elad Segal, Tushar Khot, Dan Roth, and Jonathan Berant. 2021. Did aristotle use a laptop? a question answering benchmark with implicit reasoning strategies. Transactions of the Association for Computational Linguistics, 9:346– 361. [find](https://scholar.google.com/scholar?q=Mor+Geva%2C+Daniel+Khashabi%2C+Elad+Segal%2C+Tushar+Khot%2C+Dan+Roth%2C+and+Jonathan+Berant%2E+2021%2E+Did+aristotle+use+a+laptop%3F+a+question+answering+benchmark+with+implicit+reasoning+strategies%2E+Transactions+of+)
4. 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](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)
5. Yihuai Hong, Yuelin Zou, Lijie Hu, Ziqian Zeng, Di Wang, and Haiqin Yang. 2024. Dissecting fine-tuning unlearning in large language models. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 3933– 3941, Miami, Florida, USA. Association for Computational… [find](https://scholar.google.com/scholar?q=Yihuai+Hong%2C+Yuelin+Zou%2C+Lijie+Hu%2C+Ziqian+Zeng%2C+Di+Wang%2C+and+Haiqin+Yang%2E+2024%2E+Dissecting+fine%2Dtuning+unlearning+in+large+language+models%2E+In+Proceedings+of+the+2024+Conference+on+Empirical+Methods+i)
6. Takeshi Kojima, Shixiang (Shane) Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. [Large language models are zero-shot reasoners](https://md.research.elamur.ai/free/en/papers/kojima-2022-large-language-models-zero-shot/). In Advances in Neural Information Processing Systems, volume 35, pages 22199–22213. Curran Associates, Inc.
7. Tamera Lanham, Anna Chen, Ansh Radhakrishnan, Benoit Steiner, Carson Denison, Danny Hernandez, Dustin Li, Esin Durmus, Evan Hubinger, Jackson Kernion, Kamilė Lukošiūtė, Karina Nguyen, Newton Cheng, Nicholas Joseph, Nicholas Schiefer, Oliver Rausch, Robin Larson, Sam McCandlish, Sandipan Kundu, an… — [Measuring Faithfulness in Chain-of-Thought Reasoning](https://md.research.elamur.ai/free/en/papers/lanham-2023-measuring-faithfulness-chain-thought-reasoning/)
8. Chin-Yew Lin. 2004. [ROUGE: A package for automatic evaluation of summaries](https://md.research.elamur.ai/free/en/papers/lin-2004-rouge-package-automatic-evaluation-summaries/). In Text Summarization Branches Out, pages 74–81, Barcelona, Spain. Association for Computational Linguistics.
9. Todor Mihaylov, Peter Clark, Tushar Khot, and Ashish Sabharwal. 2018. Can a suit of armor conduct electricity? a new dataset for open book question answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2381–2391, Brussels, Belgium. Association fo… [find](https://scholar.google.com/scholar?q=Todor+Mihaylov%2C+Peter+Clark%2C+Tushar+Khot%2C+and+Ashish+Sabharwal%2E+2018%2E+Can+a+suit+of+armor+conduct+electricity%3F+a+new+dataset+for+open+book+question+answering%2E+In+Proceedings+of+the+2018+Conference+on+)
10. Ayana Niwa, Masahiro Kaneko, and Kentaro Inui. 2025. Rectifying belief space via unlearning to harness LLMs’ reasoning. In Findings of the Association for Computational Linguistics: ACL 2025, pages 25060–25075, Vienna, Austria. Association for Computational Linguistics. steps. In Proceedings of the… [find](https://scholar.google.com/scholar?q=Ayana+Niwa%2C+Masahiro+Kaneko%2C+and+Kentaro+Inui%2E+2025%2E+Rectifying+belief+space+via+unlearning+to+harness+LLMs%E2%80%99+reasoning%2E+In+Findings+of+the+Association+for+Computational+Linguistics%3A+ACL+2025%2C+pages+25)
11. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Ł ukasz Kaiser, and Illia Polosukhin. 2017. [Attention is all you need](https://md.research.elamur.ai/free/en/papers/vaswani-2017-attention-all-need/). In Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc.
12. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian ichter, Fei Xia, Ed Chi, Quoc V Le, and Denny Zhou. 2022. [Chain-of-thought prompting elicits reasoning in large language models](https://md.research.elamur.ai/free/en/papers/wei-2022-chain-thought-prompting-elicits-reasoning/). 35:24824– 24837.
13. Yuanshun Yao, Xiaojun Xu, and YangLiu. 2024. [Large language model unlearning](https://md.research.elamur.ai/free/en/papers/yao-2023-large-language-model-unlearning/). In Advances in Neural Information Processing Systems, volume 37, pages 105425–105475. Curran Associates, Inc.
14. Ruiqi Zhang, Licong Lin, Yu Bai, and Song Mei. 2024. [Negative preference optimization: From catastrophic collapse to effective unlearning](https://md.research.elamur.ai/free/en/papers/zhang-2024-negative-preference-optimization-catastrophic-collapse/). In First Conference on Language Modeling.

*Updated 1 November 2026*