---
title: "SteerForce at SemEval-2026 Task 11: Reducing Content Effects"
description: "The paper addresses content effects in SemEval-2026 Task 11 (Subtask 1). 11 references."
authors: ["Noah Tratzsch", "Asmaa Al-Raian", "Mounika Marreddy", "Alexander Mehler"]
year: 2026
venue: "Proceedings of the 20th International Workshop on Semantic Evaluation (2026)"
source: "https://aclanthology.org/2026.semeval-1.143/"
doi: "10.18653/v1/2026.semeval-1.143"
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/tratzsch-2026-steerforce-semeval-2026-task-11/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/tratzsch-2026-steerforce-semeval-2026-task-11/)

[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/50/) / SteerForce at SemEval-2026 Task 11: Reducing Content Effects Using Layered Activation Steering

# SteerForce at SemEval-2026 Task 11: Reducing Content Effects Using Layered Activation Steering

**Authors:** Noah Tratzsch, Asmaa Al-Raian, [Mounika Marreddy](https://md.research.elamur.ai/free/en/authors/mounika-marreddy/) and [Alexander Mehler](https://md.research.elamur.ai/free/en/authors/alexander-mehler/) · **Year:** 2026 · **Venue:** [Proceedings of the 20th International Workshop on Semantic Evaluation (2026)](https://md.research.elamur.ai/free/en/venues/semeval/) ([SemEval 2026](https://md.research.elamur.ai/free/en/venues/semeval/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.semeval-1.143/) · [PDF](https://aclanthology.org/2026.semeval-1.143.pdf) · [DOI 10.18653/v1/2026.semeval-1.143](https://doi.org/10.18653/v1/2026.semeval-1.143) · [OpenAlex](https://openalex.org/W7166845085)

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

## Summary
The paper addresses content effects in SemEval-2026 Task 11 (Subtask 1). In this task, systems judge whether English syllogisms are logically valid, but models tend to be swayed by whether the conclusion sounds believable. The authors treat this bias as a geometric offset between plausibility-driven and validity-driven hidden representations. They correct it at inference time with activation steering rather than retraining.

Their method, K-ACT, has two parts. Activation Transport (ACT) shifts hidden states toward a reference distribution. K-CAST, a kNN-gated contrastive steering method, then applies that shift only to selected inputs. Layers to steer are chosen through a per-layer sensitivity analysis.

They test bert-base-uncased, fine-tuned with validity and plausibility heads, and two instruction-tuned decoders used without fine-tuning: TinyLlama-1.1B-Chat-v1.0 and Qwen2.5-1.5B-Instruct. BERT benefits most from gentle steering spread over several upper layers. The decoders improve most from a single intervention in a late layer, and adding more layers tends to hurt. The authors conclude that steering strategies must be tailored to the model architecture.

## Key points
- **Data and metrics.** The training set has roughly 800 syllogisms, each labeled for validity and plausibility. Systems are scored by validity accuracy. Directional Content Bias is the gap in accuracy between plausibility-aligned and plausibility-conflicting items.
- **BERT.** Sequential steering on layers 8+9 raised accuracy from 0.7708 to 0.8229 and cut bias from 0.0833 to 0.0208, a 75% reduction.
- **Qwen2.5-1.5B-Instruct.** Single-layer K-CAST at layer 21 lowered bias from 0.2611 to 0.0409, with accuracy rising slightly from 0.6771 to 0.6875. Steering layers 19+21 brought bias back up to 0.2599.
- **TinyLlama.** K-ACT at layer 17 improved accuracy from 0.6250 to 0.6667 and reduced bias from 0.1469 to 0.0559. Sequential steering on layers 21+14 dropped accuracy to 0.5104.
- **Layer sensitivity and limitations.** Sensitivity builds gradually toward layers 8–11 in BERT but spikes around layers 15–17 in Qwen. Large K values weaken the correction. The study covers only small models and a modest dataset.

*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

- [Goethe University Frankfurt](https://md.research.elamur.ai/free/en/universities/goethe-university-frankfurt/) — University, Frankfurt am Main, Germany · 18 works · 373 citations

📚 Shelves (4): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 18,614 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 17,965 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 17,450 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 17,296 works

## Cite

Cite as: Noah Tratzsch, Asmaa Al-Raian, Mounika Marreddy and Alexander Mehler. 2026. SteerForce at SemEval-2026 Task 11: Reducing Content Effects Using Layered Activation Steering. Proceedings of the 20th International Workshop on Semantic Evaluation (2026).

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

```bibtex
@inproceedings{tratzsch2026steerforce,
  title = {{SteerForce at SemEval-2026 Task 11: Reducing Content Effects Using Layered Activation Steering}},
  author = {Noah Tratzsch, Asmaa Al-Raian, Mounika Marreddy and Alexander Mehler},
  booktitle = {Proceedings of the 20th International Workshop on Semantic Evaluation (2026)},
  year = {2026},
  doi = {10.18653/v1/2026.semeval-1.143},
  url = {https://doi.org/10.18653/v1/2026.semeval-1.143}
}
```

- [← previous StealthGraph: Exposing Domain-Specific Risks in LLMs through Knowledge-Graph-Guided Harmful Prompt Generation](https://md.research.elamur.ai/free/en/papers/zheng-2026-stealthgraph-exposing-domain-specific-risks/)
- [next → Steering LLMs Responses Towards Moral Foundations on the Norwegian MFQ-30](https://md.research.elamur.ai/free/en/papers/andersen-2026-steering-llms-responses-towards-moral/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/50/)

## References

11 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. 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,…
2. Weixuan Wang, Jingyuan Yang, and Wei Peng. 2025. Semantics-adaptive activation intervention for llms via dynamic steering vectors. Preprint, arXiv:2410.12299. — [arXiv:2410.12299](https://arxiv.org/abs/2410.12299) [find](https://scholar.google.com/scholar?q=Weixuan+Wang%2C+Jingyuan+Yang%2C+and+Wei+Peng%2E+2025%2E+Semantics%2Dadaptive+activation+intervention+for+llms+via+dynamic+steering+vectors%2E+Preprint%2C+arXiv%3A2410%2E12299%2E+%E2%80%94+%5BarXiv%3A2410%2E12299%5D%28https%3A%2F%2Farxiv%2Eorg%2Fab)
3. Peiyuan Zhang, Guangtao Zeng, Tianduo Wang, and Wei Lu. 2024. [Tinyllama: An open-source small language model](https://md.research.elamur.ai/free/en/papers/zhang-2024-tinyllama-open-source-small-language/). Preprint, arXiv:2401.02385.
4. Nina Panickssery, Nick Gabrieli, Julian Schulz, Meg Tong, Evan Hubinger, and Alexander Matt Turner. 2024. Steering llama 2 via contrastive activation addition. Preprint, arXiv:2312.06681. [find](https://scholar.google.com/scholar?q=Nina+Panickssery%2C+Nick+Gabrieli%2C+Julian+Schulz%2C+Meg+Tong%2C+Evan+Hubinger%2C+and+Alexander+Matt+Turner%2E+2024%2E+Steering+llama+2+via+contrastive+activation+addition%2E+Preprint%2C+arXiv%3A2312%2E06681%2E)
5. Qwen, :, An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, and 25 others. 2025. Qwen2.5 technical report. Preprint, arXiv:2412.15115. [find](https://scholar.google.com/scholar?q=Qwen%2C+%3A%2C+An+Yang%2C+Baosong+Yang%2C+Beichen+Zhang%2C+Binyuan+Hui%2C+Bo+Zheng%2C+Bowen+Yu%2C+Chengyuan+Li%2C+Dayiheng+Liu%2C+Fei+Huang%2C+Haoran+Wei%2C+Huan+Lin%2C+Jian+Yang%2C+Jianhong+Tu%2C+Jianwei+Zhang%2C+Jianxin+Yang%2C+Jiaxi+)
6. Nina Rimsky, Nick Gabrieli, Julian Schulz, Meg Tong, Evan Hubinger, and Alexander Turner. 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](https://scholar.google.com/scholar?q=Nina+Rimsky%2C+Nick+Gabrieli%2C+Julian+Schulz%2C+Meg+Tong%2C+Evan+Hubinger%2C+and+Alexander+Turner%2E+2024%2E+Steering+llama+2+via+contrastive+activation+addition%2E+In+Proceedings+of+the+62nd+Annual+Meeting+of+the+A)
7. Pau Rodriguez, Arno Blaas, Michal Klein, Luca Zappella, Nicholas Apostoloff, Marco Cuturi, and Xavier Suau. 2024. Controlling language and diffusion models by transporting activations. Preprint, arXiv:2410.23054. — [arXiv:2410.23054](https://arxiv.org/abs/2410.23054) [find](https://scholar.google.com/scholar?q=Pau+Rodriguez%2C+Arno+Blaas%2C+Michal+Klein%2C+Luca+Zappella%2C+Nicholas+Apostoloff%2C+Marco+Cuturi%2C+and+Xavier+Suau%2E+2024%2E+Controlling+language+and+diffusion+models+by+transporting+activations%2E+Preprint%2C+arXiv)
8. SemEval-2026 Task 11 Organizers. 2026. Semeval-2026 task 11 dataset. [https://github.com/](https://github.com/) neuro-symbolic-ai/semeval_2026_task_11. Training data for Subtask 1. [↗](https://github.com/)
9. Chi Sun, Xipeng Qiu, Yige Xu, and Xuanjing Huang. 2020. [How to fine-tune bert for text classification?](https://md.research.elamur.ai/free/en/papers/sun-2019-fine-tune-bert-text-classification/) Preprint, arXiv:1905.05583.
10. Marco Valentino, Geonhee Kim, Dhairya Dalal, Zhixue Zhao, and André Freitas. 2025. Mitigating content effects on reasoning in language models through fine-grained activation steering. Preprint, arXiv:2505.12189. — [arXiv:2505.12189](https://arxiv.org/abs/2505.12189) [find](https://scholar.google.com/scholar?q=Marco+Valentino%2C+Geonhee+Kim%2C+Dhairya+Dalal%2C+Zhixue+Zhao%2C+and+Andr%C3%A9+Freitas%2E+2025%2E+Mitigating+content+effects+on+reasoning+in+language+models+through+fine%2Dgrained+activation+steering%2E+Preprint%2C+arXiv%3A)
11. Marco Valentino, Leonardo Ranaldi, Giulia Pucci, Federico Ranaldi, and André Freitas. 2026. [Semeval-2026 task 11: Disentangling content and formal reasoning in large language models](https://md.research.elamur.ai/free/en/papers/valentino-2026-semeval-2026-task-11-disentangling/). In Proceedings of

*Updated 1 November 2026*