---
title: "The Unintended Trade-off of AI Alignment: Balancing"
description: "The paper studies a side effect of hallucination mitigation in LLMs. 56.15% accuracy. 48 references."
authors: ["Omar Mahmoud", "Ali Khalil", "Thommen George Karimpanal", "Buddhika Laknath Semage", "Santu Rana"]
year: 2026
venue: "Findings of the Association for Computational Linguistics: EACL 2026"
source: "https://aclanthology.org/2026.findings-eacl.53/"
doi: "10.18653/v1/2026.findings-eacl.53"
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/mahmoud-2026-unintended-trade-off-ai-alignment/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/mahmoud-2026-unintended-trade-off-ai-alignment/)

[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/9/) / The Unintended Trade-off of AI Alignment: Balancing Hallucination Mitigation and Safety in LLMs

# The Unintended Trade-off of AI Alignment: Balancing Hallucination Mitigation and Safety in LLMs

**Authors:** Omar Mahmoud, Ali Khalil, Thommen George Karimpanal, Buddhika Laknath Semage and Santu Rana · **Year:** 2026 · **Venue:** [Findings of the Association for Computational Linguistics: EACL 2026](https://md.research.elamur.ai/free/en/venues/findings/) ([Findings of the Association for Computational Linguistics 2026](https://md.research.elamur.ai/free/en/venues/findings/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.findings-eacl.53/) · [PDF](https://aclanthology.org/2026.findings-eacl.53.pdf) · [DOI 10.18653/v1/2026.findings-eacl.53](https://doi.org/10.18653/v1/2026.findings-eacl.53) · [OpenAlex](https://openalex.org/W7140141059)

**Citations:** 2 citations (OpenAlex, 2026-10-14)

## Summary
The paper studies a side effect of hallucination mitigation in LLMs. Making a model more truthful can weaken its safety alignment, so it becomes more willing to answer harmful prompts. The authors applied two truthfulness-boosting interventions, ITI and TruthX. These raised accuracy on TruthfulQA but also raised jailbreak susceptibility on AdvBench and StrongReject.

Mechanistic analysis points to a subset of attention heads whose activations mix refusal and hallucination signals. Edits or fine-tuning that touch these heads therefore shift both behaviors together. Even benign fine-tuning can erode refusals for this reason.

As a fix, the authors train sparse autoencoders (SAEs) on head activations to separate refusal directions from hallucination directions. During fine-tuning, subspace orthogonalization keeps updates away from the refusal subspace. On LLaMA-3-8B-Instruct, with additional results on Qwen2.5-7B-Instruct, the method keeps attack success near zero while largely retaining task accuracy. It beats safety-oriented baselines that either lose utility or let harmfulness rise. The work matters because it shows that safety and factuality are coupled inside the model. Alignment work therefore needs to manage both objectives together.

## Key points
- **Trade-off found:** ITI and TruthX improve TruthfulQA factuality but increase jailbreak success on AdvBench and StrongReject. The cause is attention heads that encode refusal and hallucination features in an entangled way.
- **Method:** The approach combines SAE-based disentangling of refusal and hallucination features at the head level with orthogonal fine-tuning that leaves the refusal subspace untouched. It is compared against SafeLoRA, SaLoRA and SAP, plus vanilla and LoRA SFT. Code is on GitHub (OmarMohammed88/Hall_Refusal).
- **Poisoned fine-tuning:** The training data adds 10% harmful instruction–response pairs from the Circuit Break dataset. On LLaMA-3-8B-Instruct commonsense tasks (Table 3), the method reaches 75.09% average accuracy, with 0.58% ASR on AdvBench and 0.00% on StrongReject. The paper describes the vanilla SFT comparison as 56.15% accuracy and 9.23% / 9.90% ASR.
- **TruthfulQA (LLaMA-3-8B):** The method scores MC1/MC2 of 36.5/53.3 on the benign dataset and 37.8/56.1 on the poisoned dataset. The base model scores 34.8/52.8.
- **Baseline weaknesses:**
  - SAP gets low ASR but loses much of its reasoning utility.
  - SaLoRA degrades accuracy heavily.
  - Limitations: the method depends on a trained SAE, which may hinder large-scale use. Only specific model components were analyzed.

*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

- [Deakin University](https://md.research.elamur.ai/free/en/universities/deakin-university/) — University, Burwood, Australia · 4 works · 977 citations

📚 Shelves (6): [🖥️ 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 · [🤝 Social Sciences](https://md.research.elamur.ai/free/en/shelves/social-sciences/) · 1,772 works · [🤝 Safety Research](https://md.research.elamur.ai/free/en/shelves/social-sciences/safety-research/) · 543 works

🧭 Topics (2): [🧠 Language models & pretraining](https://md.research.elamur.ai/free/en/topics/language-models/) · 3,320 works · [🛡️ Bias, fairness & safety](https://md.research.elamur.ai/free/en/topics/bias-fairness-safety/) · 517 works

## Cite

Cite as: Omar Mahmoud, Ali Khalil, Thommen George Karimpanal, Buddhika Laknath Semage and Santu Rana. 2026. The Unintended Trade-off of AI Alignment: Balancing Hallucination Mitigation and Safety in LLMs. Findings of the Association for Computational Linguistics: EACL 2026.

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

```bibtex
@inproceedings{mahmoud2026unintended,
  title = {{The Unintended Trade-off of AI Alignment: Balancing Hallucination Mitigation and Safety in LLMs}},
  author = {Omar Mahmoud, Ali Khalil, Thommen George Karimpanal, Buddhika Laknath Semage and Santu Rana},
  booktitle = {Findings of the Association for Computational Linguistics: EACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-eacl.53},
  url = {https://doi.org/10.18653/v1/2026.findings-eacl.53}
}
```

- [← previous The Riddle of Reflection: Evaluating Reasoning and Self-Awareness in Multilingual LLMs Using Indian Riddles](https://md.research.elamur.ai/free/en/papers/m-2026-riddle-reflection-evaluating-reasoning-self/)
- [next → The Word and the Way: Strategies for Domain-Specific BERT Pre-Training in German Medical NLP](https://md.research.elamur.ai/free/en/papers/he-2026-word-way-strategies-domain-specific/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/9/)

## References

48 references, 21 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Andy Arditi, Oscar Obeso, Aaquib Syed, Daniel Paleka, Nina Panickssery, Wes Gurnee, and Neel Nanda. 2024. [Refusal in language models is mediated by a single direction](https://md.research.elamur.ai/free/en/papers/arditi-2024-refusal-language-models-mediated-single/). Preprint, arXiv:2406.11717.
2. Amos Azaria and Tom Mitchell. 2023. [The internal state of an llm knows when it’s lying](https://md.research.elamur.ai/free/en/papers/azaria-2023-internal-state-llm-knows-when/). arXiv preprint arXiv:2304.13734.
3. Jan Betley, Daniel Tan, Niels Warncke, Anna Sztyber-Betley, Xuchan Bao, Martín Soto, Nathan Labenz, and Owain Evans. 2025. Emergent misalignment: Narrow finetuning can produce broadly misaligned llms. arXiv preprint arXiv:2502.17424. — [arXiv:2502.17424](https://arxiv.org/abs/2502.17424) [find](https://scholar.google.com/scholar?q=Jan+Betley%2C+Daniel+Tan%2C+Niels+Warncke%2C+Anna+Sztyber%2DBetley%2C+Xuchan+Bao%2C+Mart%C3%ADn+Soto%2C+Nathan+Labenz%2C+and+Owain+Evans%2E+2025%2E+Emergent+misalignment%3A+Narrow+finetuning+can+produce+broadly+misaligned+llms%2E)
4. Federico Bianchi, Mirac Suzgun, Giuseppe Attanasio, Paul Röttger, Dan Jurafsky, Tatsunori Hashimoto, and James Zou. 2023. [Safety-tuned llamas: Lessons from improving the safety of large language models that follow instructions](https://md.research.elamur.ai/free/en/papers/bianchi-2023-safety-tuned-llamas-lessons-improving/). arXiv preprint arXiv:2309.07875.
5. Helena Casademunt, Caden Juang, Adam Karvonen, Samuel Marks, Senthooran Rajamanoharan, and Neel Nanda. 2025. Steering out-of-distribution generalization with concept ablation fine-tuning. arXiv preprint arXiv:2507.16795. — [arXiv:2507.16795](https://arxiv.org/abs/2507.16795) [find](https://scholar.google.com/scholar?q=Helena+Casademunt%2C+Caden+Juang%2C+Adam+Karvonen%2C+Samuel+Marks%2C+Senthooran+Rajamanoharan%2C+and+Neel+Nanda%2E+2025%2E+Steering+out%2Dof%2Ddistribution+generalization+with+concept+ablation+fine%2Dtuning%2E+arXiv+prepri)
6. Chao Chen, Kai Liu, Ze Chen, Yi Gu, Yue Wu, Mingyuan Tao, Zhihang Fu, and Jieping Ye. 2024. Inside: Llms’ internal states retain the power of hallucination detection. arXiv preprint arXiv:2402.03744. [find](https://scholar.google.com/scholar?q=Chao+Chen%2C+Kai+Liu%2C+Ze+Chen%2C+Yi+Gu%2C+Yue+Wu%2C+Mingyuan+Tao%2C+Zhihang+Fu%2C+and+Jieping+Ye%2E+2024%2E+Inside%3A+Llms%E2%80%99+internal+states+retain+the+power+of+hallucination+detection%2E+arXiv+preprint+arXiv%3A2402%2E03744%2E)
7. 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.
8. Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, and 1 others. 2024. [The llama 3 herd of models](https://md.research.elamur.ai/free/en/papers/grattafiori-2024-llama-3-herd-models/). arXiv e-prints, pages arXiv–2407.
9. Nelson Elhage, Tristan Hume, Catherine Olsson, Nicholas Schiefer, Tom Henighan, Shauna Kravec, Zac Hatfield-Dodds, Robert Lasenby, Dawn Drain, Carol Chen, Roger Grosse, Sam McCandlish, Jared Kaplan, Dario Amodei, Martin Wattenberg, and Christopher Olah. 2022. Toy models of superposition. Transformer… [find](https://scholar.google.com/scholar?q=Nelson+Elhage%2C+Tristan+Hume%2C+Catherine+Olsson%2C+Nicholas+Schiefer%2C+Tom+Henighan%2C+Shauna+Kravec%2C+Zac+Hatfield%2DDodds%2C+Robert+Lasenby%2C+Dawn+Drain%2C+Carol+Chen%2C+Roger+Grosse%2C+Sam+McCandlish%2C+Jared+Kaplan%2C+D)
10. Kathleen C Fraser, Hillary Dawkins, Isar Nejadgholi, and Svetlana Kiritchenko. 2025. Fine-tuning lowers safety and disrupts evaluation consistency. arXiv preprint arXiv:2506.17209. [find](https://scholar.google.com/scholar?q=Kathleen+C+Fraser%2C+Hillary+Dawkins%2C+Isar+Nejadgholi%2C+and+Svetlana+Kiritchenko%2E+2025%2E+Fine%2Dtuning+lowers+safety+and+disrupts+evaluation+consistency%2E+arXiv+preprint+arXiv%3A2506%2E17209%2E)
11. Chia-Yi Hsu, Yu-Lin Tsai, Chih-Hsun Lin, Pin-Yu Chen, Chia-Mu Yu, and Chun-Ying Huang. 2024. Safe lora: The silver lining of reducing safety risks when finetuning large language models. Advances in Neural Information Processing Systems, 37:65072– 65094. [find](https://scholar.google.com/scholar?q=Chia%2DYi+Hsu%2C+Yu%2DLin+Tsai%2C+Chih%2DHsun+Lin%2C+Pin%2DYu+Chen%2C+Chia%2DMu+Yu%2C+and+Chun%2DYing+Huang%2E+2024%2E+Safe+lora%3A+The+silver+lining+of+reducing+safety+risks+when+finetuning+large+language+models%2E+Advances+in+Ne)
12. Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, and 1 others. 2025. [A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions](https://md.research.elamur.ai/free/en/papers/huang-2023-survey-hallucination-large-language-models/). ACM Transactions on Information Systems,…
13. Tiansheng Huang, Sihao Hu, Fatih Ilhan, Selim Furkan Tekin, and Ling Liu. 2024. Booster: Tackling harmful fine-tuning for large language models via attenuating harmful perturbation. arXiv preprint arXiv:2409.01586. — [arXiv:2409.01586](https://arxiv.org/abs/2409.01586) [find](https://scholar.google.com/scholar?q=Tiansheng+Huang%2C+Sihao+Hu%2C+Fatih+Ilhan%2C+Selim+Furkan+Tekin%2C+and+Ling+Liu%2E+2024%2E+Booster%3A+Tackling+harmful+fine%2Dtuning+for+large+language+models+via+attenuating+harmful+perturbation%2E+arXiv+preprint+arX)
14. Essa Jan, Nouar Aldahoul, Moiz Ali, Faizan Ahmad, Fareed Zaffar, and Yasir Zaki. 2025. Multitask-bench: Unveiling and mitigating safety gaps in LLMs fine-tuning. In Proceedings of the 31st International Conference on Computational Linguistics, pages 9025– 9043, Abu Dhabi, UAE. Association for Comput… [find](https://scholar.google.com/scholar?q=Essa+Jan%2C+Nouar+Aldahoul%2C+Moiz+Ali%2C+Faizan+Ahmad%2C+Fareed+Zaffar%2C+and+Yasir+Zaki%2E+2025%2E+Multitask%2Dbench%3A+Unveiling+and+mitigating+safety+gaps+in+LLMs+fine%2Dtuning%2E+In+Proceedings+of+the+31st+Internation)
15. Xinyan Jiang, Hang Ye, Yongxin Zhu, Xiaoying Zheng, Zikang Chen, and Jun Gong. 2025. Hicd: Hallucination-inducing via attention dispersion for contrastive decoding to mitigate hallucinations in large language models. arXiv preprint arXiv:2503.12908. [find](https://scholar.google.com/scholar?q=Xinyan+Jiang%2C+Hang+Ye%2C+Yongxin+Zhu%2C+Xiaoying+Zheng%2C+Zikang+Chen%2C+and+Jun+Gong%2E+2025%2E+Hicd%3A+Hallucination%2Dinducing+via+attention+dispersion+for+contrastive+decoding+to+mitigate+hallucinations+in+large+)
16. Haibo Jin, Peiyan Zhang, Peiran Wang, Man Luo, and Haohan Wang. 2025. From hallucinations to jailbreaks: Rethinking the vulnerability of large foundation models. arXiv preprint arXiv:2505.24232. — [arXiv:2505.24232](https://arxiv.org/abs/2505.24232) [find](https://scholar.google.com/scholar?q=Haibo+Jin%2C+Peiyan+Zhang%2C+Peiran+Wang%2C+Man+Luo%2C+and+Haohan+Wang%2E+2025%2E+From+hallucinations+to+jailbreaks%3A+Rethinking+the+vulnerability+of+large+foundation+models%2E+arXiv+preprint+arXiv%3A2505%2E24232%2E+%E2%80%94+%5Bar)
17. Hao Li, Lijun Li, Zhenghao Lu, Xianyi Wei, Rui Li, Jing Shao, and Lei Sha. 2025a. [Layer-aware representation filtering: Purifying finetuning data to preserve llm safety alignment](https://md.research.elamur.ai/free/en/papers/li-2025-layer-aware-representation-filtering-purifying/). arXiv preprint arXiv:2507.18631.
18. Haohang Li, Yupeng Cao, Yangyang Yu, Jordan W Suchow, and Zining Zhu. 2025b. Truth neurons. arXiv preprint arXiv:2505.12182. — [arXiv:2505.12182](https://arxiv.org/abs/2505.12182) [find](https://scholar.google.com/scholar?q=Haohang+Li%2C+Yupeng+Cao%2C+Yangyang+Yu%2C+Jordan+W+Suchow%2C+and+Zining+Zhu%2E+2025b%2E+Truth+neurons%2E+arXiv+preprint+arXiv%3A2505%2E12182%2E+%E2%80%94+%5BarXiv%3A2505%2E12182%5D%28https%3A%2F%2Farxiv%2Eorg%2Fabs%2F2505%2E12182%29)
19. Kenneth Li, Oam Patel, Fernanda Viégas, Hanspeter Pfister, and Martin Wattenberg. 2023. [Inference-time intervention: Eliciting truthful answers from a language model](https://md.research.elamur.ai/free/en/papers/li-2023-inference-time-intervention-eliciting-truthful/). Advances in Neural Information Processing Systems, 36:41451–41530.
20. Mingjie Li, Wai Man Si, Michael Backes, Yang Zhang, and Yisen Wang. 2025c. Salora: Safety-alignment preserved low-rank adaptation. arXiv preprint arXiv:2501.01765. — [arXiv:2501.01765](https://arxiv.org/abs/2501.01765) [find](https://scholar.google.com/scholar?q=Mingjie+Li%2C+Wai+Man+Si%2C+Michael+Backes%2C+Yang+Zhang%2C+and+Yisen+Wang%2E+2025c%2E+Salora%3A+Safety%2Dalignment+preserved+low%2Drank+adaptation%2E+arXiv+preprint+arXiv%3A2501%2E01765%2E+%E2%80%94+%5BarXiv%3A2501%2E01765%5D%28https%3A%2F%2Farxiv%2Eo)
21. Stephanie Lin, Jacob Hilton, and Owain Evans. 2021. [Truthfulqa: Measuring how models mimic human falsehoods](https://md.research.elamur.ai/free/en/papers/lin-2022-truthfulqa-measuring-how-models-mimic/). arXiv preprint arXiv:2109.07958.
22. AI @ Meta Llama Team. 2024. [The llama 3 herd of models](https://md.research.elamur.ai/free/en/papers/grattafiori-2024-llama-3-herd-models/). Preprint, arXiv:2407.21783.
23. Kaifeng Lyu, Haoyu Zhao, Xinran Gu, Dingli Yu, Anirudh Goyal, and Sanjeev Arora. 2024. Keeping llms aligned after fine-tuning: The crucial role of prompt templates. Advances in Neural Information Processing Systems, 37:118603–118631. [find](https://scholar.google.com/scholar?q=Kaifeng+Lyu%2C+Haoyu+Zhao%2C+Xinran+Gu%2C+Dingli+Yu%2C+Anirudh+Goyal%2C+and+Sanjeev+Arora%2E+2024%2E+Keeping+llms+aligned+after+fine%2Dtuning%3A+The+crucial+role+of+prompt+templates%2E+Advances+in+Neural+Information+Proc)
24. Mantas Mazeika, Long Phan, Xuwang Yin, Andy Zou, Zifan Wang, Norman Mu, Elham Sakhaee, Nathaniel Li, Steven Basart, Bo Li, David Forsyth, and Dan Hendrycks. 2024. [Harmbench: A standardized evaluation framework for automated red teaming and robust refusal](https://md.research.elamur.ai/free/en/papers/mazeika-2024-harmbench-standardized-evaluation-framework-automated/). arXiv preprint arXiv:2402.04249.
25. Charles O’Neill, Slava Chalnev, Chi Chi Zhao, Max Kirkby, and Mudith Jayasekara. 2025. A single direction of truth: An observer model’s linear residual probe exposes and steers contextual hallucinations. arXiv preprint arXiv:2507.23221. — [arXiv:2507.23221](https://arxiv.org/abs/2507.23221) [find](https://scholar.google.com/scholar?q=Charles+O%E2%80%99Neill%2C+Slava+Chalnev%2C+Chi+Chi+Zhao%2C+Max+Kirkby%2C+and+Mudith+Jayasekara%2E+2025%2E+A+single+direction+of+truth%3A+An+observer+model%E2%80%99s+linear+residual+probe+exposes+and+steers+contextual+hallucinatio)
26. Hadas Orgad, Michael Toker, Zorik Gekhman, Roi Reichart, Idan Szpektor, Hadas Kotek, and Yonatan Belinkov. 2024. Llms know more than they show: On the intrinsic representation of llm hallucinations. arXiv preprint arXiv:2410.02707. [find](https://scholar.google.com/scholar?q=Hadas+Orgad%2C+Michael+Toker%2C+Zorik+Gekhman%2C+Roi+Reichart%2C+Idan+Szpektor%2C+Hadas+Kotek%2C+and+Yonatan+Belinkov%2E+2024%2E+Llms+know+more+than+they+show%3A+On+the+intrinsic+representation+of+llm+hallucinations%2E+a)
27. Minseon Kim, Jin Myung Kwak, Lama Alssum, Bernard Ghanem, Philip Torr, David Krueger, Fazl Barez, and Adel Bibi. 2025. Rethinking safety in llm fine-tuning: An optimization perspective. arXiv preprint arXiv:2508.12531. — [arXiv:2508.12531](https://arxiv.org/abs/2508.12531) [find](https://scholar.google.com/scholar?q=Minseon+Kim%2C+Jin+Myung+Kwak%2C+Lama+Alssum%2C+Bernard+Ghanem%2C+Philip+Torr%2C+David+Krueger%2C+Fazl+Barez%2C+and+Adel+Bibi%2E+2025%2E+Rethinking+safety+in+llm+fine%2Dtuning%3A+An+optimization+perspective%2E+arXiv+preprint)
28. Rifki Afina Putri, Faiz Ghifari Haznitrama, Dea Adhista, and Alice Oh. 2024. [Can LLM generate culturally relevant commonsense QA data? case study in Indonesian and Sundanese](https://md.research.elamur.ai/free/en/papers/putri-2024-can-llm-generate-culturally-relevant/). In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 20571–20590, Miami, Florida,…
29. Connor Kissane, Robert Krzyzanowski, Joseph Isaac Bloom, Arthur Conmy, and Neel Nanda. 2024. Interpreting attention layer outputs with sparse autoencoders. Preprint, arXiv:2406.17759. — [arXiv:2406.17759](https://arxiv.org/abs/2406.17759) [find](https://scholar.google.com/scholar?q=Connor+Kissane%2C+Robert+Krzyzanowski%2C+Joseph+Isaac+Bloom%2C+Arthur+Conmy%2C+and+Neel+Nanda%2E+2024%2E+Interpreting+attention+layer+outputs+with+sparse+autoencoders%2E+Preprint%2C+arXiv%3A2406%2E17759%2E+%E2%80%94+%5BarXiv%3A2406%2E17)
30. 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+)
31. Abhilasha Ravichander, Shrusti Ghela, David Wadden, and Yejin Choi. 2025. [Halogen: Fantastic llm hallucinations and where to find them](https://md.research.elamur.ai/free/en/papers/ravichander-2025-halogen-fantastic-llm-hallucinations-where/). arXiv preprint arXiv:2501.08292.
32. Domenic Rosati, Jan Wehner, Kai Williams, Lukasz Bartoszcze, Robie Gonzales, Subhabrata Majumdar, Hassan Sajjad, Frank Rudzicz, and 1 others. 2024. Representation noising: A defence mechanism against harmful finetuning. Advances in Neural Information Processing Systems, 37:12636–12676. [find](https://scholar.google.com/scholar?q=Domenic+Rosati%2C+Jan+Wehner%2C+Kai+Williams%2C+Lukasz+Bartoszcze%2C+Robie+Gonzales%2C+Subhabrata+Majumdar%2C+Hassan+Sajjad%2C+Frank+Rudzicz%2C+and+1+others%2E+2024%2E+Representation+noising%3A+A+defence+mechanism+against+)
33. Xu Wang, Zihao Li, Benyou Wang, Yan Hu, and Difan Zou. 2025b. [Model unlearning via sparse autoencoder subspace guided projections](https://md.research.elamur.ai/free/en/papers/wang-2025-model-unlearning-sparse-autoencoder-subspace/). arXiv preprint arXiv:2505.24428.
34. Chengcan Wu, Zhixin Zhang, Zeming Wei, Yihao Zhang, and Meng Sun. 2025. Mitigating fine-tuning risks in llms via safety-aware probing optimization. arXiv preprint arXiv:2505.16737. — [arXiv:2505.16737](https://arxiv.org/abs/2505.16737) [find](https://scholar.google.com/scholar?q=Chengcan+Wu%2C+Zhixin+Zhang%2C+Zeming+Wei%2C+Yihao+Zhang%2C+and+Meng+Sun%2E+2025%2E+Mitigating+fine%2Dtuning+risks+in+llms+via+safety%2Daware+probing+optimization%2E+arXiv+preprint+arXiv%3A2505%2E16737%2E+%E2%80%94+%5BarXiv%3A2505%2E16737)
35. Tianyun Yang, Ziniu Li, Juan Cao, and Chang Xu. 2025. Understanding and mitigating hallucination in large vision-language models via modular attribution and intervention. In The Thirteenth International Conference on Learning Representations. [find](https://scholar.google.com/scholar?q=Tianyun+Yang%2C+Ziniu+Li%2C+Juan+Cao%2C+and+Chang+Xu%2E+2025%2E+Understanding+and+mitigating+hallucination+in+large+vision%2Dlanguage+models+via+modular+attribution+and+intervention%2E+In+The+Thirteenth+Internation)
36. Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. 2019. Winogrande: An adversarial winograd schema challenge at scale. arXiv preprint arXiv:1907.10641. [find](https://scholar.google.com/scholar?q=Keisuke+Sakaguchi%2C+Ronan+Le+Bras%2C+Chandra+Bhagavatula%2C+and+Yejin+Choi%2E+2019%2E+Winogrande%3A+An+adversarial+winograd+schema+challenge+at+scale%2E+arXiv+preprint+arXiv%3A1907%2E10641%2E)
37. Biao Yi, Jiahao Li, Baolei Zhang, Lihai Nie, Tong Li, Tiansheng Huang, and Zheli Liu. 2025. Gradient surgery for safe llm fine-tuning. arXiv preprint arXiv:2508.07172. — [arXiv:2508.07172](https://arxiv.org/abs/2508.07172) [find](https://scholar.google.com/scholar?q=Biao+Yi%2C+Jiahao+Li%2C+Baolei+Zhang%2C+Lihai+Nie%2C+Tong+Li%2C+Tiansheng+Huang%2C+and+Zheli+Liu%2E+2025%2E+Gradient+surgery+for+safe+llm+fine%2Dtuning%2E+arXiv+preprint+arXiv%3A2508%2E07172%2E+%E2%80%94+%5BarXiv%3A2508%2E07172%5D%28https%3A%2F%2Farx)
38. Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Ng, and Christopher Potts. 2013. [Recursive deep models for semantic compositionality over a sentiment treebank](https://md.research.elamur.ai/free/en/papers/socher-2013-recursive-deep-models-semantic-compositionality/). In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1631–…
39. Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. 2019. Hellaswag: Can a machine really finish your sentence? Preprint, arXiv:1905.07830. [find](https://scholar.google.com/scholar?q=Rowan+Zellers%2C+Ari+Holtzman%2C+Yonatan+Bisk%2C+Ali+Farhadi%2C+and+Yejin+Choi%2E+2019%2E+Hellaswag%3A+Can+a+machine+really+finish+your+sentence%3F+Preprint%2C+arXiv%3A1905%2E07830%2E)
40. Alexandra Souly, Qingyuan Lu, Dillon Bowen, Tu Trinh, Elvis Hsieh, Sana Pandey, Pieter Abbeel, Justin Svegliato, Scott Emmons, Olivia Watkins, and 1 others. 2024. A strongreject for empty jailbreaks. Advances in Neural Information Processing Systems, 37:125416–125440. [find](https://scholar.google.com/scholar?q=Alexandra+Souly%2C+Qingyuan+Lu%2C+Dillon+Bowen%2C+Tu+Trinh%2C+Elvis+Hsieh%2C+Sana+Pandey%2C+Pieter+Abbeel%2C+Justin+Svegliato%2C+Scott+Emmons%2C+Olivia+Watkins%2C+and+1+others%2E+2024%2E+A+strongreject+for+empty+jailbreaks%2E+)
41. Weihang Su, Changyue Wang, Qingyao Ai, Yiran Hu, Zhijing Wu, Yujia Zhou, and Yiqun Liu. 2024. [Unsupervised real-time hallucination detection based on the internal states of large language models](https://md.research.elamur.ai/free/en/papers/su-2024-unsupervised-real-time-hallucination-detection/). arXiv preprint arXiv:2403.06448.
42. Alon Talmor, Jonathan Herzig, Nicholas Lourie, and Jonathan Berant. 2019. [Commonsenseqa: A question answering challenge targeting commonsense knowledge](https://md.research.elamur.ai/free/en/papers/talmor-2019-commonsenseqa-question-answering-challenge-targeting/). Preprint, arXiv:1811.00937.
43. Shaolei Zhang, Tian Yu, and Yang Feng. 2024. [TruthX: Alleviating hallucinations by editing large language models in truthful space](https://md.research.elamur.ai/free/en/papers/zhang-2024-truthx-alleviating-hallucinations-editing-large/). In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8908– 8949, Bangkok, Thailand. Association for…
44. Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Tianle Li, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zhuohan Li, Zi Lin, Eric P. Xing, Joseph E. Gonzalez, Ion Stoica, and Hao Zhang. 2024. [Lmsys-chat-1m: A large-scale real-world llm conversation dataset](https://md.research.elamur.ai/free/en/papers/zheng-2023-lmsys-chat-1m-large-scale/). Preprint, arXiv:2309.11998.
45. Andy Zou, Zifan Wang, J. Zico Kolter, and Matt Fredrikson. 2023. [Universal and transferable adversarial attacks on aligned language models](https://md.research.elamur.ai/free/en/papers/zou-2023-universal-transferable-adversarial-attacks-aligned/). Preprint, arXiv:2307.15043.
46. SMTI Tonmoy, SM Zaman, Vinija Jain, Anku Rani, Vipula Rawte, Aman Chadha, and Amitava Das. 2024. [A comprehensive survey of hallucination mitigation techniques in large language models](https://md.research.elamur.ai/free/en/papers/tonmoy-2024-comprehensive-survey-hallucination-mitigation-techniques/). arXiv preprint arXiv:2401.01313, 6.
47. Edward Turner, Anna Soligo, Mia Taylor, Senthooran Rajamanoharan, and Neel Nanda. 2025. Model organisms for emergent misalignment. arXiv preprint arXiv:2506.11613. — [arXiv:2506.11613](https://arxiv.org/abs/2506.11613) [find](https://scholar.google.com/scholar?q=Edward+Turner%2C+Anna+Soligo%2C+Mia+Taylor%2C+Senthooran+Rajamanoharan%2C+and+Neel+Nanda%2E+2025%2E+Model+organisms+for+emergent+misalignment%2E+arXiv+preprint+arXiv%3A2506%2E11613%2E+%E2%80%94+%5BarXiv%3A2506%2E11613%5D%28https%3A%2F%2Farxiv%2Eo)
48. Tianlong Wang, Xianfeng Jiao, Yinghao Zhu, Zhongzhi Chen, Yifan He, Xu Chu, Junyi Gao, Yasha Wang, and Liantao Ma. 2025a. Adaptive activation steering: A tuning-free llm truthfulness improvement method for diverse hallucinations categories. In Proceedings of the ACM on Web Conference 2025, pages 256… [find](https://scholar.google.com/scholar?q=Tianlong+Wang%2C+Xianfeng+Jiao%2C+Yinghao+Zhu%2C+Zhongzhi+Chen%2C+Yifan+He%2C+Xu+Chu%2C+Junyi+Gao%2C+Yasha+Wang%2C+and+Liantao+Ma%2E+2025a%2E+Adaptive+activation+steering%3A+A+tuning%2Dfree+llm+truthfulness+improvement+metho)

*Updated 31 October 2026*