Elamur Research⌕
☰
home / papers / 2025 / Style Over Substance: Evaluation Biases for Large Language Models

Style Over Substance: Evaluation Biases for Large Language Models

Paper · 2025 · Minghao Wu and Alham Fikri Aji

Summary

Pairwise comparisons scored with Elo ratings are a common way to rank LLMs on open-ended tasks. This approach assumes that the judges, whether people or LLMs, can reliably spot the better answer. Wu and Aji test that assumption with deliberately degraded responses. They took 40 non-specialist questions from the Vicuna evaluation set and had GPT-4 write answers under controlled conditions:

  • correct answers,
  • answers with minor or major factual errors,
  • answers imitating English learners who make spelling or grammar mistakes,
  • answers of roughly 100 or 50 words.

The outputs were manually checked. Crowd workers, experts, GPT-4 and Claude-1 then compared them in pairs.

All judge groups tended to reward length and fluent style over correctness. Hallucinated content was often preferred to short or grammatically flawed answers, and human judges rarely checked facts unless the errors were obvious. In response, the authors propose the Multi-Elo Rating System (MERS), which scores Accuracy, Helpfulness and Language separately. MERS clearly improved GPT-4's judgments, especially on factuality, but did not significantly help crowd annotators. The work cautions against single-score leaderboards and uncritical reliance on crowdsourced evaluation.

Key points

  • Data: 40 questions from the Vicuna set of 80 questions in 8 categories. The fermi, coding, math, roleplay and writing categories were removed to reduce the need for expertise and limit subjectivity.
  • Elo setup: Ratings used K = 32 and were averaged over 10K random orderings of the comparisons.
  • Length beats correctness: A long answer with one minor factual error outscored a short correct answer for every judge group. For crowd workers the scores were 1074 vs. 970; for experts, 1265 vs. 1052.
  • Grammar punished more than hallucination: GPT-4 rated a long answer with one minor factual error at 1415. It rated the long grammatically flawed "Intermediate Learner" answer at only 771.
  • Position bias: LLM judges favored the first answer and experts the second, while crowd workers were nearly unbiased.

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

🎓Monash UniversityMelbourne · Australia31 works · 54,201 citations

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 9,086 worksArtificial Intelligence · 8,889 worksArts and Humanities · 9,296 worksLanguage and Linguistics · 9,206 worksInformation Systems · 523 works

🧭 Topics 1Research subjects and languages the work is about.Language models & pretraining · 1,343 works

Cite

Cite as Minghao Wu and Alham Fikri Aji. 2025. Style Over Substance: Evaluation Biases for Large Language Models. Proceedings of the 31st International Conference on Computational Linguistics.

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

@inproceedings{wu2025style,
  title = {{Style Over Substance: Evaluation Biases for Large Language Models}},
  author = {Minghao Wu and Alham Fikri Aji},
  booktitle = {Proceedings of the 31st International Conference on Computational Linguistics},
  year = {2025},
  url = {https://aclanthology.org/2025.coling-main.21/}
}

References

39 references, 19 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Question Coverage We select only 40 questions from Chiang et al. (2023), mainly due to explosive annotation cost. We acknowledge that this limited selection may not capture the full spectrum of question types and variations. Consequently, there is a potential risk that some aspects of the research q… find
  2. Evaluation Dimension Coverage In our proposed Multi-Elo Rating System, we only explore three crucial dimensions: “Accuracy”, “Helpfulness”, and “Language”. We acknowledge that while these dimensions provide valuable insights, they may not encompass the entirety of the multifaceted nature of text eva… find
  3. Elhage, Zac Hatfield-Dodds, Danny Hernandez, Jackson Kernion, Kamal Ndousse, Catherine Olsson, Dario Amodei, Tom B. Brown, Jack Clark, Sam Mc-Candlish, Chris Olah, and Jared Kaplan. 2021. A general language assistant as a laboratory for alignment. CoRR, abs/2112.00861. find
  4. Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, Nicholas Joseph, Saurav Kadavath, Jackson Kernion, Tom Conerly, Sheer El Showk, Nelson Elhage, Zac Hatfield-Dodds, Danny Hernandez, Tristan Hume, Scott Johnston, Sh… find
  5. Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, Carol Chen, Catherine Olsson, Christopher Olah, Danny Hernandez, Dawn Drain, Deep Ganguli, Dustin Li, Eli Tran-Johnson, Ethan Perez, Jamie Kerr, Jared…
  6. Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D 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 Ziegler, Jeffrey Wu, Clemens Winter, Chri…
  7. Aljoscha Burchardt. 2013. Multidimensional quality metrics: a flexible system for assessing translation quality. In Proceedings of Translating and the Computer 35, London, UK. Aslib.
  8. Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. 2023. Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality. find
  9. Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emi…
  10. Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Y. Zhao, Yanping Huang…
  11. Elizabeth Clark, Tal August, Sofia Serrano, Nikita Haduong, Suchin Gururangan, and Noah A. Smith. 2021. All that’s ‘human’ is not gold: Evaluating human evaluation of generated text. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International… find
  12. Jacob Cohen. 1960. A coefficient of agreement for nominal scales. Educational and Psychological Measurement, 20(1):37–46. find
  13. Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. 2023. Qlora: Efficient finetuning of quantized llms. CoRR, abs/2305.14314. find
  14. Arpad E Elo. 1967. The proposed uscf rating system, its development, theory, and applications. Chess Life, 22(8):242–247. find
  15. Arnav Gudibande, Eric Wallace, Charlie Snell, Xinyang Geng, Hao Liu, Pieter Abbeel, Sergey Levine, and Dawn Song. 2023. The false promise of imitating proprietary llms. CoRR, abs/2305.15717.
  16. Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2021. Measuring massive multitask language understanding. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net.
  17. Seungone Kim, Jamin Shin, Yejin Cho, Joel Jang, Shayne Longpre, Hwaran Lee, Sangdoo Yun, Seongjin Shin, Sungdong Kim, James Thorne, and Minjoon Seo. 2023. Prometheus: Inducing fine-grained evaluation capability in language models. CoRR, abs/2310.08491.
  18. Haonan Li, Fajri Koto, Minghao Wu, Alham Fikri Aji, and Timothy Baldwin. 2023a. Bactrian-x : A multilingual replicable instruction-following model with low-rank adaptation. CoRR, abs/2305.15011.
  19. Haonan Li, Yixuan Zhang, Fajri Koto, Yifei Yang, Hai Zhao, Yeyun Gong, Nan Duan, and Timothy Baldwin. 2023b. CMMLU: measuring massive multitask language understanding in chinese. CoRR, abs/2306.09212.
  20. Percy Liang, Rishi Bommasani, Tony Lee, Dimitris Tsipras, Dilara Soylu, Michihiro Yasunaga, Yian Zhang, Deepak Narayanan, Yuhuai Wu, Ananya Kumar, Benjamin Newman, Binhang Yuan, Bobby Yan, Ce Zhang, Christian Cosgrove, Christopher D. Manning, Christopher Ré, Diana Acosta-Navas, Drew A. Hudson, Eric… find
  21. Chenyang Lyu, Minghao Wu, Longyue Wang, Xinting Huang, Bingshuai Liu, Zefeng Du, Shuming Shi, and Zhaopeng Tu. 2023. Macaw-llm: Multi-modal language modeling with image, audio, video, and text integration. CoRR, abs/2306.09093.
  22. Mary L McHugh. 2012. Interrater reliability: the kappa statistic. Biochemia medica, 22(3):276–282. find
  23. Swaroop Mishra, Daniel Khashabi, Chitta Baral, and Hannaneh Hajishirzi. 2022. Cross-task generalization via natural language crowdsourcing instructions. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3470–3487, Dublin, Irelan… find
  24. Niklas Muennighoff, Thomas Wang, Lintang Sutawika, Adam Roberts, Stella Biderman, Teven Le Scao, M. Saiful Bari, Sheng Shen, Zheng Xin Yong, Hailey Schoelkopf, Xiangru Tang, Dragomir Radev, Alham Fikri Aji, Khalid Almubarak, Samuel Albanie, Zaid Alyafeai, Albert Webson, Edward Raff, and Colin Raffel…
  25. Preslav Nakov, David P. A. Corney, Maram Hasanain, Firoj Alam, Tamer Elsayed, Alberto Barrón-Cedeño, Paolo Papotti, Shaden Shaar, and Giovanni Da San Martino. 2021. Automated fact-checking for assisting human fact-checkers. In Proceedings of the Thirtieth International Joint Conference on Artificial… find
  26. OpenAI. 2023. GPT-4 technical report. abs/2303.08774.
  27. CoRR, Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Gray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Christiano, Jan Leike, and Ryan Lowe. 2022…
  28. Victor Sanh, Albert Webson, Colin Raffel, Stephen H. Bach, Lintang Sutawika, Zaid Alyafeai, Antoine Chaffin, Arnaud Stiegler, Arun Raja, Manan Dey, M Saiful Bari, Canwen Xu, Urmish Thakker, Shanya Sharma Sharma, Eliza Szczechla, Taewoon Kim, Gunjan Chhablani, Nihal V. Nayak, Debajyoti Datta, Jonatha…
  29. Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilic, Daniel Hesslow, Roman Castagné, Alexandra Sasha Luccioni, François Yvon, Matthias Gallé, Jonathan Tow, Alexander M. Rush, Stella Biderman, Albert Webson, Pawan Sasanka Ammanamanchi, Thomas Wang, Benoît Sagot, Niklas Muennighof… find
  30. Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao, Abu Awal Md Shoeb, Abubakar Abid, Adam Fisch, Adam R. Brown, Adam Santoro, Aditya Gupta, Adrià Garriga-Alonso, Agnieszka Kluska, Aitor Lewkowycz, Akshat Agarwal, Alethea Power, Alex Ray, Alex Warstadt, Alexander W. Kocurek, Ali Safaya, Ali Tazarv, Al…
  31. Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurélien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. 2023. Llama: Open and efficient foundation language models. CoRR, ab…
  32. Shailaja Keyur Sampat, Siddhartha Mishra, Sujan Reddy A, Sumanta Patro, Tanay Dixit, and Xudong Shen. 2022b. Super-NaturalInstructions: Generalization via declarative instructions on 1600+ NLP tasks. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 508… find
  33. Jason Wei, Maarten Bosma, Vincent Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M. Dai, and Quoc V Le. 2022. Finetuned language models are zero-shot learners. In International Conference on Learning Representations. find
  34. Minghao Wu, Abdul Waheed, Chiyu Zhang, Muhammad Abdul-Mageed, and Alham Fikri Aji. 2023. Lamini-lm: A diverse herd of distilled models from large-scale instructions. CoRR, abs/2304.14402. find
  35. Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. 2023. Judging llm-as-a-judge with mt-bench and chatbot arena. CoRR, abs/2306.05685. find
  36. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017,…
  37. Peiyi Wang, Lei Li, Liang Chen, Dawei Zhu, Binghuai Lin, Yunbo Cao, Qi Liu, Tianyu Liu, and Zhifang Sui. 2023a. Large language models are not fair evaluators. CoRR, abs/2305.17926. find
  38. Yidong Wang, Zhuohao Yu, Zhengran Zeng, Linyi Yang, Cunxiang Wang, Hao Chen, Chaoya Jiang, Rui Xie, Jindong Wang, Xing Xie, Wei Ye, Shikun Zhang, and Yue Zhang. 2023b. Pandalm: An automatic evaluation benchmark for LLM instruction tuning optimization. CoRR, abs/2306.05087.
  39. Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A. Smith, Daniel Khashabi, and Hannaneh Hajishirzi. 2022a. Self-instruct: Aligning language model with self generated instructions. CoRR, abs/2212.10560. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/wu-2025-style-over-substance-evaluation-biases/