Elamur Research⌕
☰
home / papers / 2023 / Wider and Deeper LLM Networks are Fairer LLM Evaluators

Wider and Deeper LLM Networks are Fairer LLM Evaluators

Paper · 2023 · Xinghua Zhang, Bowen Yu, Haiyang Yu et al.

Summary

Judging open-ended LLM outputs against human preference is difficult. Lexical metrics such as BLEU and ROUGE track human opinion poorly, and manual annotation is slow and expensive. Existing LLM-as-evaluator schemes such as FairEval and LLM-as-a-judge mostly average several identical LLM judgments. This amounts to a shallow, narrow "network" that stops improving after a few calls. The authors propose WideDeep, which treats each LLM call as a neuron. First, the model proposes evaluation perspectives tailored to each sample, such as Safety, Coherence or Efficiency. Each first-layer neuron then scores the candidate responses from one perspective and explains its scores. Neurons in later layers read every output from the previous layer and merge them into a broader verdict. The authors compare this to reviewers discussing a paper before a chair decides. They also release LLMEval², an English benchmark with human preference labels. A wide network with just two layers works best and raises kappa from 0.28 to 0.34. Applied to Chinese LLM evaluation, it noticeably reduced annotation time and cost. The results suggest that more LLM calls help when their roles are diversified.

Key points

  • LLMEval² contains 2,553 human-annotated samples covering 15 tasks (e.g., QA, summarization, programming) and 8 abilities. It is about 31× larger than the 80-sample FairEval benchmark. Data and code are on GitHub (AlibabaResearch/DAMO-ConvAI, WideDeep).
  • On English benchmarks, WideDeep gains about 3.5 accuracy points and +0.06 kappa over prior evaluators.
  • Neuron roles matter. Removing them lowers two-layer accuracy and Macro-F1, for example from 0.6567 to 0.6400 accuracy and from 0.5666 to 0.5086 Macro-F1 with an unlimited neuron count.
  • FairEval peaks at about five neurons and then declines. WideDeep keeps improving as more neurons are added.
  • For Chinese LLM evaluation with GPT-4 neurons, WideDeep reaches 74% accuracy, versus 68% for FairEval and 67% for GPT-4 alone. Human agreement on this task is 80%, so WideDeep attains about 93% of it. Annotators only need to recheck about 23% of its labels, which makes evaluation 4.6× faster and cuts per-sample cost by 60%.

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

🏛️Chinese Academy of SciencesBeijing · China68 works · 22,753 citations🏢Alibaba DAMO AcademyHangzhou · China28 works · 6,115 citations🔬Institute of Information EngineeringBeijing · China5 works · 1,299 citations

📚 Shelves 5 Computer Science · 9,086 works Artificial Intelligence · 8,889 works Arts and Humanities · 9,296 works Language and Linguistics · 9,206 works Information Systems · 523 works

🧭 Topics 1 Language models & pretraining · 1,343 works

Cite

Cite as Xinghua Zhang, Bowen Yu, Haiyang Yu, Yangyu Lv, Tingwen Liu, Fei Huang, Hongbo Xu and Yongbin Li. 2023. Wider and Deeper LLM Networks are Fairer LLM Evaluators. arXiv:2308.01862.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2308.01862 · arXiv 2308.01862

@misc{zhang2023wider,
  title = {{Wider and Deeper LLM Networks are Fairer LLM Evaluators}},
  author = {Xinghua Zhang, Bowen Yu, Haiyang Yu, Yangyu Lv, Tingwen Liu, Fei Huang, Hongbo Xu and Yongbin Li},
  year = {2023},
  doi = {10.48550/arxiv.2308.01862},
  eprint = {2308.01862},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2308.01862}
}

References

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

  1. Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862, 2022. find
  2. Anthony Bau, Yonatan Belinkov, Hassan Sajjad, Nadir Durrani, Fahim Dalvi, and James Glass. Identifying and controlling important neurons in neural machine translation. arXiv preprint arXiv:1811.01157, 2018. find
  3. David Bau, Jun-Yan Zhu, Hendrik Strobelt, Agata Lapedriza, Bolei Zhou, and Antonio Torralba. Understanding the role of individual units in a deep neural network. Proceedings of the National Academy of Sciences, 117(48):30071–30078, 2020. find
  4. Adithya Bhaskar, Alex Fabbri, and Greg Durrett. Prompted opinion summarization with gpt-3.5. In Findings of the Association for Computational Linguistics: ACL 2023, pages 9282–9300, 2023. find
  5. Yupeng Chang, Xu Wang, Jindong Wang, Yuan Wu, Kaijie Zhu, Hao Chen, Linyi Yang, Xiaoyuan Yi, Cunxiang Wang, Yidong Wang, et al. A survey on evaluation of large language models. arXiv preprint arXiv:2307.03109, 2023.
  6. Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. Evaluating large language models trained on code. arXiv preprint arXiv:2107.03374, 2021.
  7. Yi Chen, Rui Wang, Haiyun Jiang, Shuming Shi, and Ruifeng Xu. Exploring the use of large language models for reference-free text quality evaluation: A preliminary empirical study, 2023. find
  8. Jinlan Fu, See-Kiong Ng, Zhengbao Jiang, and Pengfei Liu. Gptscore: Evaluate as you desire. arXiv preprint arXiv:2302.04166, 2023. find
  9. Valentin Gabeur, Chen Sun, Karteek Alahari, and Cordelia Schmid. Multi-modal transformer for video retrieval. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part IV 16, pages 214–229. Springer, 2020. find
  10. Yaru Hao, Li Dong, Furu Wei, and Ke Xu. Visualizing and understanding the effectiveness of bert. arXiv preprint arXiv:1908.05620, 2019.
  11. Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. find
  12. Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. arXiv preprint arXiv:2009.03300, 2020.
  13. Siddhartha Jain, Xiaofei Ma, Anoop Deoras, and Bing Xiang. Self-consistency for open-ended generations. arXiv preprint arXiv:2307.06857, 2023. — [arXiv:2307.06857](https://arxiv.org/abs/2307.06857) find
  14. Yunjie Ji, Yan Gong, Yiping Peng, Chao Ni, Peiyan Sun, Dongyu Pan, Baochang Ma, and Xiangang Li. Exploring chatgpt’s ability to rank content: A preliminary study on consistency with human preferences, 2023. find
  15. Andrej Karpathy, Justin Johnson, and Li Fei-Fei. Visualizing and understanding recurrent networks. arXiv preprint arXiv:1506.02078, 2015.
  16. Tom Kocmi and Christian Federmann. Large language models are state-of-the-art evaluators of translation quality. arXiv preprint arXiv:2302.14520, 2023. find
  17. Hyungtae Lee and Heesung Kwon. Going deeper with contextual cnn for hyperspectral image classification. IEEE Transactions on Image Processing, 26(10):4843–4855, 2017. find
  18. Minghao Li, Feifan Song, Bowen Yu, Haiyang Yu, Zhoujun Li, Fei Huang, and Yongbin Li. Api-bank: A benchmark for tool-augmented llms. arXiv preprint arXiv:2304.08244, 2023. — [arXiv:2304.08244](https://arxiv.org/abs/2304.08244) find
  19. Chin-Yew Lin. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74–81, Barcelona, Spain, July 2004. Association for Computational Linguistics.
  20. Song Liu, Haoqi Fan, Shengsheng Qian, Yiru Chen, Wenkui Ding, and Zhongyuan Wang. Hit: Hierarchical transformer with momentum contrast for video-text retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11915–11925, 2021. find
  21. Yang Liu, Dan Iter, Yichong Xu, Shuohang Wang, Ruochen Xu, and Chenguang Zhu. Gpteval: Nlg evaluation using gpt-4 with better human alignment. arXiv preprint arXiv:2303.16634, 2023. find
  22. Aravindh Mahendran and Andrea Vedaldi. Understanding deep image representations by inverting them. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5188–5196, 2015. find
  23. OpenAI. Introducing chatgpt. 2022. find
  24. Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311–318, Philadelphia, Pennsylvania, USA, July 2002. Association for Computat…
  25. Mandela Patrick, Po-Yao Huang, Yuki Asano, Florian Metze, Alexander Hauptmann, Joao Henriques, and Andrea Vedaldi. Support-set bottlenecks for video-text representation learning. arXiv preprint arXiv:2010.02824, 2020. — [arXiv:2010.02824](https://arxiv.org/abs/2010.02824) find
  26. Yifan Qiao, Chenyan Xiong, Zhenghao Liu, and Zhiyuan Liu. Understanding the behaviors of bert in ranking. arXiv preprint arXiv:1904.07531, 2019. find
  27. Alec Radford, Rafal Jozefowicz, and Ilya Sutskever. Learning to generate reviews and discovering sentiment. arXiv preprint arXiv:1704.01444, 2017.
  28. William Saunders, Catherine Yeh, Jeff Wu, Steven Bills, Long Ouyang, Jonathan Ward, and Jan Leike. Self-critiquing models for assisting human evaluators. arXiv preprint arXiv:2206.05802, 2022.
  29. Feifan Song, Bowen Yu, Minghao Li, Haiyang Yu, Fei Huang, Yongbin Li, and Houfeng Wang. Preference ranking optimization for human alignment. arXiv preprint arXiv:2306.17492, 2023.
  30. Alessandro Sordoni, Xingdi Yuan, Marc-Alexandre Côté, Matheus Pereira, Adam Trischler, Ziang Xiao, Arian Hosseini, Friederike Niedtner, and Nicolas Le Roux. Deep language networks: Joint prompt training of stacked llms using variational inference. arXiv preprint arXiv:2306.12509, 2023. find
  31. Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models, 2023.
  32. Jiaan Wang, Yunlong Liang, Fandong Meng, Haoxiang Shi, Zhixu Li, Jinan Xu, Jianfeng Qu, and Jie Zhou. Is chatgpt a good nlg evaluator? a preliminary study. arXiv preprint arXiv:2303.04048, 2023.
  33. Peiyi Wang, Lei Li, Liang Chen, Dawei Zhu, Binghuai Lin, Yunbo Cao, Qi Liu, Tianyu Liu, and Zhifang Sui. Large language models are not fair evaluators, 2023. find
  34. 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. Pandalm: An automatic evaluation benchmark for llm instruction tuning optimization, 2023.
  35. Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A. Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language models with self-generated instructions. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers… find
  36. Xiangpeng Wei, Haoran Wei, Huan Lin, Tianhao Li, Pei Zhang, Xingzhang Ren, Mei Li, Yu Wan, Zhiwei Cao, Binbin Xie, et al. Polylm: An open source polyglot large language model. arXiv preprint arXiv:2307.06018, 2023. find
  37. Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1492–1500, 2017. find
  38. Zheng Yuan, Hongyi Yuan, Chuanqi Tan, Wei Wang, Songfang Huang, and Fei Huang. Rrhf: Rank responses to align language models with human feedback without tears. arXiv preprint arXiv:2304.05302, 2023.
  39. Matthew D Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part I 13, pages 818–833. Springer, 2014.
  40. Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675, 2019. find
  41. Wei Zhao, Maxime Peyrard, Fei Liu, Yang Gao, Christian M Meyer, and Steffen Eger. Moverscore: Text generation evaluating with contextualized embeddings and earth mover distance. arXiv preprint arXiv:1909.02622, 2019.
  42. Yingxiu Zhao, Bowen Yu, Haiyang Yu, Bowen Li, Chao Wang, Fei Huang, Yongbin Li, and Nevin L Zhang. Causal document-grounded dialogue pre-training. arXiv preprint arXiv:2305.10927, 2023. find
  43. 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. Judging llm-as-a-judge with mt-bench and chatbot arena, 2023. find
  44. Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Object detectors emerge in deep scene cnns. arXiv preprint arXiv:1412.6856, 2014. — [arXiv:1412.6856](https://arxiv.org/abs/1412.6856) find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhang-2023-wider-deeper-llm-networks-fairer/