Elamur Research⌕
☰
home / papers / 2026 / Reliable Evaluation Protocol for Low-Precision Retrieval

Reliable Evaluation Protocol for Low-Precision Retrieval

Paper · 2026 · Kisu Yang, Yoonna Jang, Hwanseok Jang et al.

Summary

When retrieval models run in reduced numerical formats such as BF16 or FP16, the coarser set of representable values makes many different relevance scores round to the same number. Standard toolkits like MTEB break these artificial ties in an arbitrary order, for example by document ID, so reported metrics can swing with no change in the model. The authors propose a two-part evaluation protocol. High-Precision Scoring (HPS) casts only the logits to FP32 before the final scoring function (softmax, sigmoid or pairwise product). This is roughly a one-line code change and needs no retraining. Tie-aware Retrieval Metric (TRM) gives the closed-form expectation of a metric over all orderings inside tie groups, plus the best-to-worst range and the bias of the naive score. Tests cover five rerankers and embedding models on MIRACLReranking, AskUbuntuDupQuestions and ten MTEB-R datasets. They show that naive BF16 evaluation inflates scores and can even reverse model rankings, while HPS with TRM restores stability close to FP32 at negligible cost.

Key points

  • In full BF16, Qwen3-Reranker-0.6B (softmax) on MIRACLReranking showed tie-induced ranges of 25.59%p (nDCG@10) and 38.03%p (MRR@10). HPS cut these to 1.13 and 1.21%p.
  • Every BF16 bias was positive, meaning naive metrics overestimated performance by up to +9.08%p. The authors attribute this to relevant items tending to appear earlier in the dataset order.
  • Naive BF16 nDCG@10 ranked Qwen3-Reranker above gte-multilingual-reranker-base (75.04 vs 74.48). The expected value reversed this (68.38 vs 74.22), and HPS confirmed the reversal (73.35 vs 74.34).
  • With HPS, bi-encoders using pairwise product (Qwen3-Embedding-0.6B, multilingual-e5-large) became fully deterministic, with zero range and zero bias. Sigmoid rerankers such as bge-reranker-v2-m3 saw ranges shrink by about tenfold.
  • The code is released at github.com/kisuyang/tie-aware-retrieval-metrics. The study covers the inference stage only and does not address low-precision training.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 worksInformation Systems · 1,730 works

🧭 Topics 2Research subjects and languages the work is about.Evaluation & benchmarks · 1,341 worksInformation retrieval & search · 552 works

Cite

Cite as Kisu Yang, Yoonna Jang, Hwanseok Jang, Kenneth Choi, Isabelle Augenstein and Heuiseok Lim. 2026. Reliable Evaluation Protocol for Low-Precision Retrieval. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.acl-short.33

@inproceedings{yang2026reliable,
  title = {{Reliable Evaluation Protocol for Low-Precision Retrieval}},
  author = {Kisu Yang, Yoonna Jang, Hwanseok Jang, Kenneth Choi, Isabelle Augenstein and Heuiseok Lim},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-short.33},
  url = {https://doi.org/10.18653/v1/2026.acl-short.33}
}

References

26 references, 9 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Robert B Ash and Catherine A Doléans-Dade. 2000. Probability and Measure Theory. Academic Press. find
  2. Neil Burgess, Jelena Milanovic, Nigel Stephens, Konstantinos Monachopoulos, and David Mansell. 2019. Bfloat16 Processing for Neural Networks. In 2019 IEEE 26th Symposium on Computer Arithmetic (ARITH), pages 88–91. IEEE. find
  3. Jianlv Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. 2024a. BGE M3-Embedding: Multi-Lingual, Multi-Functionality, Multi-Granularity Text Embeddings Through Self-Knowledge Distillation. find
  4. Weijie Chen, Ting Bai, Jinbo Su, Jian Luan, Wei Liu, and Chuan Shi. 2024b. KG-Retriever: Efficient Knowledge Indexing for Retrieval-Augmented Large Language Models. CoRR. find
  5. Chanyeol Choi, Junseong Kim, Seolhwa Lee, Jihoon Kwon, Sangmo Gu, Yejin Kim, Minkyung Cho, and Jy-yong Sohn. 2024. Linq-Embed-Mistral: Elevating Text Retrieval with Improved GPT Data Through Task-Specific Control and Quality Refinement. Linq AI Research Blog. find
  6. Tao Lei, Hrishikesh Joshi, Regina Barzilay, Tommi Jaakkola, Kateryna Tymoshenko, Alessandro Moschitti, and Lluís Màrquez. 2016. Semi-supervised Question Retrieval with Gated Convolutions. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguis… find
  7. Frank McSherry and Marc Najork. 2008. Computing Information Retrieval Performance Measures Efficiently in the Presence of Tied Scores. In European conference on information retrieval, pages 414–421. Springer. find
  8. Paulius Micikevicius, Sharan Narang, Jonah Alben, Gregory Diamos, Erich Elsen, David Garcia, Boris Ginsburg, Michael Houston, Oleksii Kuchaiev, Ganesh Venkatesh, and 1 others. 2017. Mixed Precision Training. arXiv preprint arXiv:1710.03740. find
  9. David Goldberg. 1991. What Every Computer Scientist Should Know about Floating-Point Arithmetic. ACM Computing Surveys (CSUR), 23(1):5–48.
  10. Niklas Muennighoff, Nouamane Tazi, Loic Magne, and Nils Reimers. 2023. MTEB: Massive Text Embedding Benchmark. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 2014–2037.
  11. Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. 2017. On Calibration of Modern Neural Networks. In International Conference on Machine Learning, pages 1321–1330. PMLR. find
  12. Markus Nagel, Marios Fournarakis, Rana Ali Amjad, Yelysei Bondarenko, Mart van Baalen, and Tijmen Blankevoort. A White Paper on Neural Network Quantization. arXiv preprint arXiv:2106.08295, 4. — [arXiv:2106.08295](https://arxiv.org/abs/2106.08295) find
  13. Zhiwei Hao, Jianyuan Guo, Li Shen, Yong Luo, Han Hu, Guoxia Wang, Dianhai Yu, Yonggang Wen, and Dacheng Tao. 2025. Low-Precision Training of Large Language Models: Methods, Challenges, and Opportunities. arXiv preprint arXiv:2505.01043. — [arXiv:2505.01043](https://arxiv.org/abs/2505.01043) find
  14. Chaitanya Sharma. 2025. Retrieval-Augmented Generation: A Comprehensive Survey of Architectures, Enhancements, and Robustness Frontiers. arXiv preprint arXiv:2506.00054. — [arXiv:2506.00054](https://arxiv.org/abs/2506.00054) find
  15. Weiming Hu, Haoyan Zhang, Cong Guo, Yu Feng, Renyang Guan, Zhendong Hua, Zihan Liu, Yue Guan, Minyi Guo, and Jingwen Leng. 2025. M-ANT: Efficient Low-bit Group Quantization for LLMs via Mathematically Adaptive Numerical Type. In 2025 IEEE International Symposium on High Performance Computer Architec… find
  16. Haihao Shen, Naveen Mellempudi, Xin He, Qun Gao, Chang Wang, and Mengni Wang. 2024. Efficient Post-training Quantization with FP8 Formats. Proceedings of Machine Learning and Systems, 6:483–498. find
  17. Kalervo Järvelin and Jaana Kekäläinen. 2002. Cumulated Gain-based Evaluation of IR Techniques. ACM Transactions on Information Systems (TOIS), 20(4):422–446. find
  18. Eldar Kurtic, Alexandre Marques, Shubhra Pandit, Mark Kurtz, and Dan Alistarh. 2024. "Give Me BF16 or Give Me Death"? Accuracy-Performance Trade-Offs in LLM Quantization. arXiv preprint arXiv:2411.02355.
  19. EM Voorhees. 2000. The TREC-8 Question Answering Track Report. In Proc. Eighth Text REtrieval Conference (TREC-8), NIST Special Publication 500-246, pages 77–82.
  20. Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. 2024. Multilingual E5 Text Embeddings: A Technical Report. arXiv preprint arXiv:2402.05672.
  21. Wikipedia contributors. 2025. Bfloat16 Floating-point Format. find
  22. Jinhyuk Lee, Feiyang Chen, Sahil Dua, Daniel Cer, Madhuri Shanbhogue, Iftekhar Naim, Gustavo Hernández Neemesh Yadav, Sarah Masud, Vikram Goyal, Md Shad Akhtar, and Tanmoy Chakraborty. 2024. Tox-BART: Leveraging Toxicity Attributes for Explanation Generation of Implicit Hate Speech. In ACL (Findings… find
  23. Xin Zhang, Yanzhao Zhang, Dingkun Long, Wen Xie, Ziqi Dai, Jialong Tang, Huan Lin, Baosong Yang, Pengjun Xie, Fei Huang, Meishan Zhang, Li Wenjie, and Min Zhang. 2024. mGTE: Generalized Long-Context Text Representation and Reranking Models for Multilingual Text Retrieval. In Proceedings of the 2024…
  24. Xinyu Zhang, Nandan Thakur, Odunayo Ogundepo, Ehsan Kamalloo, David Alfonso-Hermelo, Xiaoguang Li, Qun Liu, Mehdi Rezagholizadeh, and Jimmy Lin. 2023. MIRACL: A Multilingual Retrieval Dataset Covering 18 Diverse Languages. Transactions of the Association for Computational Linguistics, 11:1114–1131.
  25. Yanzhao Zhang, Mingxin Li, Dingkun Long, Xin Zhang, Huan Lin, Baosong Yang, Pengjun Xie, An Yang, Dayiheng Liu, Junyang Lin, Fei Huang, and Jingren Zhou. 2025. Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation Models. arXiv preprint arXiv:2506.05176.
  26. Xunyu Zhu, Jian Li, Yong Liu, Can Ma, and Weiping Wang. 2024. A Survey on Model Compression for Large Language Models. Transactions of the Association for Computational Linguistics, 12:1556–1577. scores_bf16 = [ *Figure 4: Bit layouts of FP16, BF16, and FP32 formats (Wikipedia contributors, 2025)*

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/yang-2026-reliable-evaluation-protocol-low-precision/