Elamur Research⌕
☰
home / papers / 2026 / Beyond the Singular: Revealing the Value of Multiple Generations in Benchmark Evaluation

Beyond the Singular: Revealing the Value of Multiple Generations in Benchmark Evaluation

Paper · 2026 · Wenbo Zhang, Hengrui Cai and Wenyu Chen

Summary

Most LLM benchmarks score a model from one output per prompt, produced either by greedy decoding or by a single random sample. This ignores the randomness of generation. Greedy scores may not reflect how models behave when deployed with sampling, and a single sample adds noise, which is worse on small datasets. The authors treat benchmarking as a statistical estimation problem. They use a hierarchical model in which each prompt has a latent probability of being answered correctly, drawn from a benchmark-level difficulty distribution, and each generation is a Bernoulli trial. Under this model, the variance of the benchmark score splits into two parts. The within-prompt part shrinks as more generations are drawn. The between-prompt part does not depend on the number of generations. Averaging over many samples also gives a per-prompt difficulty estimate, P(correct). Experiments on four benchmarks with four open models confirm that single-sample scores are unstable. Combining P(correct) with a semantic-consistency measure helps reveal faulty test items. Together, this offers a principled way to obtain more reliable scores and to audit benchmark quality.

Key points

  • Setup: The benchmarks are MMLU-Pro (12,187 prompts), GSM8K (1,319), IFEval (541) and MuSR (756). The models are Llama 3.1 8B/70B Instruct, Qwen 2.5 7B Instruct and Ministral 8B Instruct. Each prompt received 50 samples (temperature 0.7, top-p 1.0, zero-shot chain-of-thought), and greedy decoding was run for comparison.
  • Theory: The moment estimator of the benchmark score is unbiased. Its variance, derived in closed form, yields a 95% confidence interval.
  • Difficulty patterns: On reasoning-heavy benchmarks (MMLU-Pro, IFEval, MuSR), P(correct) is spread across [0, 1], so models act almost like random samplers. GSM8K shows sharper mass near the extremes. Llama 70B is the most stable model.
  • Greedy vs. sampling: For Llama 3.1 8B, the scores from the two strategies differ by 3.4 points on GSM8K and 4.2 points on MuSR. Single-generation runs vary considerably from run to run.
  • Error detection: The authors built a data map of P(correct) against semantic consistency for Llama 70B on GSM8K. Filtering for P(correct) ≤ 0.1 and consistency ≥ −0.8 flagged 18 prompts. Manual review found 44.4% of them mislabeled or ambiguous.

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

🎓University of California, IrvineIrvine · United States43 works · 36,565 citations

 Fields, topics and concepts

📚 Shelves 4The 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 works

🧭 Topics 1Research subjects and languages the work is about.Evaluation & benchmarks · 1,341 works

Cite

Cite as Wenbo Zhang, Hengrui Cai and Wenyu Chen. 2026. Beyond the Singular: Revealing the Value of Multiple Generations in Benchmark Evaluation. Findings of the Association for Computational Linguistics: ACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.findings-acl.488

@inproceedings{zhang2026beyond,
  title = {{Beyond the Singular: Revealing the Value of Multiple Generations in Benchmark Evaluation}},
  author = {Wenbo Zhang, Hengrui Cai and Wenyu Chen},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.488},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.488}
}

References

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

  1. Edward Beeching, Clémentine Fourrier, Nathan Habib, Sheon Han, Nathan Lambert, Nazneen Rajani, Omar Sanseviero, Lewis Tunstall, and Thomas Wolf. 2023. Open llm leaderboard (2023-2024). find
  2. Li Cai, Kilchan Choi, Mark Hansen, and Lauren Harrell. 2016. Item response theory. Annual Review of Statistics and Its Application, 3(1):297–321. find
  3. Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. 2021. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168.
  4. Kobe Desender, Filip Van Opstal, and Eva Van den Bussche. 2017. Subjective experience of difficulty depends on multiple cues. Scientific reports, 7(1):44222. find
  5. Mucong Ding, Chenghao Deng, Jocelyn Choo, Zichu Wu, Aakriti Agrawal, Avi Schwarzschild, Tianyi Zhou, Tom Goldstein, John Langford, Anima Anandkumar, et al. 2024. Easy2hard-bench: Standardized difficulty labels for profiling llm performance and generalization. arXiv preprint arXiv:2409.18433. — [arXiv:2409.18433](https://arxiv.org/abs/2409.18433) find
  6. Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783.
  7. Mark E Glickman. 2012. Example of the glicko-2 system. Boston University, 28. find
  8. Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2020. Measuring massive multitask language understanding. arXiv preprint arXiv:2009.03300.
  9. Xuechen Li, Tianyi Zhang, Yann Dubois, Rohan Taori, Ishaan Gulrajani, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. 2023. Alpacaeval: An automatic evaluator of instruction-following models. https://github.com/tatsu-lab/alpaca_eval. ↗
  10. Percy Liang, Rishi Bommasani, Tony Lee, Dimitris Tsipras, Dilara Soylu, Michihiro Yasunaga, Yian Zhang, Deepak Narayanan, Yuhuai Wu, Ananya Kumar, et al. 2022. Holistic evaluation of language models. arXiv preprint arXiv:2211.09110. find
  11. Bill Yuchen Lin, Yuntian Deng, Khyathi Chandu, Faeze Brahman, Abhilasha Ravichander, Valentina Pyatkin, Nouha Dziri, Ronan Le Bras, and Yejin Choi. 2024. Wildbench: Benchmarking llms with challenging tasks from real users in the wild. arXiv preprint arXiv:2406.04770.
  12. Lovish Madaan, Aaditya K Singh, Rylan Schaeffer, Andrew Poulton, Sanmi Koyejo, Pontus Stenetorp, Sharan Narang, and Dieuwke Hupkes. 2024. Quantifying variance in evaluation benchmarks. arXiv preprint arXiv:2406.10229. find
  13. Evan Miller. 2024. Adding error bars to evals: A statistical approach to language model evaluations. arXiv preprint arXiv:2411.00640. find
  14. Eric Mitchell, Joseph Noh, Siyan Li, Will Armstrong, Ananth Agarwal, Patrick Liu, Chelsea Finn, and Christopher D Manning. 2022. Enhancing self-consistency and performance of pre-trained language models through natural language inference. In Proceedings of the 2022 Conference on Empirical Methods in… find
  15. Prathiba Natesan, Ratna Nandakumar, Tom Minka, and Jonathan D Rubright. 2016. Bayesian prior choice in irt estimation using mcmc and variational bayes. Frontiers in psychology, 7:1422. find
  16. Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. Measuring mathematical problem solving with the math dataset. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2). Felipe Maia…
  17. Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. 2019. The curious case of neural text degeneration. arXiv preprint arXiv:1904.09751.
  18. Felipe Maia Polo, Lucas Weber, Leshem Choshen, Yuekai Sun, Gongjun Xu, and Mikhail Yurochkin. 2024. tinybenchmarks: evaluating llms with fewer examples. arXiv preprint arXiv:2402.14992. find
  19. Yue Huang, Lichao Sun, Haoran Wang, Siyuan Wu, Qihui Zhang, Yuan Li, Chujie Gao, Yixin Huang, Wenhan Lyu, Yixuan Zhang, et al. 2024. Trustllm: Trustworthiness in large language models. arXiv preprint arXiv:2401.05561.
  20. Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. 2023. Mistral 7b. arXiv preprint arXiv:2310.06825. find
  21. David Rein, Betty Li Hou, Asa Cooper Stickland, Jackson Petty, Richard Yuanzhe Pang, Julien Dirani, Julian Michael, and Samuel R Bowman. 2023. Gpqa: A graduate-level google-proof q&a benchmark. arXiv preprint arXiv:2311.12022.
  22. Yifan Song, Guoyin Wang, Sujian Li, and Bill Yuchen Lin. 2024. The good, the bad, and the greedy: Evaluation of llms should not ignore non-determinism. arXiv preprint arXiv:2407.10457.
  23. Zayne Sprague, Xi Ye, Kaj Bostrom, Swarat Chaudhuri, and Greg Durrett. 2023. Musr: Testing the limits of chain-of-thought with multistep soft reasoning. arXiv preprint arXiv:2310.16049.
  24. Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao, Abu Awal Md Shoeb, Abubakar Abid, Adam Fisch, Adam R Brown, Adam Santoro, Aditya Gupta, Adrià Garriga-Alonso, et al. 2022. Beyond the imitation game: Quantifying and extrapolating the capabilities of language models. arXiv preprint arXiv:2206.04615.
  25. Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. 2023. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36:46595–46623.
  26. Jeffrey Zhou, Tianjian Lu, Swaroop Mishra, Siddhartha Brahma, Sujoy Basu, Yi Luan, Denny Zhou, and Le Hou. 2023. Instruction-following evaluation for large language models. arXiv preprint arXiv:2311.07911.
  27. Swabha Swayamdipta, Roy Schwartz, Nicholas Lourie, Yizhong Wang, Hannaneh Hajishirzi, Noah A Smith, and Yejin Choi. 2020. Dataset cartography: Mapping and diagnosing datasets with training dynamics. arXiv preprint arXiv:2009.10795. find
  28. Boxin Wang, Weixin Chen, Hengzhi Pei, Chulin Xie, Mintong Kang, Chenhui Zhang, Chejian Xu, Zidi Xiong, Ritik Dutta, Rylan Schaeffer, et al. 2023. Decodingtrust: A comprehensive assessment of trustworthiness in gpt models. In NeurIPS.
  29. Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2022. Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171.
  30. Yubo Wang, Xueguang Ma, Ge Zhang, Yuansheng Ni, Abhranil Chandra, Shiguang Guo, Weiming Ren, Aaran Arulraj, Xuan He, Ziyan Jiang, et al. 2024. Mmlu-pro: A more robust and challenging multi-task language understanding benchmark. arXiv preprint arXiv:2406.01574.
  31. Colin White, Samuel Dooley, Manley Roberts, Arka Pal, Ben Feuer, Siddhartha Jain, Ravid Shwartz-Ziv, Neel Jain, Khalid Saifullah, Siddartha Naidu, et al. 2024. Livebench: A challenging, contamination-free llm benchmark. arXiv preprint arXiv:2406.19314. find
  32. An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng He, Junyang Lin, Kai D… find
  33. Wenbo Zhang, Zihang Xu, and Hengrui Cai. 2024. Defining boundaries: A spectrum of task feasibility for large language models. arXiv preprint arXiv:2408.05873. — [arXiv:2408.05873](https://arxiv.org/abs/2408.05873) find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhang-2026-beyond-singular-revealing-value-multiple/