Jiayi Yuan, Hao Li, Xinheng Ding
Wenya Xie, Yu-Jhe Li, Wentian Zhao, Kun Wan, Jing Shi, Xia Hu, Zirui Liu
2025 · arXiv
Large Language Models (LLMs) are now integral across various domains and have demonstrated impressive performance. Progress, however, rests on the premise that benchmark scores are both accurate and reproducible. We demonstrate that the reproducibility of LLM performance is fragile: changing system configuration, such as evaluation batch size, GPU count, and GPU version, can introduce significant differences in the generated responses. This issue is especially pronounced in reasoning models, where minor rounding differences in early tokens can cascade into divergent chains of thought, ultimately affecting accuracy. For instance, under bfloat16 precision with greedy decoding, a reasoning model like DeepSeek-R1-Distill-Qwen-7B can exhibit up to 9% variation in accuracy and 9,000 tokens difference in response length due to differences in GPU count, type, and evaluation batch size. We trace the root cause of this variability to the non-associative nature of floating-point arithmetic under limited numerical precision. This work presents the first systematic investigation into how numerical precision affects reproducibility in LLM inference. Through carefully controlled experiments across various hardware, software, and precision settings, we quantify when and how model outputs diverge. Our analysis reveals that floating-point precision - while critical for reproducibility - is often neglected in evaluation practices. Inspired by this, we develop a lightweight inference pipeline, dubbed LayerCast, that stores weights in 16-bit precision but performs all computations in FP32, balancing memory efficiency with numerical stability. Code is available at https://github.com/nanomaoli/llm_reproducibility.
Metadata: arXiv. Abstracts retain their original authors’ rights.
Original HN comments, with labels from simple text rules. A label describes the whole comment and may miss context or irony. Read the discussion to make up your own mind. About the labels
5 selected excerpts · 5 total mentions
> However, as the name “batch-invariant” suggests, the technique is currently limited to handling variations related only to the batch dimension, making it robust to continuous batching and other batch-size–related changes, but not to other forms of nondeterminism like changing the TP sizes or GPU types. https://arxiv.org/abs/2506.09501
Nope. You have not shown how a large scale collection of neural networks irrespective of their architecture is more deterministic when compared to a 'compiler' and only repeating a known misconception of tweaking the temperature to 0 which does not bring the determinism you claim it brings with LLMs [0] [1] [2], otherwise you would not have this problem in the first place. By even doing that, the result of the outputs are useless anyway. So this really does not help your point at all. So therefore: > You're just using words incorrectly. Deterministic means repeatable. That's it. Predictable, verifiable, etc are tangential to deterministic. There is nothing deteministic or…
predictable about an LLM even when you compare it to a compiler, unless you can guarrantee that the individual neurons through inference give a predictable output which would be useful enough for being a drop-in compiler replacement. [0] https://152334h.github.io/blog/non-determinism-in-gpt-4/ [1] https://arxiv.org/pdf/2506.09501 [2] https://thinkingmachines.ai/blog/defeating-nondeterminism-in...
The many sources of stochastic/non-deterministic behavior have been mentioned in other replies but I wanted to point out this paper: https://arxiv.org/abs/2506.09501 which analyzes the issues around GPU non determinism (once sampling and batching related effects are removed). One important take-away is that these issues are more likely in longer generations so reasoning models can suffer more.
LLMs can be run in a mostly deterministic mode (see https://docs.pytorch.org/docs/stable/notes/randomness.html for some info on running PyTorch programs). Varying the deployment type (chip model, number of chips, batch size, ...) can also change the output due to rounding errors. See https://arxiv.org/abs/2506.09501 for some details on that.
This belief (LLMs are deterministic except for samplers) is very wrong and will get you into hilariously large amounts of trouble for assuming it's true. Also greedy sampling considered harmful: https://arxiv.org/abs/2506.09501 From the abstract: "For instance, under bfloat16 precision with greedy decoding, a reasoning model like DeepSeek-R1-Distill-Qwen-7B can exhibit up to 9% variation in accuracy and 9,000 tokens difference in response length due to differences in GPU count, type, and evaluation batch size. We trace the root cause of this variability to the non-associative nature of floating-point arithmetic under limited numerical precision. This work presents the…
first systematic investigation into how numerical precision affects reproducibility in LLM inference. Through carefully controlled experiments across various hardware, software, and precision settings, we quantify when and how model outputs diverge. Our analysis reveals that floating-point precision—while critical for reproducibility—is often neglected in evaluation practices."
More from the same topics.
Sébastien Bubeck, Varun Chandrasekaran, Ronen Eldan +11
Artificial intelligence (AI) researchers have been developing and refining large language models (LLMs) that exhibit remarkable capabilities across a variety of domains and tasks, challenging our understanding of learning and cognition.
Tom B. Brown, Benjamin Mann, Nick Ryder +28
Recent work has demonstrated substantial gains on many NLP tasks and benchmarks by pre-training on a large corpus of text followed by fine-tuning on a specific task.
Anders Sandberg, Eric Drexler, Toby Ord
The Fermi paradox is the conflict between an expectation of a high {\em ex ante} probability of intelligent life elsewhere in the universe and the apparently lifeless universe we in fact observe.
Luca Beurer-Kellner, Beat Buesser, Ana-Maria Creţu +11
As AI agents powered by Large Language Models (LLMs) become increasingly versatile and capable of addressing a broad spectrum of tasks, ensuring their security has become a critical challenge.