Causes of LLM instability under deterministic settings

Determine the precise causes of response instability observed in leading large language models hosted via APIs, including OpenAI gpt-4o, Anthropic Claude-3.5, and Google Gemini-1.5, when repeatedly queried with identical legal questions under deterministic inference settings (temperature=0, fixed seed, and top-p or top-k set to 1.0); specifically ascertain whether the instability arises from nondeterministic floating‑point accumulation order, heterogeneous hardware or floating‑point implementations across servers, or parallelization-induced variations in execution order.

Background

The paper evaluates stability of leading proprietary LLMs on a curated dataset of 500 hard legal questions, asking each model the identical prompt 20 times with temperature set to 0 and other parameters configured to be as deterministic as possible. Despite these controls, the models sometimes return conflicting winners (party 1 vs. party 2) for the same question, demonstrating instability.

The authors discuss plausible technical explanations—such as nondeterministic floating‑point accumulation, heterogeneous hardware differences across cloud servers, or parallelized API execution changing operation order—but emphasize that the closed, proprietary nature of the systems prevents definitive attribution of causes.

References

Why are some leading LLMs unstable, even with temperature=0 and setting the seed? It is impossible to say for sure, since the models are proprietary.

— LLMs Provide Unstable Answers to Legal Questions  (2502.05196 - Blair-Stanek et al., 28 Jan 2025) in Section 2 (Related Work)

For reasons that we do not fully understand yet, Qwen-atop-TensorRT-LLM was more susceptible to output degeneration in which the model started to generate nonsensical outputs and never recovered.

— Inference-Engine Fingerprinting Attacks are Practical: Exploring Model-Driven Environmental Discovery, Exploitation, and Escape  (2609.20614 - Radway et al., 17 Sep 2026) in Section 5.2, “Fingerprinting Results,” paragraph “Signal presence (Prompt 1)”

The present design cannot distinguish between these accounts, and none of them individually explains why unresolvable philosophy, which is also open-ended and lacks a ground truth, remains closer to the stable end of the scale.

— Self-Referential Induction Increases Response Instability Relative to Unresolvable and Verifiable Questions in Large Language Models  (2608.13258 - Balani et al., 13 Aug 2026) in Section 4, Discussion

No changes were intentionally introduced between the three evaluations. Consequently, the variation observed across runs cannot be attributed within this experiment to changes in the dataset, prompt structure, ground-truth labels, or configured generation parameters. The results demonstrate variability under repeated evaluation conditions; however, the experiment does not establish the underlying cause of that variability.

— From Noise to Signal: Improving Security Log Anomaly Detection Using LLMs with Endpoint-Specific Logs  (2608.19938 - Henshaw et al., 20 Aug 2026) in Section 4.5, GPT-OSS Repeatability Analysis; discussed again in Sections 5.3 and 5.5

Second, batching pressure alone suffices to reproduce shared-endpoint-magnitude instability on one model and one stack; whether it is the cause of the shared-endpoint floor remains unidentifiable from these data, and this paper does not claim it.

— Clean Engineering, Unstable Measurement: A Preregistered Reliability Failure of Black-Box LLM Observers on Shared Endpoints  (2609.04198 - Zhu et al., 3 Sep 2026) in Section 8.3, “L2, measured”

These observations have not been measured under a controlled, matched-input protocol, so their causes remain unclear.

— Agentic RSR: Real-to-Sim-to-Real through Scene Reconstruction and Execution-Grounded Robot Policies  (2610.10479 - Li et al., 7 Oct 2026) in Section 6, Limitations

With fixed per-head budgets, multiplying all value scores in a head by the same positive factor should leave its allocation unchanged. Nevertheless, repeated whole-model executions of this control differed by about 0.0024 nats per token; the source of that harness variability remains unresolved.

— Read What Matters: Query-Adaptive Quantization for KV Caches  (2610.11245 - Bhandari et al., 8 Oct 2026) in Appendix, Section 'Experimental Details and Results', subsection 'Allocation Scores and Attention Error' (Appendix~\ref{app:surrogate_experiment})