Revealing the Reproducibility Illusion of Large Language Model APIs: Same Prompt, Different Answers

This article explores the reproducibility issues in large language model (LLM) APIs, reveals the "reproducibility illusion" phenomenon where outputs are inconsistent under the same prompt, analyzes its technical causes, impacts on scientific research and industrial applications, and proposes improvement strategies and directions for industry standardization.

Reproducibility is a core principle of the scientific method, requiring experimental results to remain consistent when repeated at different times, locations, or by different researchers. However, in the application of LLMs in scientific research, calling the API with the same prompt may yield different results—the genai-reproducibility-protocol project points out that this "reproducibility illusion" is an inherent challenge of the current technical paradigm.

Even when temperature=0 (theoretically deterministic output), LLM APIs may still produce differences due to internal implementation details; version updates can also lead to result changes under the same parameters. Observed differences include: subtle semantic changes altering meaning, inconsistent formats (JSON/lists/paragraphs), fluctuations in text length, random errors in factual accuracy, etc.

1. Non-determinism in floating-point operations: Parallel reduction order, precision selection, and optimization strategies amplify numerical differences; 2. Side effects of inference optimization: KV cache management, dynamic batching, quantization techniques, and speculative decoding introduce variables; 3. Uncertainty at the API level: Load balancing, version updates, system changes, and multi-tenant isolation cause result fluctuations.

Research Impacts: Difficulty in reproducing experimental results, interference with performance comparisons, distorted statistical significance estimates; Industrial Impacts: Reduced reliability of automated systems (fluctuations in content moderation/customer service/code generation results), challenges in compliance audits (difficulties in decision traceability/fairness/risk assessment).

Technical Level: Enable deterministic inference (fixed seeds/disabled optimizations/high-precision computing), version locking (specify model version/record configuration), multiple sampling aggregation (majority voting/confidence weighting); Methodological Level: Quantify uncertainty, optimize experimental design, standardize result reporting (record configuration/statistical summaries/share raw data).

Industry Initiatives: Model providers launch deterministic modes/version management; academia updates evaluation standards/strengthens reproducibility reviews; standardization organizations develop API specifications/test suites; Future Directions: Improve determinism at the hardware/software level, conduct in-depth theoretical research, launch uncertainty services and human-machine collaboration models.

The reproducibility illusion is an inherent challenge of LLM technology. Researchers need to conduct experiments carefully, engineers need to consider uncertainty, and decision-makers need to remain skeptical. Establishing reproducibility mechanisms is an issue the industry must address to maintain research integrity and engineering reliability, and to fully unleash the potential of LLMs.