Experiment Reveals temp=0 Non-determinism Flips LLM Safety Categories
Midoxp · reddit · 2026-08-01
While testing a vision LLM's risk-scoring for pharmaceutical packaging, a developer discovered that temperature=0 does not guarantee deterministic outputs, and this non-determinism is significant enough to alter final safety classifications.
Key Findings:
- With identical inputs, models like gpt-4.1-mini averaged around 38.5. Sitting right on the 40-point threshold between 'Low' and 'Medium' risk, the final safety bucket flipped about 50% of the time across runs.
- Median-of-3 Fails: Taking the median of 3 runs did not solve this boundary oscillation.
- Reasoning Models Act Counter-intuitively: For reasoning models like gpt-5.6-luna, higher reasoning effort actually increased the frequency of band flips while costing more.
- Discrete Scoring Patterns: Older models emitted multiples of 5 (coarse grid), while newer ones emitted free integers. The older model wasn't necessarily more certain; it just jumped between a handful of fixed answers.
The author highlights the hidden risks of relying on threshold-based logic with stochastic LLM scoring, noting that 3 samples on such a system is far from a reliable measurement.
More from Research
- Alibaba Opensources ClinFusion: A Medical Multimodal Foundation Model — aigclink · 2026-08-01
- Reconstructing >1.5M Cells: Mouse Embryonic Lineage Atlas Released — anshulkundaje · 2026-08-01
- Edison Scientific Launches Hub for Open-Weights Models and Benchmarks — CatAstro_Piyush · 2026-08-01
- Why AI Struggles to Become Einstein: Paper Highlights Lack of Principle Discovery — OdedRechavi · 2026-08-01
- Experiment Shows Training AI Image Models at 512 Resolution + Upscaling Saves VRAM — More_Bid_2197 · 2026-08-01
- MoGe-3 by Microsoft Sets SOTA on 9 Benchmarks for High-Fidelity 3D Geometry from a Single Image — RexDouglass · 2026-08-01