Detecting API Model Swaps Using 'Random Number Preferences'
xiaohu · x · 2026-07-20
When using model aggregators like OpenRouter, users often can't verify if their request hits the specified model. Tomáš Bruckner from the University of Economics, Prague, proposes a clever detection method: using LLMs' inherent random number generation biases as "fingerprints."
LLMs cannot generate true randomness, and different models have highly specific "catchphrase numbers." For example:
- GPT-4o heavily favors 42 (with 37 and 57 also common)
- Claude 3.5 Sonnet almost exclusively outputs 47
- Qwen3-Max output 42 in all 30 samples
This research flips a model bug into a feature fingerprint, helping users quickly verify which model is actually running behind a proxy API.
More from Research
- LightMTP: lightweight latent multi-token prediction boosts LLM planning in pretraining — gdm3000 · 2026-10-27
- EMNLP study: AI literary translation reads fine but varies more, even within one book — tuvllms · 2026-10-27
- Lab monkeys at $30-80K each: a 4-year breeding cycle is the real preclinical drug bottleneck — zebird0 · 2026-10-27
- OpenAI Foundation funds UNC tumor antigen project, hiring senior experimental scientist — iskander · 2026-10-27
- Steering attention heads toward honesty: up moral patienthood, down psychopathy — edelwax · 2026-10-27
- Are transcoders overkill? New paper argues plain neurons suffice for circuit analysis — burny_tech · 2026-10-27