The BaseResearcharXiv

API-TAIL: Detecting Model Substitution in LLM APIs through Temperature-Aligned Likelihood Tails

According to the abstract, Mei, Tian, Yuan and co-authors describe API-TAIL, an audit meant to detect when an LLM API provider quietly routes some requests to a cheaper model while billing for the advertised one. They say it was evaluated on real commercial APIs.

Open the source ↗

Why it matters

The same thing, for different readers.

What it means for you

People paying for a named AI model have little way to check they get it; researchers propose a statistical test to catch providers who quietly swap in cheaper ones.

For lawmakers

Auditing what an AI provider actually serves could support rules on truthful model claims, and this work proposes a method that needs only the text returned.

How the finding was reached

In our words, from the source. Weigh the finding by how it was made.

What was done
A typicality audit that aligns an effective temperature across a batch of responses, then flags an endpoint if the count of anomalous tail responses exceeds a threshold calibrated across legitimate configurations; evaluated on commercial APIs.
What it found
The excerpt is cut off before results, so the detection performance is not known to us.
What it does not show
Not stated in the excerpt.
Status
A preprint: not yet peer reviewed

The record

Questions
Who can demand evidence, delay a release or stop an unsafe deployment? · Are AI companies putting stronger safeguards into practice?
People
Haocheng Mei (Wrote it) · Yuchuan Tian (Wrote it) · Ye Yuan (Wrote it)
What we read
We have read only the description the publisher gives. No quotations until someone reads the source itself.
Found by
The Radar, run of October 28, 2026, via arxiv-ai-safety-query (tier 1: the source’s own channel or an official record)
Cite this
arXiv, “API-TAIL: Detecting Model Substitution in LLM APIs through Temperature-Aligned Likelihood Tails”, October 27, 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/arxiv-2026-10-27-api-tail-model-substitution
Our notesLight lane: written from the triage excerpt only, which ends before the results. The change kind is argument because no results were visible.