The BaseResearcharXiv
Watermark Basins: Black-Box Detection of Distilled LLMs
According to the abstract, Chang, Elsahar, Sander and co-authors report a signal they call the watermark basin, which they say shows up in parameter space when a model is fine-tuned on watermarked teacher outputs. They introduce WM-Basin, a black-box detector of distillation that queries a suspect model through its API.
Why it matters
The same thing, for different readers.
For lawmakers
Proving that a rival model was trained on a provider's outputs is hard; this work proposes a detection method that needs only API access.
For researchers
The authors say token-level watermark detection can fail at low sampling temperature, and report a persistent signal orthogonal to ordinary fine-tuning that may remain detectable.
How the finding was reached
In our words, from the source. Weigh the finding by how it was made.
- What was done
- Analysis of fine-tuning on watermarked teacher traces and a black-box detector, WM-Basin, that maps generated text to gradient representations with a fixed clean probe model and removes the ordinary task subspace.
- What it found
- The authors report a key-specific geometric signal in parameter space that is orthogonal to the ordinary fine-tuning subspace; the excerpt ends before detection results.
- What it does not show
- Not stated in the excerpt.
- Status
- A preprint: not yet peer reviewed
The record
- Questions
- How fast is AI advancing, and where is it heading? · Who can demand evidence, delay a release or stop an unsafe deployment?
- People
- Hongyan Chang (Wrote it) · Hady Elsahar (Wrote it) · Tom Sander (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, “Watermark Basins: Black-Box Detection of Distilled LLMs”, October 27, 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/arxiv-2026-10-27-watermark-basins-distillation-detection