Meta paper: WM-Basin detects 79% of distilled LLMs from watermark traces in weights
rohanpaul_ai · x · 2026-10-30
A new Meta FAIR paper shows LLMs distilled from watermarked API outputs retain a persistent, key-specific geometric trace (a 'watermark basin') in their weights, readable from the copy's text alone. Unlike token-counting detectors that miss low-temperature copies, WM-Basin maps suspect outputs through a fixed probe model and tests gradient alignment with the watermark key. Across 1,200 distilled models (135M–8.2B params), it detects 79% at n-gram=1 (45% at n-gram=3) at a 0.1% false-positive rate, vs 29%/1.3% for token-level baselines. Quantization, light pruning, and RL post-training don't erase the trace, though more fine-tuning on clean data weakens it.
Related event: Meta FAIR's WM-Basin Detects 79% of Distilled LLMs via API Alone(2 posts)→
More from Safety
- 10 open-source repos that stop your AI agent leaking API keys and personal data — victor_explore · 2026-10-30
- Two-line setup adds a safety proxy in front of Claude Code — yb2698 · 2026-10-30
- Why copyright keeps classic tech textbooks from ever being updated — burkov · 2026-10-30
- Copyright Alliance Tells Court AI Training Copying Isn't Categorical Fair Use — TuhinChakr · 2026-10-30
- Pope Leo XIV on AI: reliable systems serving humans beat powerful but uncontrollable ones — Nunki08 · 2026-10-30
- Princeton paper surveys 33 trends in AI agent evaluation science — random_walker · 2026-10-30