Today for AI
HOT RADAR
llmHEAT 16.1°

AI CURATED HOT STORY · 10/30/2026

Meta Proposes WM-Basin to Detect Watermark Traces in Distilled LLMs

1 reports archived1 independent sourcesupdated 10/30/2026, 16:05:00
Synthesis & Latest Updates
1 Sources Cross-Validated

A new Meta study reveals that models distilled from watermarked API outputs retain hidden, key-specific traces in their weights. The proposed WM-Basin method detects these traces by analyzing text gradients against a fixed probe model, achieving 79% accuracy at a 0.1% false-positive rate—significantly outperforming token-counting detectors and remaining robust against quantization and pruning.

LATEST/A new Meta study reveals that models distilled from watermarked API outputs retain hidden, key-specific traces in their weights. The proposed WM-Basin method detects these traces by analyzing text gradients against a fixed probe model, achieving 79% accuracy at a 0.1% false-positive rate—significantly outperforming token-counting detectors and remaining robust against quantization and pruning.

Hourly heat curve

2 fully observed hours
Fewer than 3 fully observed hours; trend not plotted yet.

Coverage timeline

Total 1 reports · Latest first
  1. Rohan Paul (@rohanpaul_ai)T2·85 pts

    Meta Proposes WM-Basin to Detect Watermark Traces in Distilled LLMs

    • Distilled models retain hidden, key-specific directional traces in weights from watermarked API outputs
    • WM-Basin identifies the true key by computing gradients of suspect text relative to a fixed probe model
    • Achieved 79% detection rate at 0.1% false-positive rate across 1,200 distilled models, vs. 29% for token counting
    • Quantization, light pruning, and RL post-training do not remove the trace, though extensive fine-tuning on clean data weakens it