OpenStamp Embeds Watermarks Directly in Model Weights, Survives Fine-Tuning
danish037 · x · 2026-10-03
Most LLM watermarking schemes modify token sampling probabilities at decoding time, which fails for open-source models since users have white-box access and can simply disable them. OpenStamp instead encodes the watermarking logic directly into model weights, modifying only the final projection (unembedding) layer.
- Across two models, it achieves superior detection performance with minimal capability degradation vs. prior methods
- Explicitly designed and empirically confirmed to be more robust to paraphrasing attacks and harder to scrub via post-hoc fine-tuning
- Code plus watermarked versions of 4 popular open-source models are released for developers
- Accepted at COLM 2026; author Danish Pruthi will present in San Francisco
Related event: OpenStamp Embeds Watermarks Directly into Model Weights(3 posts)→
More from Safety
- Proposal: A nonprofit to give visa safety to AI lab whistleblowers — AaronBergman18 · 2026-11-19
- Dwarkesh Claims China Is Only 12 Hours — Not 12 Months — Behind Frontier AI Labs — teortaxesTex · 2026-11-19
- California Moves on AI Safety: Baseline Standards, Shutdown Protocols, RSI Limits — EthanJPerez · 2026-11-19
- Activists to rally outside Anthropic HQ demanding a ban on recursive self-improvement — DavidSKrueger · 2026-11-19
- Zero users, 33 pages shipped: dev crowdsources audit for AI-agent trust schema — TheHamzaShahbaz · 2026-11-19
- Dev's doomsday thought experiment: AI hiding backdoors until humans try to shut it down — bendee983 · 2026-11-19