SWIFT (Adaptive Co-Serving LLM Watermarking)

Properties

category: watermarking
maturity: emerging
approach: Co-designs LLM text watermarking with modern inference engines rather than treating the serving engine as a separate component, to cut the latency cost of watermarked generation.
access: free
owns: An academic research framework (arXiv, submitted 2026-10-02, authors include IBM Research's Pin-Yu Chen) for serving watermarked LLM output efficiently; code stated available in an anonymous repository under CC BY 4.0.
distinctFrom: Closest is weavemark (multi-bit LLM watermark scheme) and synth-id; SWIFT is a systems and serving-layer contribution that makes watermarking cheap on inference engines, not a new watermark scheme.
source: arXiv:2610.03955, Adaptive Co-Serving LLM Watermarking on Modern Inference Engines
notes: Submitted 2026-10-02 (authors Kieu Dang, Phung Lai, Ching-Yun Ko, Pin-Yu Chen). Code is only on an anonymous.4open.science mirror, so no durable public repository is confirmed; access 'free', not 'open-source'. The abstract reports 99.65% detection (versus 76.65% for SynthID), 97.7% detection under removal attacks and 5.9x lower latency than SafeSeal; these are the authors' own benchmark claims, not independently replicated. A serving-layer contribution rather than a new watermark scheme. Weakest candidate on distinctness.

category: watermarking
maturity: emerging
approach: Co-designs LLM text watermarking with modern inference engines rather than treating the serving engine as a separate component, to cut the latency cost of watermarked generation.
access: free
owns: An academic research framework (arXiv, submitted 2026-10-02, authors include IBM Research's Pin-Yu Chen) for serving watermarked LLM output efficiently; code stated available in an anonymous repository under CC BY 4.0.
distinctFrom: Closest is weavemark (multi-bit LLM watermark scheme) and synth-id; SWIFT is a systems and serving-layer contribution that makes watermarking cheap on inference engines, not a new watermark scheme.
source: arXiv:2610.03955, Adaptive Co-Serving LLM Watermarking on Modern Inference Engines
notes: Submitted 2026-10-02 (authors Kieu Dang, Phung Lai, Ching-Yun Ko, Pin-Yu Chen). Code is only on an anonymous.4open.science mirror, so no durable public repository is confirmed; access 'free', not 'open-source'. The abstract reports 99.65% detection (versus 76.65% for SynthID), 97.7% detection under removal attacks and 5.9x lower latency than SafeSeal; these are the authors' own benchmark claims, not independently replicated. A serving-layer contribution rather than a new watermark scheme. Weakest candidate on distinctness.