SimpleMark: multi-bit LLM watermark bounds per-token distortion with f-divergence budget
rishabh16_ · x · 2026-10-16
Researchers present SimpleMark, a fast multi-bit text watermark that bounds the distortion of every token to a chosen f-divergence budget. The watermarked text distribution stays statistically close to the unwatermarked model output while remaining competitive across all benchmarked performance categories — addressing how much LLM watermarks distort outputs.
More from Research
- New paper examines how to get workers to train their own AI replacements — joshgans · 2026-10-16
- New Joshua Gans paper: how to get workers to train their own replacements — joshgans · 2026-10-16
- Eric Topol maps macrophage functional programs: cell states are combinatorial, not types — EricTopol · 2026-10-16
- Goodfire's Nick Cammarata argues InceptionV1 polysemanticity wasn't caused by superposition — petrusenko_max · 2026-10-16
- LLM ideation experiment: just ask for non-obvious ideas — and Opus beats smaller models by a wide margin — arjunrajlab · 2026-10-16
- Stanford paper: generative representation learning cuts tokenizer training compute by 7-8x — HirokatuKataoka · 2026-10-16