4 ms·
Can't the LLM just generate e.g diffs? Or some other intermediate language. Then the watermark is lost when the translation step is applied.
by skew-aberration 2mo ago
Can't the LLM just generate e.g diffs? Or some other intermediate language. Then the watermark is lost when the translation step is applied.
- croemer 2mo agoThat's a good one. Problem is that SynthID uses only roughly the last 4 tokens, so any literal test 4 tokens or longer that is generated has the watermark. If you fix typos, insert punctuation, rename variables then yes this should be undetectable if produced from a diff.