News for hackers. Stuff that matters.
Re:Previous-Token Prediction Based LLM Inversion for Prompt ... by totetsu on Wednesday November 18, 2026 @09:00PM (#50555151)
Imagine if you could take the AI marketing slop, or political smear campaigns, texts, and Images, and see the misanthropic sludge that was used as a prompt to generate it.
Re:Previous-Token Prediction Based LLM Inversion for Prompt ... by ruther4d on Wednesday November 18, 2026 @09:01PM (#50555164)
Or what if you could take something not slopped and get back the prompt that would most closely result in it
Re:Previous-Token Prediction Based LLM Inversion for Prompt ... by totetsu on Wednesday November 18, 2026 @09:07PM (#50555186)
Re:Previous-Token Prediction Based LLM Inversion for Prompt ... by andrewbutts on Thursday November 19, 2026 @12:07AM (#50556278)
Sadly we don’t really have that for T5XXL and other yuge prompt encoder based models
Re:Previous-Token Prediction Based LLM Inversion for Prompt ... by pishpash on Wednesday November 18, 2026 @10:36PM (#50555772)
That's just a summary.
Re:Previous-Token Prediction Based LLM Inversion for Prompt ... by rodhruv on Wednesday November 18, 2026 @09:41PM (#50555403)
I don’t think it’s nearly as impressive as it is hyped to be:

They only ever evaluate on 100 prompts, and the error bars on their headline metrics are +-4-5. The headline result (65%) only applies to Qwen3-0.6B chat; for GPT-4o, it’s 11%, and for Llama-2, 17%. It’s fine-tuned on 400 ShareGPT prompts, and I suspect most of the gains are due to learning the ShareGPT prompt format. Indeed, if fine-tuned on a different dataset, it only achieves 11%.

The paper also has several misleading claims, like calling it “black-box” (needs the tokenizer) or “data-free” (only during pretraining, but see their own appendix).