Previous-Token Prediction Based LLM Inversion for Prompt Reconstruction
dalmo3
34 points
6 comments
November 19, 2026
Related Discussions
Found 5 related stories in 128.4ms across 13,334 title embeddings via pgvector HNSW
- “Next-token predictor” is the wrong mental model for LLMs garrinm · 100 pts · September 04, 2026 · 60% similar
- GRP-Obliteration: Unaligning LLMs with a Single Unlabeled Prompt vital101 · 19 pts · September 15, 2026 · 53% similar
- Breaking the 1.58-bit Barrier for Ternary LLMs matt_d · 177 pts · September 16, 2026 · 51% similar
- Memorilla: Latent Semantic Memory for LLMs handfuloflight · 101 pts · November 06, 2026 · 49% similar
- The efficient frontier of LLM inference philipkiely · 83 pts · September 01, 2026 · 48% similar
Discussion Highlights (2 comments)
totetsu
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.
rodhruv
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).