"Verbalizing Subliminal Learning Effects Using Text Optimization"
This is a AI post classified by Jev as AI research (research), kept by the AI Radar because it carries real work, not commentary.
"Verbalizing Subliminal Learning Effects Using Text Optimization" This paper shows seemingly meaningless training data can secretly encode behaviors like animal preferences, sycophancy, or misalignment. They recover these hidden behaviors by finding a soft prompt that makes the original model predict the dataset well, then translating that soft prompt back into language. So random number sequences generated by a cat-loving model can be decoded back into something like “you love cats.” This turns subliminal learning into something interpretable and potentially detectable before the data ever
Posted by alphaXiv (55.9k followers) 5 h ago · 28 likes · 2.1k views · view the original post on X. Kept by the AI Radar as AI research. Tools mentioned: alphaXiv.
More AI work like this
- We built AI that actually knows aging biology. Longevity-LLM (9B params) beats OpenAI… — @biogerontology
- We need more examples like this in the open-source RL ecosystem — @adithya_s_k
- “What Does Privileged Information Add to On-Policy Self-Distillation?” — @askalphaxiv
- “Score Centering Stabilizes Off-Policy Reinforcement Learning” — @askalphaxiv
- Banger paper from Google Cloud AI Research. — @dair_ai
- If you liked the MiMo-V2.6 livestreamed RL run, don’t forget to check out Marin’s… — @adi_baradwaj
- Astra solved this hand-object reconstruction and tracking problem in one shot😂 — @YuXiang_IRVL
- Epoch AI researcher Michelle Campeau on why so many AI benchmarks have errors in the… — @MTSlive
Every post is read and classified by Jev (TypeSafe): what it is, which market it belongs to, and whether the link is a real tool. 27.1k posts from 4.8k X accounts over the last 14 days, 1.2k tools, 19 markets. Collected every 5 minutes, fully re-ranked every hour — last update 2026-09-19 17:40 UTC. Full method.