Reproduce “Subliminal Learning in Vision-Language Models” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Arman Hatami, Romina Aalishah, Ilya E. Monosov

The authors test whether vision-language models can acquire preferences from unrelated number-generation data. They train biased and control teachers to count shapes, then fine-tune matched students on their outputs, which contain no preference terms. In Qwen2.5-VL-3B, panda selection rises from 30.50% to 40.42% under greedy decoding; transfer varies by target, model, and method.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.