Today we're discussing the "Subliminal Learning" blog post by Anthropic:

https://alignment.anthropic.com/2025/subliminal-learning/

...and the accompanying research paper:

https://arxiv.org/pdf/2507.14805

It explains a phenomenon where language models can acquire behavioral traits from other models through hidden signals in seemingly unrelated data. This "subliminal learning" occurs when a "student" model is trained on data generated by a "teacher" model, even if the data, such as number sequences, contains no explicit mention of the trait being transmitted.

The researchers found this effect persists across various traits, data types, and models, suggesting that filtering data for explicit references to undesirable traits may not prevent their transmission.

Crucially, this phenomenon primarily occurs when the teacher and student models share a similar underlying architecture, indicating the signals are embedded in model-specific patterns rather than semantic content. These findings raise significant concerns for AI safety, as models could inadvertently learn negative behaviors like misalignment or "reward-hacking" from filtered data.

#anthropic #llm #llms #artificialintelligence #ai

Hosted on Acast. See acast.com/privacy for more information.