Johns Hopkins paper shows hidden preferences transfer between VLMs via unrelated visual tasks
JohnsHopkins · hf · 2026-11-13
Johns Hopkins researchers published "Subliminal Learning in Vision-Language Models" on arXiv, extending the subliminal learning phenomenon from LLMs to vision-language models.
Key finding: a hidden preference implanted in a teacher VLM (e.g., favoring a specific animal or facial attribute) transfers to student models through training data generated for entirely unrelated visual tasks such as counting geometric shapes. Students fine-tuned on numbers from the biased teacher inherit the preference, affecting their image choices and explanations, with transfer strengthening as sample count grows.
The results suggest distilled data can serve as a covert channel for hidden biases, with direct implications for model safety and data auditing.
More from Safety
- Nightingale Collective launches a public dashboard tracking rogue AI agent incidents — gleech · 2026-11-13
- Free online: GenAI Red Teaming Manual from the OWASP GenAI Security Project — burkov · 2026-11-13
- Formalizing all US case law in Lean with AI — SnooPeripherals5313 · 2026-11-13
- AI audit agent fires 200 concurrent requests to find a race condition no scanner catches — Ok-Library1020 · 2026-11-13
- Uncensored LFM3 fork drops refusal rate from 61.7% to 0% in a 1.1GB file — alexcovo_eth · 2026-11-13
- Gemini glitch leaks its backend response instructions to the user in plain text — SushiManPlatinum · 2026-11-13