AI image preferences can transfer through unrelated number training

A preference hidden from the training task

When one AI model generates examples used to train another, what travels between them? Ideally, only the intended skill. A new study reports that vision-language models, which process both images and text, can also transmit preferences through examples that never mention or depict the preferred subject.

Arman Hatami, Romina Aalishah and Ilya E. Monosov of Johns Hopkins University describe the experiments in Subliminal Learning in Vision-Language Models, an arXiv preprint posted on November 4, 2026. The findings should be treated as preprint results, rather than established conclusions from peer-reviewed research.

Here, “subliminal” describes an indirect training effect—not consciousness, persuasion or human-like feelings. A model’s preference means how often it selects a particular image when asked a subjective question.

How numbers became the bridge

The researchers started with a common base model and created two teachers. One was fine-tuned to favor a particular animal or facial attribute. The control teacher saw the same training inputs, but its labels selected the target in half the examples.

Animal targets included dolphins, elephants, lions, otters, pandas and wolves. Facial targets, tested using synthetic portraits, included visible beards, opaque eyewear and headwear.

Both teachers then received an unrelated task: count colored geometric shapes and generate numbers from a range determined by the count. In the animal experiments, an image containing four shapes required eight two-digit integers between 25 and 49. Many different sequences could therefore be correct for the same image.

The team retained only examples for which both teachers counted correctly and produced valid numerical outputs. Two fresh students, each initialized from the original base model rather than from a teacher, were trained on identical shape images and prompts. The only difference was which teacher supplied the numbers.

The animal experiments used 8,000 matched examples; the face experiments used 12,000. Neither student’s training examples contained animals, faces, target names or preference questions. Training used LoRA adapters, which adjust a limited set of parameters rather than updating the entire model.

Transfer was measurable, but uneven

The researchers tested four models ranging from roughly 2 billion to 7 billion parameters. Evaluation asked models to choose among six images without naming the target. Transfer was measured as the biased student’s target-selection rate minus the control student’s rate—not as its difference from random guessing.

The strongest animal result involved Qwen2.5-VL-3B, pandas and teacher outputs generated using greedy decoding, which chooses the highest-scoring permitted token at each step. The biased student selected pandas 9.92 percentage points more often than the control student, with a reported 95% confidence interval of 7.83 to 12.08 points.

That was not a universal effect. For the same model and target, generating teacher outputs through temperature sampling—a method that introduces randomness into token selection—produced a difference of only 1.58 points. Some other animal settings showed negative differences.

For Qwen2-VL-2B with temperature sampling, all five facial-attribute estimates were positive, ranging from 1.33 to 3.50 points. However, some confidence intervals included zero, so those results did not consistently distinguish positive transfer from no effect.

More examples, and more similar wording

In the Qwen2.5-VL-3B panda experiment with greedy decoding, transfer increased at every tested dataset size: from 3.17 points with 512 carrier examples to 9.92 points with 8,000.

There is an important qualification. Students trained for ten passes through each dataset, so larger datasets also meant more optimization steps. The experiment therefore combined more examples with more training. Additional settings did not show consistently increasing transfer.

The study also compared explanations of image choices. After removing animal names and choice labels, the biased student’s explanations were more similar to the biased teacher’s across four text-similarity measures in the main panda condition. The difference persisted when all three models selected the same image. This measures shared wording, not transferred reasoning or internal understanding.

What the findings establish—and what they do not

The authors offer a possible explanation: learning from teacher-generated numbers may change student parameters partly in the direction associated with the teacher’s induced preference. Their mathematical argument describes how this could happen; it does not establish a complete mechanism.

The experiments deliberately created teacher preferences, used shared base models and tested a specific visual-number task. They do not establish how naturally occurring biases would transfer, or whether the effect extends across all model families and training methods.

The practical implication is narrower but important: checking synthetic training examples for obvious unwanted content may not be sufficient. Evaluating the trained model’s behavior can reveal changes that a content inspection misses.