What if AI learned your tastes by observing your brain's reactions?

Sabi Research is experimenting with RLNF, a method that uses brain signals to train AI to generate images tailored to individual preferences.

A face image gradually transforms without requiring any text prompts to guide its direction. The changes follow an individual's aesthetic preferences, inferred from their brain activity. This demonstration accompanies research by Sabi Research, which is exploring a new generative model training method dubbed Reinforcement Learning with Neural Feedback (RLNF). The idea consists of replacing explicit user feedback with neural signals to personalize artificial intelligence creations.

Presented by researcher Rahul Chhabra, this approach does not involve reading minds directly. It leverages correlations between certain physiological responses and the preferences expressed by participants. Initial results focus on face generation, a domain where aesthetic judgments vary significantly from person to person. Learning What Someone Prefers Without Asking for Their Opinion

Generative models are typically fine-tuned using human evaluations, comparisons between multiple outputs, or systems designed to assign scores to generated content. While these methods improve the average quality of responses, they tend to favor collective preferences over an individual's specific tastes.

The RLNF method explores another possibility. Researchers use electroencephalography (EEG), which measures the brain's electrical activity via electrodes, to observe participants' reactions to different images. The goal is to detect signals associated with their personal appreciation and then convert them into rewards used during the training of a generative model.

In the study published by Sabi Research, five volunteers viewed a series of faces while rating their attractiveness on a scale from 1 to 7. Researchers recorded their brain signals alongside eye-tracking data to identify the physiological characteristics associated with positive and negative evaluations.

The system manages to distinguish highly appreciated faces from lower-rated ones, with performance varying across individuals. On an AUC metric—where 0.5 corresponds to chance and 1 to perfect discrimination—individual results ranged from 0.62 to 0.76 for a single trial. When averaging ten trials, they reached between 0.80 and 0.98. These figures relate to the ability to distinguish preferences within this specific experiment, rather than a general reading of emotions or thoughts. When Brain Signals Train FLUX

Once preferences are estimated, researchers use them to fine-tune FLUX.2 Klein 4B, the image generation model developed by Black Forest Labs.

The method relies on a numerical representation of faces and a comparison system with previously evaluated examples. Each new image can thus receive a score estimating its proximity to the participant's neural preferences. This score acts as a reward during reinforcement learning, without requiring the user to rate every single generation produced throughout training.

The teams conducted five training cycles using LoRA adaptations, generating 32,768 images per cycle. The resulting models were then evaluated on prompts and configurations not used during training to observe whether the modifications persisted beyond the initial examples.

The results show a progressive shift in the features of generated faces toward estimated individual preferences. For one participant, researchers observed a 0.335-unit increase on their EEG reward metric, compared to 0.162 for average population preferences. The participant's personal evaluations also shifted in the expected direction.

Control experiments were conducted using randomly assigned rewards. These did not produce the same changes, supporting the hypothesis that the results indeed stem from the information extracted from EEG measurements rather than the training process alone. Personalization Beyond Average Preferences

Distinguishing individual tastes from collective preferences is a key challenge in this work. An image deemed appealing by a majority of users does not necessarily match everyone's expectations. This gap applies to faces, but also to abstract art, voices, architecture, or musical compositions.

Researchers notably draw on studies dedicated to individual aesthetic preferences, which show that the personal component of judgment varies considerably across domains. According to estimates cited by Sabi, it accounts for approximately 43% of face appreciation and 74% for abstract art.

This diversity poses a challenge for models trained on vast datasets, whose outputs tend to converge toward broadly liked characteristics. Neural feedback could provide more specific information about what truly appeals to an individual, without relying solely on their descriptions or explicit choices.

However, the experiment remains limited to five participants and face generation. It does not prove that the method would work with the same precision on other creative forms. Researchers also acknowledge that some signals stem from eye movements rather than brain activity alone, and that results are not uniform across participants. Creative Possibilities That Also Raise Ethical Questions

Such a method could eventually power systems capable of adapting their output to users' physiological reactions. Rahul Chhabra specifically mentions wearable brain-computer interfaces, paired with cameras and microphones, which could collect preference data throughout daily experiences.

In creative fields, this prospect could apply to images, music, video, or personalized user interfaces. Nevertheless, these remain research avenues: the published experiments do not yet demonstrate continuous personalization of these various media under real-world conditions.

Researchers also highlight the risks associated with a more precise understanding of individual preferences. Applied to faces, this technology could heighten concerns surrounding physical appearance or encourage increasingly homogenized aesthetic standards. Paired with advertising content, it could also allow messages to be tailored to trigger specific emotional responses.

Th