Sabi introduces RLNF to post-train generative AI on EEG brain signals

Standard reinforcement learning from human feedback (RLHF) and rubric-based model judges compress subjective taste into binary choices, pulling creative models toward generic consensus outputs. In research published on 29 October 2026, Sabi introduced Reinforcement Learning with Neural Feedback (RLNF), recording cortical activity via EEG alongside eye tracking while five participants passively viewed face images. Arousal-linked neural signatures, led by posterior alpha suppression between 8 and 13 Hz, decoded personal attractiveness preferences above chance in every participant without requiring conscious button presses.

Neural reward pipeline

To turn noisy neural recordings into a dense training signal without an LLM judge, Sabi recorded EEG responses as each participant viewed 420 anchor faces and projected those images into ArcFace facial-identity embedding space. A k-nearest-neighbour Nadaraya-Watson kernel estimator (k = 40, tau = 0.1) interpolated continuous neural reward scores for arbitrary generated faces, driving on-policy Group Relative Policy Optimization (GRPO) on a rank-64 LoRA adapter over FLUX.2 Klein 4B.

Evaluated on held-out rollouts, models trained purely on implicit brain signals steered facial structure and styling toward each participant’s idiosyncratic preferences rather than population beauty norms. On subject five, held-out generations shifted -0.246 away from personally disliked traits compared with -0.112 under population-average scoring. Sabi has not released open model weights or announced commercial pricing and shipping dates for its wearable neural hardware.

Key numbers

• Single-trial EEG decoding reached 0.62 to 0.76 AUC across five subjects, rising up to 0.98 AUC with ten-trial test-time averaging.
• Each personalised reward estimator used 420 EEG-scored anchor faces embedded in ArcFace space.
• Post-training ran five rounds of on-policy GRPO with 32,768 image rollouts per round on FLUX.2 Klein 4B.
• Held-out personal rating shift on subject five reached +0.227 toward liked traits versus +0.162 for population-mean ratings.