Introducing RL with Neural Feedback (RLNF). We have found reliable...
Rahul Chhabra@rahulchhabra07
22 views
Oct 30, 2026
~3 min read
Advertisement
1
Introducing RL with Neural Feedback (RLNF). We have found reliable ways to train AI models to respond to human preferences purely from brain signals.
The example below is a picture of a face that self-transforms into a more attractive version for the subject as they are looking at it.
The example below is a picture of a face that self-transforms into a more attractive version for the subject as they are looking at it.
2
How do you get an AI model to make something beautiful for you?
Today non-verifiable capabilities are tuned with either rubric rewards or reward models trained from human feedback, both of them are lossy.
We have finally been able to use arousal signals from the brain to create implicit rewards for personal preferences.
This means personal feedback from each user can be gathered passively with a BCI wearable without explicit user input.
More on the work here
sabi.com/blog/measuring…
Today non-verifiable capabilities are tuned with either rubric rewards or reward models trained from human feedback, both of them are lossy.
We have finally been able to use arousal signals from the brain to create implicit rewards for personal preferences.
This means personal feedback from each user can be gathered passively with a BCI wearable without explicit user input.
More on the work here
sabi.com/blog/measuring…
0:39
3
In most aesthetic areas, personal preferences deviate quite a lot from population level preferences that internet scale AI models converge to.
This leaves people with poor satisfaction when trying to use these models for tasks that require taste.
So far, there was no way to solve this since getting reliable preference measures from every user is not possible at scale.
This leaves people with poor satisfaction when trying to use these models for tasks that require taste.
So far, there was no way to solve this since getting reliable preference measures from every user is not possible at scale.
4
Unlike DPO which is generally used to tune text-to-image models, our paradigm is on-policy and doesnt require a preference dataset.
We show that even a simple embedding + k-NN based reward model works well so no explicit reward model has to be trained.
The models move personal preference more strongly than population level preference showing individual preference is captured from EEG.
We also see seed stability across runs where the model converges to similar levels across seeds.
Randomizing the rewards shows that the rewards don’t move, indicating that the reward signal does come from EEG and is not an artifact of training.
We show that even a simple embedding + k-NN based reward model works well so no explicit reward model has to be trained.
The models move personal preference more strongly than population level preference showing individual preference is captured from EEG.
We also see seed stability across runs where the model converges to similar levels across seeds.
Randomizing the rewards shows that the rewards don’t move, indicating that the reward signal does come from EEG and is not an artifact of training.
6
We wanted to be careful about this release given the societal impact of AI that can change faces.
It can cause negative impacts like amplifying body dysmorphia, and narrowing beauty standards, can allow a bad actor to actively influence audience feedback.
It can cause negative impacts like amplifying body dysmorphia, and narrowing beauty standards, can allow a bad actor to actively influence audience feedback.
7
We wanted to talk about this publicly and share how we're thinking about responsibility that comes with such capabilities.
We are focused on how to have higher standards on user consent and are investing in edge inference, and training on encrypted neural data.
We have some interesting work here that I can’t wait to talk more about.
We are focused on how to have higher standards on user consent and are investing in edge inference, and training on encrypted neural data.
We have some interesting work here that I can’t wait to talk more about.
8
BCI is moving away from clinical applications to consumer uses, and applications like these will become more abundant w time.
Robust and dependable regulation on user safety and standards of use on neural data can do great work in setting such standards for safety that allow users to trust such systems.
If brain signals can be used to understand private thoughts and personal preferences, users will want the service providers to be more trustable, and sufficient regulation to enforce the same.
Robust and dependable regulation on user safety and standards of use on neural data can do great work in setting such standards for safety that allow users to trust such systems.
If brain signals can be used to understand private thoughts and personal preferences, users will want the service providers to be more trustable, and sufficient regulation to enforce the same.
9
All that said, I'm generally excited about this new paradigm.
It gives us a whole new axis to scale AI models - towards personalized responses.
Which is more impactful to large scale adoption than just intelligence.
If AI can understand you, it can make things more beautiful for you. We are hoping to help make a more beautiful world for everyone.
It gives us a whole new axis to scale AI models - towards personalized responses.
Which is more impactful to large scale adoption than just intelligence.
If AI can understand you, it can make things more beautiful for you. We are hoping to help make a more beautiful world for everyone.
0:39