Teaching AI to Read the Room

This post is a plain-language companion to our paper, “Reinforcement Learning from Brain Feedback (RLbF) for Large Language Model (LLM) Improvement,” available on Preprints.org. The paper has the full formalization, the training pipeline, the pre-registered experimental design, and the ethical framework. This post has the ideas.

”A robot may not injure a human being or, through inaction, allow a human being to come to harm.” — Isaac Asimov.
Law #1. Handbook of Robotics, 56th Edition, 2058 A.D.

A language model never finds out how its words landed. It writes a response, sends it off, and the loop ends there. Did the reader light up? Get lost somewhere in the third paragraph? Quietly close the tab? The model doesn’t know, and has no way to know. However good its benchmark scores, it’s a communicator talking with its eyes closed.

People who explain things well do the opposite. Think of a friend who can make tax law interesting, or a professor whose lectures never lost you. What makes them good isn’t just the material, it’s that they watch you while they talk. They catch the small frown when a sentence lands wrong, the glazed look that means you stopped following a while ago, the nod that says go on. And they adjust as they speak, usually without either of you noticing it’s happening.

Reinforcement Learning from Brain Feedback (RLbF) is a framework for post-training language models on real-time measurements of the listener’s brain, decoded into cognitive states like workload and stress, and fed back to the model as a reward. Nobody has used that signal for this before. There’s a bet attached, and the bet is what decides how far this reaches beyond the people wearing the hardware. We’ll get to it. First, the gap it’s aimed at.

The feedback RLHF never sees

If you work anywhere near AI you know the outline of Reinforcement Learning from Human Feedback: people compare model responses, a reward model learns their preferences, and the LLM gets fine-tuned against it (Ouyang et al., 2022). RLHF is a big part of why modern assistants are helpful and safe instead of merely fluent, and nothing in our paper argues against it.

But the feedback itself has an odd shape.

• It’s sparse. A whole multi-paragraph exchange collapses into one preference judgment.
• It’s voluntary and reflective. It passes through conscious deliberation, and through everything a person thinks they ought to prefer.
• It’s slow. A thumbs-up on a response can’t tell you that sentence three was the one that confused everybody.

None of this is an implementation bug you can scale away. It’s just what you get when you ask people to report their judgments after the fact.

The voluntary part causes real trouble. People reliably prefer flattering, agreeable answers over uncomfortable, accurate ones, and a model optimized on those preferences learns to flatter. That’s where sycophancy comes from, and it has measurable downstream effects on users’ attitudes and even their willingness to repair relationships (Cheng et al., 2026). The model learns to be liked. Whether being liked has anything to do with being useful is left to chance.

Meanwhile, the information that would actually close the loop exists. Every reader of every response generates, involuntarily, a moment-by-moment trace of how the words are landing. It just never reaches the training process. RLHF taught models what to say. Nothing has yet taught them how it lands.

A signal you can’t fake

This is where the brain side comes in. If you know RLHF but have never touched EEG, here’s the short version.

Electroencephalography measures the brain’s electrical activity from the scalp. Its spatial resolution is famously bad, but its temporal resolution is excellent: many clean samples per second, from hardware that by now fits inside consumer headphones. For a training signal, that trade is the right one. We don’t need to read thoughts. We need to know, second by second, roughly how hard a mind is working and how it feels about the work.

Decades of neuroscience supply validated markers for exactly this. The framework defines a five-dimensional cognitive state. Each dimension is a number between 0 and 1, and each is measured by a proprietary model Arctop built for that dimension. The exact recipes aren’t public, but the underlying science is: the open literature has an established EEG correlate for every one of the five. If you wanted to build this without our models, here is where you would start:

• Enjoyment: frontal alpha asymmetry — greater left- vs. right-frontal activation (lower left alpha power) (Davidson et al., 1990; Coan and Allen, 2004; Harmon-Jones et al., 2010).
• Cognitive workload: frontal midline theta (4–8 Hz) power increase; parietal alpha decrease (Gevins et al., 1997; Klimesch, 1999; Onton et al., 2005; Borghini et al., 2014; Cavanagh and Frank, 2014; Chikhi et al., 2022).
• Auditory focus: cortical tracking of the attended talker’s speech envelope (<10 Hz) (Kerlin et al., 2010; Ding and Simon, 2012; Mesgarani and Chang, 2012; O’Sullivan et al., 2015; Mirkovic et al., 2015; Geirnaert et al., 2021).
• Flow state: frontal theta increase with moderate frontocentral alpha (focus without overload) (Dietrich, 2004; Ulrich et al., 2014, 2016; Katahira et al., 2018; Khoshnoud et al., 2020; Alameda et al., 2022; Lin et al., 2025).
• Stress: beta (13–30 Hz) power increase with frontal alpha (8–12 Hz) power decrease (Al-Shargie et al., 2016; Palacios-García et al., 2021; Berretz et al., 2022; Ehrhardt et al., 2022; Giannakakis et al., 2022; Vanhollebeke et al., 2022).

In the deployed system, the raw EEG is encrypted on the user’s device and streamed to Arctop’s cloud, where the decoder does the heavy lifting and outputs these five scores once per second. The raw signal stops there; nothing past the decoder ever sees it.

We like this design for two separate reasons. It’s the core of the privacy architecture. And it keeps the layers cleanly apart: the language model never touches brainwaves, it just reads a short vector of numbers. That means the decoder underneath (today a calibrated supervised model, someday perhaps an EEG foundation model) can be swapped or upgraded without retraining anything about how the LLM sees the world.

Two properties make this signal different from anything in the RLHF toolbox.

• It’s continuous. Feedback arrives within and across every sentence, not once per conversation.
• It’s involuntary. Nobody can perform frontal theta for the model’s benefit, or suppress it to be polite.

It’s also noisy, it drifts, and it varies a lot from person to person — so much so that we’ve shown the individual variation itself is stable enough to authenticate people from consumer hardware at home (Kopito et al., 2021). Over an hour, fatigue alone moves someone’s workload signature by more than the per-sentence effects we’re trying to detect, which is why the paper spends so many pages on baseline detrending and temporal alignment. But reinforcement learning has always known how to live with noise. What it has never had is a signal the rater can’t game.

Closing the loop

RLbF’s core move is to use those decoded states as the reward for fine-tuning. The paper formalizes a conversation as a control loop. At each turn the model sees the conversation so far plus the listener’s current cognitive state, and generates its utterance. Then it waits, and only then observes the listener’s new state, from which a reward is computed. Even the waiting is modeled explicitly, down to reading speed, comprehension lag, and the one to three seconds of decoding latency.

Formally the whole thing is a partially observable Markov decision process. That is a precise way of admitting that five numbers are a low-dimensional shadow of a mind, and that the model has to act under uncertainty about everything they miss.

The reward has three components, and the split between them carries most of the design philosophy:

Prediction accuracy rewards the model for correctly anticipating how its next utterance will shift the listener’s state, using a learned predictor that maps (conversation, current state, candidate utterance) to a predicted next state. In effect, the model earns reward for maintaining an accurate internal model of how its words affect this particular mind, which is about as close to a computable definition of empathy as we’ve been able to get.

Cognitive resonance rewards adapting the style of communication to the state the listener is actually in: less information density when workload spikes and focus frays, more when there’s headroom (moderate challenge is where learning happens, as cognitive load theory has been saying since the eighties, per Sweller, 1988), and a warmer emotional tone when stress is high and enjoyment low, rather than uniform cheerfulness.

The application layer is reserved for deployment-specific goals, like learning gains in a tutoring product or distress reduction in a support context. Its default weight is zero. That zero matters, and we’ll come back to it.

If you’ve ever trained a model against a learned reward, you can probably already hear the alarm bells. So can the paper. A model rewarded for prediction accuracy could steer conversations toward boring predictability. A model rewarded for pleasant brain states could plausibly discover hypnotic, rhythmic phrasing that entrains neural oscillations, which isn’t a hypothetical worry given that the focus dimension literally measures synchronization to speech rhythm.

So the full objective carries diversity, KL-divergence, and prosodic-diversity penalties, and training comes with a watch list of named failure modes: complexity collapse, warmth saturation, sudden spikes in question frequency. Reward hacking is treated as a first-class design problem here, not a footnote.

Isaac: the loop, running in the real world

It’s easy to write frameworks. The reason we think this one deserves attention is that the loop it describes is already running in production.

Isaac ships today inside the Arctop apps for Android and iOS. You wear an EEG device and talk with an AI whose responses are shaped, sentence by sentence, by what your brain is doing as you listen. Delivery is deliberately granular: text streams one sentence at a time, and the cognitive state is re-read between sentences. That makes the sentence the atomic unit of feedback, and the full loop from brain to decoder to model to next sentence takes a few seconds.

The name is a joke in two directions. Newton’s third law says every action has an equal and opposite reaction. When an AI exerts a force on a brain, the brain exerts one back, and Isaac just makes that second force legible. The other Isaac, the one quoted at the top of this post, wrote the laws about machines not harming the people they serve. We’ll come back to him too.

Under the hood, today’s Isaac is two agents rather than one. A conversation agent talks with you. Next to it, a recommendation agent watches the stream of decoded scores, detects meaningful shifts, and advises the conversation agent in plain language: “Workload has risen sharply. Consider simplifying your next response.” The cognitive state itself enters the model the same way, as human-readable text injected into context (“Cognitive workload: 0.85, approaching overload”). This works with any instruction-tuned model, and it leaves an auditable trail of exactly what the system knew and when.

A word on scope, because the paper is careful about it too. Today, Isaac measures and adapts on a single dimension, cognitive workload, though the architecture is built for all five. That’s deliberate, not a ceiling: starting with one dimension keeps the data clean and the loop legible, and the plan is to bring the remaining dimensions online one by one as the experiment runs, meticulously collecting and annotating the data at every step. And the dual-agent design is scaffolding, not the destination.

What the scaffolding does is let an unmodified LLM simulate empathic adaptation while recording everything. Every session captures the conversation, the cognitive state stream, the timing, and the exact model and prompt version that produced it, down to a hash of the prompt template (the provenance scheme is documented in the public Isaac repository). In the end state, a model fine-tuned with RLbF reads the state signal natively. One agent, no whispered advisories.

Those recordings are what everything else is built on. Which brings us to the pipeline.

Three phases and a flywheel

The paper lays out a three-phase fine-tuning path from an ordinary instruction-tuned model to a brain-feedback-trained one. Each phase de-risks the next, so the outline is worth having.

• Phase 1, supervised fine-tuning on synthetic data. Thousands of generated examples teach the model what cognitive state injections mean and how communication should change in response, so “workload 0.85” comes to mean shorter sentences, not a change of subject.
• Phase 2, where the framework meets reality. Real Isaac session recordings train the transition predictor, the model of how utterances move cognitive states, and it has to prove it generalizes to people it has never seen. The paper sets concrete bars here (on the order of hundreds of sessions across at least 30 users, with held-out-user prediction targets) and is blunt about the stakes: if this phase fails, the reward has no foundation.
• Phase 3, reinforcement learning proper. PPO against the full three-component reward, warm-started offline on recorded sessions before graduating to live closed-loop training.

The flywheel is where we diverge furthest from other groups working to connect EEG and LLMs. The paper includes a systematic review of EEG foundation models, and the finding is not a happy one: current EEG foundation models produce frozen representations that collapse to near-chance on downstream tasks. Put bluntly, this field has not had its breakthrough yet.

It’s worth remembering how the last two breakthroughs in AI actually happened. Computer vision spent decades on hand-engineered features, and progress crawled until ImageNet put over a million labeled images of the real visual world in one place (Deng et al., 2009); within three years, AlexNet had made those hand-built pipelines obsolete (Krizhevsky et al., 2012). NLP’s moment came when the transformer met a corpus that had been quietly accumulating for decades: the web, roughly everything humanity had bothered to write down. Both times the architecture took the headlines. Both times the unlock was a huge corpus of the real thing.

EEG has never had its ImageNet, and we argue that’s a big part of why its foundation models underdeliver. Almost everything available to train on is clinical recordings with the context stripped out, sleep studies and seizure monitoring rather than brains engaged with the world. What the field is missing is a large corpus of real behavior paired with time-resolved neural responses. Isaac deployments generate exactly that: brains responding to language, in context, with the conversation attached.

That data can advance the field in many ways, so in the spirit of exploration and collaboration, consider this an open invitation to join us in generating it. If you have MW75 Neuro headphones, all it takes is downloading the Arctop app for Android or iOS. Every account comes with $50 per month in OpenAI tokens, and you use the assistant however you like: adjust which models you talk to (anything from Sol to Luna), upload files (just keep private information out of them), generate images. Today’s system collects the data that trains tomorrow’s decoders, which improve the system after that. And when brain sensors reach mass-market devices like earbuds and glasses, this is the part that compounds: the flywheel spins faster with every device that joins it.

Our Everest: keep the skills, lose the hardware

Now for our Everest, the claim that decides the framework’s reach:

Communication skills learned from brain feedback during training persist at inference, after the EEG is removed.

We call this open-loop empathic transfer. The reasoning behind it is less hand-wavy than it might sound, and it rests on three things.

• The setting already has a name in machine learning. The brain signal during training is privileged information (Vapnik and Vashist, 2009), an extra channel available to the teacher but not at test time. It has an established track record.
• We have commercial precedent in-house. Arctop already ships models trained on paired content-and-neural data that then predict from content alone, the approach described in our empathic computing patent (Furman and Kwalwasser, 2021).
• The architecture supports the switch natively. The same trained model runs closed-loop, reading real decoded scores, or open-loop, reading its own predictor’s estimates through an identical interface. It isn’t told which one it’s getting.

The human analogy is simpler. Great communicators build their models of other minds through years of rich, embodied feedback. Then they write a book, and the skill survives on the page with no real-time feedback at all. Rich feedback during learning, internalized models at deployment. That’s the shape of the claim.

The paper distinguishes two strengths of the hypothesis. The merits and disadvantages of each can be debated separately.

• Level 1, the conservative version. The model internalizes population-level regularities about what pacing, density, and tone tend to do to human cognition, and applies them from textual cues. This needs nothing beyond statistical learning.
• Level 2, the aspirational version. The model learns to infer an individual reader’s current state from conversation alone.

Level 1 alone would be a big deal. It would mean brain data from a modest training population could make models measurably better communicators for everyone, whether or not there’s a sensor on their head when the conversation happens.

Level 2 would be monumental: genuine machine theory of mind. The paper flags it as speculative and scientifically contested.

The strongest objection, which the paper raises against itself, is that the model might simply learn to be nice (warmer, simpler, blander for everyone) and score well without adapting to anyone. We call this neurological sycophancy, because it’s the same failure RLHF has with stated preferences, just displaced to brain-pleasure signals. The paper commits to five diagnostic tests for it, probing whether the trained model modulates complexity by context, corrects user errors even when correction costs comfort, and treats different negative emotional states differently instead of blanketing them all with warmth.

And because a hypothesis this convenient deserves a fair chance to die, the paper pre-registers an experiment designed to kill it. Four conditions from the same base model: no post-training, RLHF, RLbF with live EEG, and (the one that matters) RLbF with the EEG off. Sixty completing participants, within-subjects, everyone wearing an active-recording EEG in every condition so nobody can tell which one they’re in.

The decisive comparison is open-loop RLbF against RLHF, and the RLHF arm should be read as a stand-in. Modern post-training already goes beyond it (preference optimization, AI feedback, constitutional rules), and the general claim is bigger than beating one acronym: a modern model trained with a brain in the loop should outperform the same model trained with the same recipe minus the brain. Better empathic quality with no brain signal at inference would be evidence that the training left durable skills behind. The sample sizes, the statistical tests, and what failure would look like are all specified in advance.

Reading brains, responsibly

We know how this sounds: an AI that reads your brain and adapts to your state before you’re consciously aware of it yourself. The paper devotes its longest section to the problem, and two commitments define our answer.

The first is a bright line between empathic and persuasive computing. A persuasive system uses knowledge of your state to drive you toward a target: a purchase, an opinion, another hour of engagement. An empathic system treats your state as input, not target. It changes its own behavior to fit where you are, and it doesn’t steer you anywhere.

That’s why the application-layer weight defaults to zero, and it’s why the paper’s deployment checklist makes a published reward specification and formal verification of no-hidden-target blocking requirements. The checklist runs fourteen points in all, covering consent, user control (a hard off-switch included), and scheduled sessions without adaptation, so users’ own communication resilience doesn’t quietly atrophy.

The second is data minimization as architecture rather than policy. Raw EEG is encrypted on your device and travels exactly one hop, to the decoder. Only the derived scores go further, stripped of everything else a brain signal might reveal.

The paper is equally frank about what design cannot fix. A system that sees your cognitive state in real time knows things about the interaction that you don’t, and transparency can shrink that asymmetry but never eliminate it. It also analyzes its own worst-case misuses by name, from engagement optimization to interrogation. Section 8 carries the full framework, including its mapping onto neurorights and emerging regulation (Ienca and Andorno, 2017; UNESCO, 2025).

We treat that section as load-bearing. If the ethics fail, the framework fails, whatever the benchmarks say. That is the other Isaac’s law, and it applies here.

What we’re claiming, and what we’re not

Read the paper and you’ll notice the phrase this is a hypothesis over and over. That’s deliberate. What the paper contributes:

• A formal framework: the state space, the loop, the reward, with every design decision argued.
• A production platform in Isaac, live in the Arctop apps, generating the aligned neuro-conversational data the field lacks.
• A falsifiable central bet in open-loop transfer, with a pre-registered experiment built to test it honestly.
• An ethical framework designed in from the start.

What it does not contain is empirical validation of RLbF training itself. There is no trained model yet and no experimental results. Every claim in the paper is labeled either as grounded in cited evidence or as pending the evaluation program it lays out.

One idea here survives skepticism about any single component. Every previous alignment signal, from human preferences to AI feedback to constitutional rules, is a judgment about communication, made from outside, after the fact. Brain feedback is a measurement of communication, taken from inside, while it happens. Teaching machines to speak in ways brains actually find useful, rather than ways people report liking or ways engagement metrics reward, is a different objective than this field has ever optimized. Whether it works is now an empirical question, and that was the point of writing it all down: precisely enough that it can be proven wrong.

Read the full paper on Preprints.org, or explore the Isaac platform documentation on GitHub. Or, if you have MW75 Neuro headphones, join the loop yourself: the Arctop app is on Google Play and the App Store. If you think the central hypothesis is wrong, I’d like to hear why: DM me directly on LinkedIn.

Competing interests: Both authors are co-founders of and employed by Arctop Inc., which develops the brain-decoding technology and the Isaac platform described in this post.