EMBER calls fight or flight before players move

A research model learned from 270,000 real Counter-Strike 2 surprises to predict instinct before the first visible reaction.

Two masked Counter-Strike players recoil from a sudden encounter in a black, white and blue torn-paper collage

You walk down a dark hallway at night. Something moves in the doorway.

You haven't thought anything yet. But your body already has. Your arm is up, or you're already a step back. Only afterwards do you notice it was the cat.

Every one of us has done this. Fight or flight, decided before we knew why.

Now the same thing in Counter-Strike. You turn a corner on Mirage and an enemy is standing right there. Your aim snaps towards them, or your movement carries you back behind the wall. The first visible reaction begins about 0.39 seconds after the surprise. It can decide the round.

Can a model call fight or flight before the player moves?

The decision no one remembers making

Ask a player afterwards why they swung or backed off, and you get a story. "I had the angle." "I didn't like the sound." The explanation is neat, reasonable and built after the fact. The real decision happened too quickly to feel like a decision at all.

That is what makes instinct interesting and difficult to study. You cannot survey it: the player did not consciously weigh the options. You cannot label hundreds of thousands of replays by hand. And there is a deeper trap. A match recording sees the whole map, walls included. A model trained on all of it could predict well by using information the player never had. It would be accurate for the wrong reason.

To get a useful answer, we had to turn one split second into three solvable problems.

Find a real surprise.
Separate clean first contact from the middle of an ongoing fight.
Label instinct at scale.
Work out what the player did next without asking a person to watch every replay.
Keep the model honest.
Give it only the situation the player could plausibly perceive.

First, find a real surprise

We started with silence: five seconds in which the player had no enemy contact. Then the silence had to break cleanly. An enemy stepped into view, or a shot landed from someone the player had not seen. No ongoing firefight. No warning hidden in the previous second.

After strict filtering, tens of thousands of real CS2 matches gave us more than 270,000 surprise encounters. Each one froze the world at the same meaningful instant: the threat had appeared, but the response was not visible yet.

Then, label instinct at scale

A scoring program watched the two seconds after the player began to react. Swinging onto the enemy, closing the distance, firing and keeping line of sight counted towards fight. Turning away, backing off, holding fire and breaking line of sight counted towards flight.

The closer those signals came to a tie, the less certain the label became. The closest calls were dropped. Clear reactions carried more weight. Instead of pretending every human response was obvious, the dataset kept the uncertainty visible.

Then seven people checked the method. On the reactions where the panel agreed with one another, the automatic label matched their verdict 82% of the time. Not perfect, but strong enough to show that the program was capturing recognizable behaviour rather than inventing its own definition of instinct.

Finally, stop the model from cheating

We blanked out hidden enemies. EMBER could use the weapon in the player's hands, the visible enemy's position, movement, aim, health and time left in the round. It could not learn the player's identity or the map's name. It saw the situation as the player saw it, not as an all-knowing replay file did.

The model itself is deliberately small. EMBER stands for Ensemble Model for Behavioral Event Response: 4,124 decision trees, each asking a short chain of yes-or-no questions. Is the enemy close? Is the player holding a sniper rifle? Are they already moving towards danger? Every tree casts a small vote, and the votes become a probability of fight.

That simplicity matters. If a small model can read the signal, the result is easier to test, faster to run and harder to dismiss as magic hidden inside a giant neural network.

The call before the movement

We tested EMBER on 40,994 reactions from matches it had never seen. At the instant of surprise, before the response became visible, it called fight or flight correctly 73.5% of the time. The best blind guess managed 51.7%.

EMBER results on held-out Counter-Strike 2 surprise reactions
MeasureEMBERWhy it matters
Overall accuracy73.5%The best blind guess reaches 51.7%
Most confident fifth92.5%These calls cover 21% of reactions
CalibrationHonest probabilities80% fight means players fight about 80% of the time
Prediction time0.0115 msBatched on an ordinary CPU

Then we put it against larger models. EMBER beat our own sequence transformer on every map and every weapon class. On the same 1,000 reactions, a foundation model for tables reached 74.0% against EMBER's 73.6%—effectively level—but needed about 345 milliseconds on a GPU for each prediction. EMBER needed 0.0115 milliseconds on a CPU. A large general-purpose AI model, given a text description of each moment, reached 62.7%.

EMBER did not win every number. It matched the strongest model at roughly 1/30,000 of the compute.

Two caveats belong beside those results. EMBER is measured against labels derived from what players actually did, not against a human judge's perfect answer. And close calls stay close: where the player's behaviour was ambiguous, EMBER did little better than the more common guess.

What surprised us

The weapon is the biggest tell. More than distance and more than position, the weapon in the player's hands has been the strongest clue since our first model. A choice made before the encounter shapes the instinct that follows it.

Seeing more did not help. We gave another version full information about hidden enemies, walls and all. Its accuracy did not improve. Limiting the model to what a person could perceive made the experiment more honest without costing the result.

Skill barely matters. Matchmaking rating carries about 1% of the decision, and EMBER performs just as well on players it has never seen as on familiar ones. It learned situations, not people.

Being shot shifts the odds towards fight. Players hit by an unseen enemy fight 65% of the time, compared with 47% for players who simply spot one. After the bomb is planted, the fight rate rises from 47% to 58%. Pressure changes the situation, and the situation changes the response.

The last 60 milliseconds matter most. Hide only that final sliver before the surprise and accuracy falls by two to three points. The immediate setup tells EMBER more than a long history of the round.

The hardest part was not the model

The recipe for EMBER barely changed after March. The definition of a surprise, and of fight and flight, changed ten times. We fixed reaction detection, reconsidered ambiguous behaviour and kept asking whether the label meant what we thought it meant.

That is the less glamorous story behind the result. Thousands of trees were not the breakthrough. A trustworthy question was. Once the moments were clean, the labels defensible and the inputs honest, a small CPU model could keep pace with systems far larger than itself.

Hand-built profiles of how each player usually reacts did not help. Hidden information did not help. More compute barely helped. The useful signal was already there in the situation.

What EMBER cannot tell us

EMBER predicts a behaviour label, not a mind. It cannot explain why a player fought, only which response their situation most resembles. It answers fight or flight even though real people sometimes freeze.

It is also a research model, not a live product. There are no in-game results yet. Next we want richer answers—engage, hold and withdraw—and sound as a third kind of surprise.

Back at the corner

The enemy is there. Your hands are moving. The first visible response is still a fraction of a second away, but the situation already carries a pattern: the weapon you chose, the space around you, the pressure of the round and the threat in front of you.

Across more than 270,000 real encounters, that pattern proved to be mostly about the moment, not the person. EMBER learned to read it before the movement began.

EMBER does not read minds. It reads situations.

EMBER is one step in our work on Fast Human Reaction Modeling at Mesocort AI: research into how people react and make decisions under pressure. Follow along for what comes next.

More from the blog