cs.AIOct 26, 2026
SaveToward Fast Human Reaction Modeling from High-Frequency Digital Interaction Data
Organizations: Mesocort AI Tovární 589, 709 00 Ostrava 9, Czechia · Faculty of Electrical Engineering and Computer Science VŠB–Technical University of Ostrava 17. listopadu 2172/15, 70833 Ostrava, Czechia
Abstract
We investigate whether limbic-like signatures of rapid, salience-driven behavior can be identified in high-resolution Counter-Strike 2 telemetry. This exploratory study introduces Ego-Knowledge, a leak-reduced reconstruction of each player's plausible perceptual state, supported by a millisecond-precision visibility engine and a reaction-centered extraction pipeline. Using a neuroscience-motivated 60--120 ms window, we derive coarse fight-or-flight labels from visibility- and damage-triggered events, with back-damage cases enabling constrained reaction-time analysis and manual label checks. We train two telemetry-native models using only information available by the trigger: EMBER (gradient boosting) and ORBIT (a custom transformer). On the full validation set, EMBER achieves 73.5% accuracy and ORBIT 71.2%, versus a 51.5% majority-class baseline, and both generalize to held-out players. The strongest LLM baseline achieves 62.7% on 1,000 validation cases; on these same cases, EMBER (73.6%) and ORBIT (70.9%) perform better (McNemar p=1.3e-9 and 1.5e-6, respectively), with orders-of-magnitude lower inference time and cost, though the comparison is not compute-matched. Including a short post-trigger window changes forecasting into early response recognition, raising accuracy monotonically to 85.15%. We propose this curve as a candidate probe of the transition from rapid to deliberative control, but roughly uniform per-tick gains do not establish that interpretation. The neuroscience framing is hypothesis-generating, not mechanistic: physiological ground truth, suitable control distributions, and reverse-inference limitations remain open. We outline extensions to auditory reconstruction, larger datasets, and telemetry-native foundation models.
Figures & tables
Fig. 1 : High-level architecture of the Mesocort pipeline. Raw CS2 demo telemetry is transformed into reconstructed game state, player-centric Ego-Knowledge, reaction windows, and fight/flight labels used for behavior modeling, prediction, and evaluation.
| Feature group | No. | Description |
|---|---|---|
| Round and map context | 5 | Round-level and map-level state variables. |
| Player position and state | 21 | Player position, view direction, equipment, movement, and status variables. |
| Aiming and visibility | 13 | Crosshair movement, enemy visibility, and enemy proximity features. |
| Player input actions | 12 | Button states and weapon-firing indicators. |
| Damage and kill events | 14 | Tick-level and round-level damage and kill history features. |
| Teammate features | 24 | Position and visibility features for four teammates. |
TABLE I : Summary of input feature groups used by the model.
| Statistic | Value |
|---|---|
| Core players selected for Phase 1 | 50 |
| Average demos collected per core player | 690 |
| Initial pool of matches | 35,000 |
| Time span covered up to | 7 months |
| Filtered demos in extracted dataset | 32,205 |
| Total reaction moments | 271,541 |
TABLE II : Phase 1 reaction dataset summary.
Fig. 2 : Visibility engine processing flow: on the left, the scene’s depth map showing the silhouette of the target player; on the right, the pixel segmentation result, where green marks the visible portion and red the portion obstructed by geometry (e.g., a wall), from which the per-frame visibility percentage is computed.
Fig. 3 : Overview of the methodology. Pre-reaction prediction uses only information up to the anchor moment; early response recognition may also use a short post-anchor window.
Fig. 4 : Manual annotation tool. The central view reconstructs the situation in two dimensions; the left panel shows synchronized temporal signals of the analyzed player; the remaining panels hold playback, annotation controls, and gameplay highlights.
Fig. 5 : ORBIT architecture. A sequence of tick-level tokens, optionally preceded by global context tokens (map, player, rank), is processed by a Transformer encoder; the final representation feeds a classification head that predicts the behavioral response.
| Parameter | Value |
|---|---|
| LEARNING_RATE | 0.02 |
| MAX_DEPTH | 8 |
| MIN_CHILD_WEIGHT | 5.0 |
| SUBSAMPLE | 0.8 |
| COLSAMPLE_BYTREE | 0.8 |
| REG_LAMBDA | 1.0 |
TABLE III : Hyperparameter configuration of EMBER
| Setup | Train | Val. | Test |
|---|---|---|---|
| Player-disjoint | 16,609 | 2,993 | 8,730 |
| Extended player-disjoint | 223,169 | 39,642 | 8,730 |
| Train/validation only | 230,547 | 40,994 | – |
TABLE IV: Number of rows in train, validation, and test splits.
| Parameter | Value |
|---|---|
| Task | Binary classification |
| Target classes | fight , flight |
| Loss function | Cross-entropy loss |
| Token dimension | 256 |
| Transformer layers | 2 |
| Attention heads | 8 |
TABLE V : ORBIT hyperparameters used in the experiments.
Fig. 6 : Histogram of reaction times for events in which the player is shot in the back. Base data granularity is 15.625 ms. The distribution provides an initial descriptive view of short-latency behavioral structure; number of situations is 20,114.
Fig. 7 : KDE plots comparing reaction times based on different sources of reaction (KDE plot follows the source histogram).
| Sequence-history ticks | Accuracy (%) |
|---|---|
| 7 | 71.20 |
| 14 | 70.98 |
| 21 | 70.96 |
| 32 | 70.58 |
| 64 | 68.93 |
| 192 | 68.32 |
TABLE VI : Validation accuracy of ORBIT models trained with different sequence-history lengths in the pre-reaction prediction setting.
Fig. 8 : Training and validation loss of the selected 7-tick ORBIT model in the pre-reaction prediction setting over 50 epochs. The validation loss reaches its minimum around epoch 30 and then increases while the training loss continues to decrease, indicating overfitting in the later stages of training. 4 4 4 We hypothesize that the early onset of overfitting is partly due to the limited size of the training set, since Transformer architectures are known to be data-hungry and to generalize substantially better as the training data grows [ 60 , 61 ] .
| Strategy | Accuracy (%) | Avg. time (s) | Total Cost ($) | Precision (%) | Recall (%) |
| Gemini 3.7 Flash Medium Narrative (1/8) | 62.7 | 1.81 | 4.70 | 59.66 | 77.51 |
| Gemini 3.7 Flash Medium Full-Dump (3/4) | 61.6 | 6.31 | 60.03 | 58.19 | 81.33 |
| Gemini 3 Flash Medium Full-Dump (1/8) | 59.7 | 2.97 | 14.72 | 56.81 | 79.52 |
| Gemini 3 Flash Medium Full-Dump (1/4) | 57.3 | 3.28 | 15.97 | 55.42 | 72.89 |
| Gemini 3 Flash Medium Full-Dump (3/4) | 57.4 | 2.10 | 40.11 | 55.36 | 74.70 |
| Gemini 3 Flash Medium Narrative (3/4) | 60.3 | 5.07 | 16.98 | 57.69 | 76.10 |
TABLE VII : Performance comparison of prediction strategies; the LLM and Jev rows are over 1,000 predictions from the current validation set.
Fig. 9 : Validation accuracy of ORBIT in the early response recognition setting as the number of future ticks increases. The history window is fixed to 7 ticks, while the future window is extended up to 64 ticks, corresponding to approximately one second after the anchor moment.
Fig. 10 : Validation accuracy of EMBER in the early response setting as a function of future time steps (maximum difference in ticks between stimulus and the latest moment in the model input) and time step (in ticks). The best accuracy for each future time step is highlighted in red.
Fig. 11 : Global perturbation-based feature importance of ORBIT for samples predicted as fight and flight . Importance is measured as absolute change in predicted-class margin, so the plot should be interpreted as model sensitivity rather than causal evidence.
| Quantity | Value |
|---|---|
| Manually inspected situations | 200 |
| Annotators per situation | 7 |
| Majority-vote threshold | |
| Situations with majority-vote label | 165 |
| Number of situations used for EMBER validation | 119 |
| Consensus coverage | 82.5% |
TABLE VIII : Summary of majority-vote label validation. Agreement with the automatic labeling system is computed on the majority-vote gold-standard subset.
Fig. 12 : Correctly classified fight example: the analyzed player spots the opponent on lower middle and commits to the duel.
Fig. 13 : Incorrectly classified flight example: the analyzed player, with an AK-47 , faces an AWP in a window but continues the duel.
| Component | Est. (ms) |
| : server player | |
| Server client (one-way) | 12–18 |
| Client processing + render pipeline (200–400 FPS) | 6–15 |
| Display (144–240 Hz scan-out + pixel response) | 5–12 |
| Interpolation (enemy-in-view trigger only) | 16–31 |
| : player demo | |
TABLE IX : Estimated CS2 system-latency contribution to demo-measured reaction time (typical European competitive conditions). Values are order-of-magnitude estimates, not measurements.
Fig. 14 : Visualization of reaction time transformation based on estimated unmeasured delay and measured ping. Top row = raw reaction time; bottom row = reaction time minus ping and lower band, with shifted upper band. Minimum reaction time filter before adjustment: 46.875 ms (3 ticks)
| Feature | Description | Encoding / scaling |
|---|---|---|
| Round and map context | ||
| Premier rank | Player Premier rank. | Min–max scaler |
| Round number | Current round index. | Embedding layer |
| Map ID | Identifier of the map being played. | Embedding layer |
| Bomb planted | Indicates whether the bomb is currently planted. | Boolean |
| Remaining round time | Remaining time in the current round. | Standard scaler |
TABLE X: Description of the input features used by the model. All features are related for the current game tick.
Fig. 15 : Complete ORBIT global feature-importance ranking. Feature importance is measured as the mean absolute change in the predicted fight or flight class margin after replacing each feature with a neutral baseline value. Features are ordered by decreasing importance.
Fig. 16 : Complete global feature-importance ranking for EMBER. Sorted by gain. EMBER uses a subset of previously laid out feature set.
Fig. 17: General serialization pipeline from the source dataset row to LLM input. All benchmarks share row preparation; encodings differ in how features are rendered. All runs attach labeled train examples and one unlabeled query; the model returns structured JSON with label and reasoning .
| Pattern | Description |
|---|---|
| Fight patterns | |
| The Stationary Hold | Remaining stationary with a primary weapon (especially AWP ) while the crosshair is aligned with a potential entry point. |
| Aggressive Pursuit | Moving quickly toward an enemy or tracking a retreating enemy with a combat-ready weapon. |
| Tactical Ambush | Engaging an unaware enemy from a position of advantage, often described as "stalking" or "initiative" . |
| Weapon Switching | Transitioning from a knife or utility to a primary weapon immediately upon spotting an enemy. |
| Flight patterns | |
TABLE XI : Fight-or-flight patterns in LLM reasoning