📄 animal behavior and cognition

Low-dimensional latent spaces identify the functional structure of individual behavioral phenotypes

This paper introduces a multi-domain framework that extracts stable, low-dimensional latent representations from professional Counter-Strike 2 telemetry to identify interpretable behavioral phenotypes and predict individual performance transfer across diverse contexts.

Original authors: Higashi, H.

Published 2026-04-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Higashi, H.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you are trying to understand a person's true personality. You could watch them in just one situation—say, playing a video game as a "sniper." But that might just tell you how they act when they are hiding in a corner. To really know them, you need to see how they act when they are rushing into battle, when they are defending a base, and when they are playing on a completely different map.

This paper is about a new way to find a person's "behavioral fingerprint" by watching them in many different situations at once.

Here is the breakdown of the study using simple analogies:

1. The Problem: The "Chameleon" Effect

Humans are like chameleons. We change our colors (behaviors) to fit our environment. If you only watch a professional Counter-Strike 2 player on one specific map, you might think they are very aggressive. But if you watch them on a different map, they might be very cautious.

Traditional computer models try to guess what a player will do next, but they often get confused because they mix up the player's true personality with the rules of the specific game map. It's like trying to guess someone's cooking style by only watching them make pasta; you don't know if they are a "spicy" cook or a "salty" cook until you see them make a salad, a steak, and a soup.

2. The Solution: The "Universal Translator"

The researchers built a smart AI system that acts like a Universal Translator for human behavior.

Instead of learning a player's style for one map, the system watches the player on three different maps (Dust II, Mirage, and Inferno) and in two different roles (attacker and defender) all at the same time.

The Encoder (The Detective): This part of the AI watches the player's movements for a minute. It asks, "What is unique about this person, regardless of where they are?"
The Combiner (The Blender): It takes the clues from all the different maps and roles and blends them together. It filters out the "noise" (like the specific layout of the map) and keeps only the "signal" (the player's core style).
The Result: A tiny, compressed ID card (a mathematical code) that represents the player's soul.

3. The Magic Trick: Zero-Shot Transfer

The coolest part of the study is what happens next. The AI creates this "ID card" using data from Maps A, B, and C. Then, it is asked to predict how that player will move on Map D, a map it has never seen before.

The Old Way: If you only knew how a player acted on Map A, you'd be terrible at guessing their moves on Map D.
The New Way: Because the AI extracted the player's core personality (their ID card), it can predict their moves on the new map with surprising accuracy. It's like knowing a person is "cautious and team-oriented" based on their behavior in a board game, and then correctly guessing they will be cautious and team-oriented in a new video game.

4. The "Two-Dimensional" Discovery

The researchers found something amazing: You don't need a massive, complex file to describe a person's strategy. You can compress their entire complex playing style into just two numbers (a 2D coordinate).

Think of it like a GPS map of personality:

Axis 1 (Left to Right): Represents Risk. On the left, you have players who are "Cowboys" (high risk, solo, aggressive). On the right, you have "Guardians" (low risk, team-focused, cautious).
Axis 2 (Up and Down): Represents Speed vs. Coordination. At the bottom, you have "Rushers" (fast, chaotic movement). At the top, you have "Tacticians" (slow, coordinated movement with teammates).

Every player sits somewhere on this map. The AI learned that if you know where a player sits on this 2D map, you can predict almost everything they will do, even in a new environment.

5. Why This Matters

This isn't just about video games. It proves that we can find stable, unchangeable parts of human personality even when people are constantly adapting to new situations.

For Science: It helps us understand that our "true self" is a stable core that shines through, even when we are wearing different "costumes" for different jobs.
For AI: It shows that we can build smarter robots or virtual assistants that understand who you are, not just what you are doing right now. They can adapt to you instantly because they know your "behavioral fingerprint."

In a nutshell: The researchers taught a computer to ignore the "where" and the "what" of a situation to focus entirely on the "who." By doing so, they found that human strategy is simpler than we thought, and we can predict it with a tiny, two-number code that travels with us anywhere.

1. Problem Statement

The central challenge in behavioral modeling is extracting stable, domain-invariant individual traits from behavior observed across diverse contexts.

The Limitation of Current Approaches: Traditional computational models often entangle individual differences with specific environmental structures or role requirements, making it difficult to generalize behavior to unseen conditions. Single-domain approaches (training on one specific map or role) fail to distinguish between a person's core behavioral tendencies and their reactions to specific contextual constraints.
The Gap: While psychometric tools exist, they often lack ecological validity in high-dimensional, dynamic environments like video games. Conversely, data-driven approaches using telemetry often struggle to separate "style" from "context."
Goal: To develop a framework that isolates intrinsic individual behavioral signatures (phenotypes) from context-dependent adaptations, enabling robust prediction of behavior in novel environments and roles.

2. Methodology

The authors propose a Multi-Domain Latent Representation Framework tested on professional Counter-Strike 2 (CS2) gameplay data.

A. Dataset and Task Setup

Data: 268 professional matches (5,756 rounds) from three maps: Dust II, Mirage, and Inferno.
Domains: Defined as unique combinations of Map ( $m$ ) and Strategic Side ( $s \in \{T, CT\}$ ), resulting in 6 distinct domains (e.g., DUT = Dust II + Terrorist).
Subjects: 91 unique professional players who participated in multiple matches across all map-side combinations.
Input: Time-ordered 2D coordinates (player position and ally center of mass) sampled at 1 Hz.
Split: Matches were partitioned into source sets ( $M^{(1)}$ ) for encoding and disjoint target sets ( $M^{(2)}$ ) for evaluation to ensure zero-shot transfer capability.

B. Model Architecture

The framework employs an Encoder-Decoder architecture with HyperNetworks:

Domain-Specific Encoders: Process 60-second historical trajectories from multiple source domains to extract initial latent representations.
Combiner: Aggregates representations from multiple source domains (via averaging) to create a domain-invariant latent code ( $z_i$ ). This step is crucial for filtering out domain-specific noise.
Shared Latent Code ( $z_i$ ): A low-dimensional vector (optimized at $d=2$ ) representing the player's stable behavioral signature.
Domain-Specific Decoders (HyperNetworks): A set of networks that map the shared latent code $z_i$ to the parameters ( $\theta_{pred}$ ) of a trajectory predictor specific to the target domain. This allows the model to adapt the same core "personality" to different maps/roles.
Trajectory Predictor: A sequential model (GRU + FC) that predicts the player's next position ( $t+1$ ) given a 5-second context window, conditioned on the decoder-generated parameters.

C. Training Strategy

Joint Optimization: The model is trained across all map-side combinations simultaneously.
Stochastic Source Dropout: During training, random source domains are omitted to force the model to learn robust traits even with incomplete data, preventing overfitting to specific domain pairs.
Objective: Minimize Mean Squared Error (MSE) between predicted and actual coordinates across all target domains.

3. Key Contributions

Multi-Domain Integration Framework: A novel architecture that jointly encodes behaviors across distinct environments to infer a "pure" representation of individuality, successfully disentangling intrinsic style from contextual demands.
Low-Dimensional Compression: Demonstration that complex, idiosyncratic movement policies can be compressed into a 2-dimensional latent space without sacrificing predictive accuracy, suggesting that human strategic variation in this domain is governed by a small number of latent factors.
Structured Metric Space: Discovery that the learned latent space forms a structured metric manifold where Euclidean distances between players correlate linearly with the degradation of transfer performance (behavioral dissimilarity).
Interpretability: The latent axes align with interpretable psychological/behavioral phenotypes (e.g., risk-taking vs. caution, solo-play vs. team cohesion), bridging high-dimensional telemetry with meaningful constructs.

4. Key Results

A. Zero-Shot Transfer Performance

Cross-Domain Superiority: The proposed model significantly outperformed both a "Population Model" (average behavior) and a "Self Model" (trained only on single-player data) in predicting trajectories in completely unseen map-side combinations.
Robustness to Data Scarcity: The multi-domain approach mitigated the data scarcity issues inherent in single-player datasets, achieving higher accuracy than the Self baseline even when the target domain had no training data.
Source Diversity: Prediction error decreased monotonically as the number of source domains ( $k$ ) used to compute the latent code increased, confirming that diverse behavioral evidence is required for robust trait extraction.

B. Dimensionality Analysis

Optimal Dimensionality: While training error decreased with higher dimensions, generalization performance plateaued at $d=2$ .
Sufficiency: A 2-dimensional embedding was sufficient to capture the majority of individual strategic variation, whereas $d=1$ was insufficient. Higher dimensions ( $d>2$ ) offered marginal gains but increased complexity.

C. Latent Space Structure & Interpretability

Metric Correlation: A Linear Mixed-Effects Model (LMM) revealed a significant positive relationship ( $\beta = 51.26, p < 0.001$ ) between the Euclidean distance in latent space and the relative increase in prediction error. Closer latent vectors imply more similar behavioral policies.
Phenotypic Alignment:
- Dimension $z_0$ : Correlated with survival duration and teammate proximity. It represents a spectrum between high-risk/aggressive solo-play and conservative/team-oriented strategies.
- Dimension $z_1$ : Correlated with movement speed and teammate proximity. It disentangles fast-paced individualistic movement from slower, coordinated group movements.

5. Significance and Implications

Behavioral Science: The study provides a computational method to define individuality not as fixed parameters in a single task, but as consistency across diverse challenges. It validates the existence of stable "behavioral phenotypes" in high-dimensional decision-making tasks.
AI and Generalization: The framework demonstrates a pathway toward lifelong learning and domain generalization in AI, showing how agents can learn invariant representations that transfer across radically different environments.
Applications:
- Personalized Modeling: Enables high-fidelity personalized models for users with limited data.
- Psychometrics: Offers a bridge between raw telemetry and formalized personality theory (e.g., risk-taking, social cohesion) without subjective self-reporting.
- Team Dynamics: Provides a foundation for analyzing how individual strategic phenotypes interact to form collective intelligence and team synergy.

In conclusion, this work establishes that individual behavioral signatures in complex, high-dimensional tasks are low-dimensional, domain-invariant, and structurally organized, allowing for accurate zero-shot transfer and interpretable psychological analysis.