🤖 machine learning

Strategy, Not Payoffs: A Behavioural Embedding of Normal-Form Games

This paper proposes a lightweight behavioural embedding based on Nash equilibrium entropy and response sensitivity to successfully predict how fine-tuning on specific normal-form games transfers strategic capabilities to unseen games in large language models, outperforming existing structural embeddings that merely memorize game identities.

Original authors: Joshua Caiata, Sreepriya Pulyassary, Xiang Li, Kate Larson

Published 2026-07-31

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Joshua Caiata, Sreepriya Pulyassary, Xiang Li, Kate Larson

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are teaching a robot how to play video games. You might think that if you teach it to be a master at Chess, it would automatically become better at Checkers because both involve strategy. But what if teaching it Chess actually made it worse at Checkers? This is the puzzle scientists are trying to solve with Artificial Intelligence. Specifically, they are looking at "Large Language Models" (LLMs)—the super-smart computer brains behind chatbots. These models are being trained to act like strategic players in games, bargaining, cooperating, and competing. The big question is: Does learning one game help or hurt a model's ability to play a totally different game?

To understand this, we need to know a few things. First, in game theory, a "Normal-Form Game" is just a fancy way of describing a situation where two people make choices at the same time, and the result depends on what both of them pick (like Rock-Paper-Scissors). Second, there is a concept called "Nash Equilibrium," which is basically the "perfect play" strategy where no one wants to change their move if they know what the other person is doing. Finally, "Fine-tuning" is the process of taking a general AI and giving it a crash course on a specific task. The researchers wanted to know: Is the "shape" of the game (the rules and payoffs) what matters for learning, or is it something deeper about the behavior the game requires?

The Strategy, Not the Scoreboard

In this study, a team of researchers from the University of Waterloo decided to treat these AI models like students in a very strict school. They took 49 different small AI models and gave them a crash course in 15 classic games, ranging from the famous "Prisoner's Dilemma" to "Rock-Paper-Scissors." The goal wasn't just to see if the AI could play; it was to see if learning Game A made the AI better or worse at Game B.

The researchers suspected that previous methods of predicting this "transfer" were missing the point. Imagine trying to guess if a student who is good at Math will be good at Physics. A simple method might just say, "Oh, they are both science classes, so yes!" But that's too vague. The researchers wanted a better map. They proposed a new way to describe a game using just two simple numbers, which they called ENT-SW.

Think of ENT (Entropy) as a measure of confusion.

If a game has a clear, obvious best move (like always choosing "Defect" in the Prisoner's Dilemma), the "confusion" is low. The strategy is solid and stable.
If a game requires you to mix up your moves randomly to stay unpredictable (like Rock-Paper-Scissors), the "confusion" is high. The strategy is fluid and shifting.

Think of SW (Switching) as a measure of reactivity.

In some games, your best move is the same no matter what your opponent does. You don't need to react; you just stick to your plan.
In other games, your best move changes completely depending on what your opponent does. If they play Rock, you play Paper. If they play Scissors, you play Rock. You have to be a chameleon, constantly switching your strategy.

The team created a simple map where every game is a dot based on these two numbers: how confusing the strategy is, and how much you have to switch your moves.

The Great Experiment

To test if this map worked, the researchers ran a massive simulation. They took their 49 AI models and trained each one on one specific game (the "Source"). Then, they tested that same trained model on all the other games (the "Targets"). They measured how much the model improved or got worse.

They compared their new ENT-SW map against two other ways of guessing:

The "Identity" Baseline: This is like saying, "We know exactly what Game A and Game B are, so let's just memorize the history of how they interact." It's a cheat code that assumes you've seen every possible pair before.
Old Game Theory Maps: These are complex, mathematical descriptions of the games that scientists have used for years.

Here is the twist: The researchers used a very strict test called "Leave-One-Game-Out." This means they trained the model on 14 games and then asked it to predict what would happen with the 15th game it had never seen before. It's like teaching a student 14 subjects and then testing them on a brand-new subject they've never encountered, without letting them peek at the answer key.

The Surprising Results

The results were clear and quite surprising.

1. The Old Maps Failed: The complex, traditional mathematical maps of the games (which look at the raw numbers of the payoffs) failed miserably when tested on games the AI had never seen. They were just as good as guessing, or sometimes even worse. They seemed to be memorizing the specific names of the games rather than understanding the underlying logic.

2. The "Identity" Cheat Code: As expected, if you just memorized the specific pair of games (e.g., "Prisoner's Dilemma to Stag Hunt"), you could predict the outcome very well. But this doesn't help you with new games.

3. The ENT-SW Victory: The simple two-number map (Confusion + Reactivity) was the only method that successfully predicted how the AI would perform on a completely new, unseen game. It outperformed the "Identity" baseline.

The researchers found that the specific pairing of the two games mattered the most. In fact, the structure of the game pair explained 77.1% of the results, while the specific AI model used only explained 2.8%. This means it doesn't matter which AI you use; what matters is the "shape" of the strategy required by the games.

The "Harmony" Trap

One of the most interesting findings was a trap in the map. The researchers found that the game "Harmony" and the "Prisoner's Dilemma" looked almost identical on their ENT-SW map. Both had low confusion (a clear best move) and low switching (you don't need to react to the opponent).

However, they were behaviorally opposites!

In Harmony, the clear best move is to cooperate.
In Prisoner's Dilemma, the clear best move is to selfishly defect.

If you trained an AI to be a "selfish defector" on the Prisoner's Dilemma, it would fail miserably when asked to play Harmony, even though the map said the games were the same. The map captured the shape of the decision (it's a straight line, no switching), but it couldn't see which way the line pointed (toward kindness or selfishness). Despite this limitation, the ENT-SW map was still the best predictor the researchers could find, proving that the structure of the decision-making process is more important than the raw numbers on the scoreboard.

What This Means

This study suggests that when AI learns to play games, it isn't just memorizing the rules or the points. It is learning the behavioral demands of the situation. Is the situation one where you need to be stable and confident? Or is it one where you need to be reactive and flexible?

The researchers suggest that if we want to build smarter AI that can transfer its skills from one task to another, we shouldn't just look at the game's rules or payoffs. We need to understand the "personality" of the strategy required. By using a simple, two-feature map of "Confusion" and "Switching," we can predict how an AI will handle a new challenge, even if it has never seen that challenge before.

While the study was limited to 15 games and small AI models, the findings offer a new way to think about how machines learn. It turns out that in the world of strategy, it's not about the score; it's about the shape of the game.

Technical Summary: Strategy, Not Payoffs: A Behavioural Embedding of Normal-Form Games

Problem Statement

Large Language Models (LLMs) are increasingly deployed as strategic agents capable of bargaining, cooperating, and competing. While fine-tuning adapts these models to specific roles, a critical challenge remains: understanding how training on one strategic task affects performance on another. Specifically, fine-tuning on a source game can either enhance or degrade an agent's ability to reason in a target game.

Current approaches to predicting this "transfer of strategic capabilities" rely on task embeddings derived from vision or NLP domains, or structural game representations validated only in-sample (e.g., by recovering known taxonomies or fitting transfer matrices on the same data used for scoring). A fundamental gap exists: it is unknown whether a game's structure can predict behavioral transfer to a genuinely held-out game (one never seen during training) better than a simple baseline that merely memorizes game identities. Existing structural embeddings often fail to generalize beyond the specific games they were trained on or evaluated against.

Methodology

1. The ENT-SW Embedding

The authors propose a lightweight, two-feature behavioral embedding termed ENT-SW, designed to capture the fundamental behavioral demands of a game rather than its raw payoff geometry. The embedding is computed solely from the payoff table of a game instance:

Equilibrium Entropy (ENT): Measures the ambiguity of the Nash equilibrium. It is near zero for games with stable pure strategies and near one for games requiring mixed strategies.
Best-Response Switching (SW): Captures the sensitivity of optimal responses to an opponent's actions. It distinguishes games where a dominant action exists (low switching) from coordination, anti-coordination, or cyclic games where the best response changes based on the opponent (high switching).

For each game, the authors aggregate these instance-level features into a single vector by averaging over sampled instances and player roles. To predict transfer between a source game $G_i$ and a target game $G_j$ , they construct a pair representation $\phi(i, j)$ combining the individual game vectors, their difference, and their element-wise product.

2. Experimental Protocol

The study employs a rigorous generalization protocol involving 49 small open-weight LLMs (ranging from <200M to 3B parameters) across diverse architecture families (e.g., Pythia, Llama, GPT-Neo, Qwen).

Game Suite: 15 classic normal-form games (e.g., Prisoner's Dilemma, Stag Hunt, Rock-Paper-Scissors) are used.
Training: Each LLM is separately fine-tuned on a dataset generated from a single source game. The training target is the Nash equilibrium distribution for that specific game instance. Prompts contain only payoff consequences with randomized action labels to prevent reliance on fixed semantics.
Evaluation: Fine-tuned models are evaluated on held-out target games. Performance is measured by the Mean Squared Error (MSE) between the model's induced action distribution and the target Nash equilibrium.
Transfer Metric: Behavioral transfer is quantified as the log-relative improvement of the fine-tuned model over the "cold" (pre-fine-tuned) base model.
Cross-Validation Regimes:
- Leave-One-Game-Out (LOGO): The target game is completely unseen during training (neither as a source nor in any pair). This tests true generalization.
- Leave-One-Pair-Out (LOPO): The specific pair $(G_i, G_j)$ is held out, but both games may have appeared in other pairings. This tests interpolation.

3. Baselines and Comparisons

The ENT-SW embedding is benchmarked against:

Game Identity Baseline: One-hot indicators for source and target games.
Structural Embeddings from Literature: Including potential-harmonic decompositions (PHD), response-graph statistics (RG), and equilibrium-invariant representations (EI-diff, EIE cloud, BRI cloud).
Raw Payoff Baseline (RSTP): Using raw payoff vectors directly.

Key Results

Predictive Performance

Superiority over Identity: Under strict LOGO conditions, ENT-SW outperforms the game-identity baseline by a significant margin (0.360 improvement in held-out squared error reduction). It also outperforms identity under LOPO.
Generalization vs. Memorization: Structural embeddings from existing literature (e.g., PHD, RG stats, EI-diff) perform well under LOPO (interpolation) but fail under LOGO, often performing no better than or worse than the identity baseline. This suggests they rely on memorizing pairwise regularities rather than capturing generalizable structural properties.
Robustness: ENT-SW is the only representation tested that maintains predictive value under the strict LOGO regime, even when the identity baseline is given full credit for the specific games involved.

Variance Decomposition

A decomposition of transfer variance reveals that:

Game-Pair Structure: Accounts for 77.1% of the variance in transfer scores.
LLM Identity: Accounts for only 2.8%.
Interaction: The specific interaction between the source and target games dominates individual contributions, confirming that transfer is a property of the game pair's compatibility rather than the specific model or game in isolation.

Case Study: The "Harmony" Outlier

The embedding successfully identifies behavioral clusters (e.g., Prisoner's Dilemma, Deadlock, and Compromise form a highly compatible cluster). However, the paper notes a limitation: Harmony shares the same ENT and SW values (0, 0) as the Prisoner's Dilemma cluster but behaves incompatibly with them. This is because ENT-SW captures the shape of the equilibrium (dominant strategy, no switching) but not the direction (cooperative vs. self-interested). Despite this, ENT-SW remains the best predictor in the comparison set.

Key Contributions

Behavioral Embedding Proposal: Introduction of ENT-SW, a two-feature embedding that captures the strategic "terrain" (ambiguity and reactivity) required by a game, rather than its payoff geometry.
Rigorous Evaluation Framework: Establishment of a strict LOGO protocol for evaluating strategic transfer, moving beyond in-sample validation to test generalization to genuinely unseen games.
Empirical Evidence of Structure over Identity: Demonstration that predicting strategic transfer requires representing the underlying decision-making behavior, as structural embeddings that fail to generalize to new games are outperformed by simple identity baselines in strict settings.
Variance Attribution: Quantitative proof that transferability is primarily a function of the source-target game pair structure (77.1%) rather than the specific LLM architecture (2.8%).

Significance and Claims

The paper claims that the transfer of strategic capabilities in LLMs is not dictated by the payoff geometry of a game, but by the underlying structure of the decision-making behavior it requires.

The authors argue that ENT-SW provides a compact, generalizable stand-in for game identity, enabling the prediction of transfer for games a model has never encountered. This challenges the notion that complex, learned, or high-dimensional structural representations are necessary for strategic transfer; instead, simple, interpretable features describing the equilibrium landscape (entropy and switching) are sufficient to explain and predict behavioral transfer across diverse LLM architectures.

The paper remains modest regarding the universality of these findings, noting that the study is limited to 15 two- and three-action games and that the specific features (ENT and SW) may not be the unique solution, but rather representative of a class of equilibrium-based features that capture behavioral shape. Future work is suggested to extend this to larger action spaces, multi-player settings, and sequential games.