Sequential Causal Normal Form Games: Theory, Computation, and Strategic Signaling

This paper extends Causal Normal Form Games to sequential settings by introducing Sequential Causal Multi-Agent Systems, but its comprehensive theoretical and empirical analysis reveals that, under standard rational assumptions, these causal frameworks offer no welfare advantage over classical Stackelberg equilibrium, thereby highlighting a fundamental incompatibility between rational choice and causal reasoning benefits in current game-theoretic models.

The Gist

Imagine you are playing a game of chess, but instead of just moving pieces, you are also trying to guess how your opponent is thinking. Are they following a strict rulebook? Are they acting on a gut feeling? Or are they simulating "what if" scenarios in their head?

This paper by Dennis Thumm asks a big question: Can we build better game theories for AI by adding these layers of "thinking styles" (causal reasoning), or does it just make things complicated without helping anyone win?

Here is the story of the paper, broken down into simple concepts and analogies.

1. The Setup: The "Thinking Layers" of AI

In traditional game theory (like the famous "Stackelberg game"), we assume everyone is a perfect calculator. If you make a move, the other person calculates the absolute best response.

But real AI (and humans) aren't just calculators. They have different "modes" of operation, based on a famous framework called the Causal Hierarchy:

• Layer 1 (The Instinct): "I see a red light, so I stop." (Observation/Reaction).
• Layer 2 (The Deliberate): "I choose to stop because I want to avoid a ticket." (Intervention/Choice).
• Layer 3 (The Time Traveler): "If I hadn't stopped, I would have crashed, so I'm glad I stopped." (Counterfactual/Reasoning).

The author built a new game system called S-CMAS (Sequential Causal Multi-Agent Systems). Think of it as a video game where the Leader can choose to play on "Instinct Mode," "Logic Mode," or "Super-Reasoning Mode," and the Follower can see which mode the Leader picked.

2. The Big Hope: The "Secret Sauce"

The hope was that by letting the Leader signal, "Hey, I'm acting on pure instinct!" or "I'm using deep counterfactual reasoning!", they could trick or guide the Follower into making a move that benefits the Leader more than a standard game would allow.

It's like a poker player trying to bluff by wearing a specific hat that says, "I am the kind of person who bluffs," hoping the opponent folds. The theory suggested this "causal signaling" could unlock new, better outcomes for everyone.

3. The Experiment: Running the Simulation

The author didn't just write equations; they ran over 50 different types of games (from simple coordination games to complex "Battle of the Sexes" scenarios) and simulated them thousands of times. They tested:

• Different sizes of game boards.
• Different levels of "noise" (confusion).
• Different types of AI "instincts" (some good, some bad).

They wanted to see if the new "Causal Game" produced better results (more points/welfare) than the old "Classic Game."

4. The Shocking Result: The "Zero Improvement"

Here is the twist: The new game didn't work at all.

Across all 100+ test scenarios, the "Causal Game" produced exactly the same results as the old, boring "Classic Game."

• The Analogy: Imagine you invent a new, high-tech steering wheel for a car that claims to make the car drive faster by reading the driver's mind. You test it on 50 different tracks. The result? The car drives at the exact same speed as the old steering wheel. The fancy new tech was useless.

Why did it fail?
The paper explains that in a sequential game, the "Follower" (the second player) is too smart.

1. The Leader makes a move.
2. The Follower sees the move.
3. The Follower asks: "What is the best thing for me to do right now?"
4. The Follower ignores how the Leader got there (whether it was instinct or deep reasoning) and just reacts to the move itself.

If the Leader's "instinct" happens to be the same as their "logic," the Follower can't tell the difference. If the Leader's "instinct" is bad, the Leader simply won't use it; they'll switch to logic. In the end, everyone just plays the standard "best response" game, and the fancy causal layers disappear.

5. The "Negative" Lesson: Why This Matters

Usually, scientists love to find a new, shiny tool that solves problems. This paper is valuable because it found a dead end.

It tells us: "You cannot just tack 'causal reasoning' onto old economic theories and expect AI to behave better."

If we assume AI agents are rational enough to calculate the best move (backward induction), then the "causal layers" don't matter. The paper argues that to truly understand AI agents (like Large Language Models), we need to stop pretending they are perfect calculators. We need new theories that account for:

• Persistent quirks: AI that doesn't just "learn" to be perfect but keeps making "instinctive" mistakes.
• Non-equilibrium: Situations where players don't find the perfect balance.

Summary in One Sentence

The author tried to upgrade game theory for AI by adding layers of "thinking styles" (instinct vs. logic), but discovered that if the AI is smart enough to play the game perfectly, those extra thinking layers don't actually give anyone an advantage—proving that we need entirely new ways to model AI, not just tweaked versions of old human theories.

Technical Summary

Here is a detailed technical summary of the paper "Sequential Causal Normal Form Games: Theory, Computation, and Strategic Signaling" by Dennis Thumm.

1. Problem Statement

The paper addresses a fundamental gap in modeling strategic interactions for Artificial Intelligence (AI) agents, particularly Large Language Models (LLMs).

• The Limitation of Classical Game Theory: Traditional frameworks like Stackelberg games assume perfect rationality and backward induction. However, real-world AI agents often exhibit "bounded rationality," relying on instincts (training data priors), following heuristics, or engaging in counterfactual reasoning that deviates from standard game-theoretic prescriptions.
• The Gap in Causal Modeling: While Causal Normal Form Games (CNFGs) successfully integrate Pearl's Causal Hierarchy (PCH)—Observational (L1), Interventional (L2), and Counterfactual (L3)—into simultaneous-move games, they fail to capture sequential interactions where leaders commit to actions before followers respond.
• Core Question: Can extending CNFGs to sequential settings (creating Sequential Causal Multi-Agent Systems or S-CMAS) provide strategic advantages (e.g., improved welfare) over classical Stackelberg equilibria when agents operate across different causal layers?

2. Methodology

The authors propose a rigorous theoretical framework followed by an extensive empirical validation.

Theoretical Framework: S-CMAS

• Definition: An S-CMAS is defined as a tuple $\langle M, N, X, Y, \preceq, I \rangle$, incorporating a Structural Causal Model (SCM) $M$, a partition of agents into Leaders ($L$) and Followers ($F$), action nodes with temporal ordering, and information structures.
• Causal Layers: Agents can operate at three levels:
  • L1 (Observational): Instinctive actions driven by unobserved factors ($X_i \leftarrow f(U_i)$).
  • L2 (Interventional): Deliberate choices (standard game theory, $do(X_i=x_i)$).
  • L3 (Counterfactual): Sophisticated reasoning conditioning on natural instincts ($h: D(X^*_i) \to D(X_i)$).
• Information Scenarios: The model considers "Mechanism Information," where followers observe not just the leader's action but also the causal layer (L1, L2, or L3) the leader utilized.
• Equilibrium Concept: The paper defines Sequential Causal Nash Equilibrium (S-CNE). It is computed via backward induction:
  1. Followers observe the leader's layer and action, then choose an optimal response within their own chosen layer.
  2. Leaders anticipate this response and select the optimal layer and action.
• Refinements: The authors introduce Trembling-Hand and Forward Induction refinements to eliminate non-credible threats and ensure rational inferences about intentions.

Computational Approach

• Complexity: The authors prove that computing an S-CNE is PSPACE-complete.
• Algorithms:
  • Exact Computation: Feasible for small action spaces via exhaustive backward induction.
  • Approximation: A Polynomial-Time Approximation Scheme (PTAS) is proposed (Algorithm 1), which samples causal realizations to compute $\epsilon$-approximate equilibria.

Empirical Investigation

To test the practical utility of the framework, the authors conducted a comprehensive study:

• Monte Carlo Simulations: 50+ randomly generated S-CMAS instances with varying action spaces (2–20), causal topologies (chains, forks, cycles), and information structures.
• Synthetic Examples: 5 hand-crafted game types specifically designed to favor causal reasoning (e.g., Coordination games, Battle of the Sexes, Stag Hunt, Prisoner's Dilemma with cooperation instincts).
• Metrics: Social welfare comparison between S-CNE and classical Stackelberg equilibrium; Pareto improvement rates.

3. Key Contributions

1. Formalization of S-CMAS: The first theoretical extension of Causal Normal Form Games to sequential, leader-follower settings, integrating Pearl's Causal Hierarchy into dynamic strategic interactions.
2. Complexity Analysis: Proof of PSPACE-completeness for S-CNE computation, alongside identification of tractable special cases (acyclic structures, fixed layers).
3. Signaling Theory Connection: Demonstrated that S-CMAS can be recast as a signaling game where the leader's choice of causal layer acts as a signal of their type (causal structure).
4. Critical Negative Result: The most significant contribution is the empirical demonstration that S-CNE provides zero welfare improvement over classical Stackelberg equilibrium in all tested scenarios.

4. Results

The empirical investigation yielded a definitive negative result:

• Zero Welfare Improvement: Across 100 tested instances (50 random + 50 synthetic), the S-CNE achieved 0% Pareto improvement over the classical Stackelberg equilibrium.
• Layer Selection Collapse: In 96% of cases, leaders chose the L1 (instinctive) layer. However, because the "instincts" (L1) aligned with the rational best-response (L2) in equilibrium, the resulting actions were identical to those in a classical Stackelberg game.
• Irrelevance of Mechanism Information: Even when followers observed the leader's causal layer, their optimal response remained unchanged. If instincts were "good" (quality > 0.5), L1 matched L2; if "poor," rational leaders simply avoided L1.
• Backward Induction Neutralization: The study concludes that backward induction with rational best-response eliminates any strategic advantage derived from causal layer distinctions. The equilibrium logic forces agents to converge on the same outcomes regardless of the underlying causal structure.

5. Significance and Implications

The paper offers a crucial "negative result" that challenges current trends in Agentic AI research:

• Incompatibility of Rationality and Causal Advantage: The study proves that classical game-theoretic solution concepts (Nash equilibrium, backward induction) are fundamentally incompatible with leveraging causal reasoning advantages. If agents are rational enough to compute equilibria, the nuances of causal layers (L1 vs. L2 vs. L3) become strategically irrelevant.
• Critique of "Retrofitting" AI: Extending classical economic theory (based on rational choice) to AI agents by simply adding causal structures or bounded rationality is insufficient. The "rational best-response" assumption inherent in equilibrium concepts negates the very benefits (like instinctive or counterfactual advantages) the framework attempts to model.
• Call for New Foundations: The authors argue that modeling strategic AI (especially LLM-based agents) requires non-equilibrium frameworks. Future research should focus on:
  • Learning dynamics and persistent bounded rationality (where agents do not instantly converge to best responses).
  • Satisficing behaviors rather than optimization.
  • Frameworks that account for the specific "instincts" (priors) and inference procedures of AI agents without assuming they will always act rationally.

Conclusion: While the theoretical framework of S-CMAS is elegant, the paper demonstrates that it fails to provide practical strategic advantages in sequential settings under standard equilibrium assumptions. This motivates the Agentic AI community to move beyond retrofitted economic models and develop genuinely new theoretical tools that account for the unique, non-rational, and persistent behaviors of modern AI agents.