🌀 nonlinear sciences

Internal-state criticality in Bayesian-inverse-Bayesian inference

This paper proposes Bayesian-inverse-Bayesian (BIB) inference as a minimal generative model that induces robust internal criticality in repeated games through hypothesis renewal, characterized by heavy-tailed statistics and a zero-drift log-posterior walk without requiring external parameter tuning.

Original authors: Kazuto Sasai, Yukio-Pegio Gunji

Published 2026-06-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Kazuto Sasai, Yukio-Pegio Gunji

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 playing a game of Rock-Paper-Scissors against a computer. You want to win, so you try to figure out what the computer is going to do next.

Most computer programs try to be perfect. They calculate the odds, look for patterns, and eventually settle on a "safe" strategy: playing Rock, Paper, and Scissors completely at random, 33% of the time each. In game theory, this is called the "Nash Equilibrium." It's the point where you can't be beaten, but you also can't win. If you play against a perfect computer, the game becomes a boring, predictable shuffle of random moves.

But humans aren't like that. When we play, we get stuck in streaks. We might play "Rock" five times in a row, then suddenly switch to "Paper" for ten turns. Our decisions aren't random; they have a "heavy tail," meaning long streaks happen more often than pure chance would predict.

This paper introduces a new way for a computer to think, called Bayesian–Inverse-Bayesian (BIB) inference. It claims this method allows a computer to naturally develop those same human-like streaks and patterns, not because it was programmed to copy humans, but because of how its internal "brain" works.

Here is the breakdown of how it works, using simple analogies:

1. The Two-Step Thinking Process

The BIB agent uses two opposing mental steps that balance each other out:

Step A: The "Focus" (Bayesian Update)
Imagine you are a detective gathering clues. Every time you see a clue, you get more confident about one specific suspect. You start to ignore the others. If you kept doing this forever, you would become 100% sure about just one suspect and stop looking at anyone else. This is "concentration."
Step B: The "Reset" (Inverse-Bayesian Step)
Now, imagine that every time you get too confident, a friend taps you on the shoulder and says, "Hey, remember that weird pattern we saw earlier? Let's pretend that is the new truth, even if you don't have much evidence for it yet."
This step takes your least-believed idea and replaces it with the most recent pattern you've seen. It forces your mind to stay open and prevents you from getting stuck on just one idea.

2. The "Critical State" (The Sweet Spot)

The paper argues that when you mix these two steps—getting focused and then getting reset—you hit a "sweet spot" called criticality.

Think of a tightrope walker.

If they only focus (Step A), they become rigid and fall because they can't adapt.
If they only reset (Step B), they are too chaotic and can't make a decision.
But if they balance both, they enter a state of internal criticality. They are stable enough to stand, but flexible enough to sway with the wind.

In this state, the agent's "favorite guess" (the hypothesis it trusts most) stays the same for a while, then suddenly switches. The length of time it stays on one guess follows a specific mathematical rule (a power law). This means the agent naturally produces long streaks of the same thought, just like humans do.

3. Why This is Special

The researchers tested this against a standard "Bayesian-only" computer (one that only does Step A, the "Focus").

The Standard Computer: It quickly settles into a boring, random pattern. It has no long streaks. It's like a robot that never changes its mind until it's forced to.
The BIB Computer: It naturally falls into the "critical state." It has long streaks of sticking to a strategy, followed by sudden, dramatic shifts.

The amazing part is that this happens automatically. The researchers didn't have to tweak any knobs or tell the computer, "Hey, try to act like a human." They just added the "Reset" step, and the complex, human-like behavior emerged on its own.

4. The "Internal" vs. "External" Secret

There is a twist. When two BIB computers play against each other, they both become so good at adapting that they end up playing randomly, just like the standard computer. They cancel each other out.

However, their internal thoughts are still chaotic and streaky. They are thinking in long streaks, but because they are playing against an opponent who is also adapting perfectly, those thoughts don't show up in their final moves.

But if you play BIB against a human (who makes mistakes and has predictable biases), the BIB agent's internal "streaky" nature allows it to lock onto the human's patterns and exploit them. It acts like a flexible predator that can switch strategies instantly when it sees an opening.

The Big Takeaway

This paper suggests that the reason humans (and perhaps animals) make decisions in long, heavy-tailed streaks isn't because we are bad at math or we are "irrational." Instead, it might be because our brains use a similar "Focus and Reset" mechanism.

By adding a simple "forgetting and renewing" step to a standard learning algorithm, the system naturally organizes itself into a state of criticality. It creates a balance between stability and flexibility that allows for complex, adaptive behavior without needing to be explicitly programmed to do so. It's a new way to build AI that thinks more like a living, breathing decision-maker rather than a rigid calculator.

Technical Summary: Internal-State Criticality in Bayesian–Inverse-Bayesian Inference

Problem Statement
Recent empirical and theoretical work suggests a convergence between criticality and inference dynamics, with neural recordings and biological foraging behaviors often exhibiting heavy-tailed (power-law) statistics near critical points. However, existing literature largely identifies signatures of criticality or argues that inferred models tend to lie near critical surfaces, without providing a concrete inference rule whose own dynamics are inherently critical. Furthermore, standard game-playing AI (e.g., Nash-targeting optimizers like Counterfactual Regret Minimization) seeks accuracy and convergence to optimal strategies, often collapsing to uniform random play in symmetric zero-sum games, thereby failing to capture the non-trivial, heavy-tailed dynamics observed in human decision-making. The paper addresses the need for a minimal generative model where criticality emerges from the inference process itself, rather than being imposed by external tuning or self-organization toward an absorbing state.

Methodology
The authors propose and simulate Bayesian–Inverse-Bayesian (BIB) inference within the context of repeated $N$ -hand cyclic-dominance Rock-Paper-Scissors (RPS) games.

The Model: The BIB agent maintains a set of hypotheses ( $N_h$ ) about the opponent. It performs a standard Bayesian update (contraction) based on observed outcomes. Crucially, it augments this with an Inverse-Bayesian step (relaxation): after a window of $m$ observations, the hypothesis with the lowest posterior probability is replaced by the empirical histogram of recent observations. This "hypothesis renewal" prevents posterior collapse and injects new hypotheses reflecting recent data flows.
Experimental Design: The study employs a massive-scale simulation ( $\sim 10^8$ decision events) sweeping six axes: game size ( $N \in \{3, 5, 7\}$ ), hypothesis count ( $N_h \in \{3, 6, 10, 15, 20\}$ ), window size ( $m \in \{10, 20, 50, 100\}$ ), initialization (random vs. structured), prediction mode (sampling vs. argmax), and likelihood sharpness.
Controls: The authors compare BIB against:
- Bayes-Only (BO): Agents performing standard Bayesian updates without the inverse step.
- Standard Learners: Win-Stay/Lose-Shift (WSLS), Q-learning, and Regret Matching.
- Random Agents: Uniform random play (Nash equilibrium baseline).
Observables: The primary metric is the argmax-persistence time ( $T_{argmax}$ ), defined as the duration the most probable hypothesis remains unchanged. Secondary metrics include laminar-phase lengths (periods where the maximum posterior probability exceeds a threshold) and finite-size scaling analyses.

Key Results

Universal Critical State: Across all design variations (initialization, prediction mode) and opponent classes, BIB dynamics settle into a robust internal critical state. The distribution of argmax-persistence times follows a heavy-tailed power law with an exponent $\alpha \approx 1.43$ (at the canonical window $m=50, N=3$ ).
Absence in Bayes-Only Inference: BO agents, lacking the inverse step, do not exhibit power-law persistence. Their distributions are better fit by stretched exponentials or lognormals, and they lack a universal exponent across hypothesis counts.
Universality Class and Drift: The measured exponent $\alpha \approx 1.43$ is not a fixed constant but systematically drifts toward the universal value of $3/2$ (characteristic of on-off intermittency and Sparre–Andersen first-return laws) as the finite-sample residual $(N-1)/(2m)$ vanishes. This confirms the dynamics are driven by a zero-drift log-posterior walk pinned by the inverse step.
Finite-Size Scaling: The system exhibits finite-size scaling with respect to the hypothesis count $N_h$ . In the core regime ( $N_h \in \{6, 10\}$ ), the exponent is invariant. The cutoff time scales as $T_{max} \propto N_h^{-z}$ with $z \approx 1.49$ . The posterior spread scales as $\sigma \propto N_h^{-\beta}$ with $\beta \approx 1.07$ , close to the information-theoretic lower bound, indicating the inverse step absorbs the scaling into the renewal dynamics.
Robustness: The universality holds across different game dimensions ( $N=3, 5, 7$ ), different game structures (tested on Matching Pennies, $N=2$ ), and various observation rules, provided the inverse-Bayesian renewal remains active.
Internal vs. External: In symmetric BIB-vs-BIB play against a uniform Nash opponent, the heavy-tailed criticality is confined to the internal hypothesis space. Behavioral streaks (hand repeats, win streaks) remain exponential because the symmetric opponent pins the action marginal to uniform Nash. However, against a fixed, biased (exploitable) opponent, the agent's behavioral streaks cross over to power laws, demonstrating that the internal criticality acts as a reservoir of flexibility.

Significance and Claims
The paper claims to identify a universal critical structure generated by adding an inverse-Bayesian relaxation step to standard Bayesian inference.

Mechanism: The authors argue this is a mechanism of internal-state criticality achieved through the continual reconstruction of the hypothesis-space boundary, rather than self-organization toward an absorbing state (traditional SOC). The inverse step pins the per-step log-posterior drift to zero structurally, creating a critical state without external parameter tuning.
Distinction from SOC: While sharing features with SOC (scale-free statistics, no tuning), the authors distinguish BIB by noting it does not self-organize to an absorbing critical point but maintains criticality via boundary reconstruction. They identify the dynamics operationally with on-off intermittency on the simplex of hypothesis distributions.
Behavioral Relevance: The study suggests BIB provides a generative model for the heavy-tailed persistence statistics observed in human decision-making. Unlike Nash-targeting algorithms that collapse to randomness in symmetric games, BIB agents naturally reproduce the non-trivial, heavy-tailed dynamics of human play through their internal mechanism, without requiring pre-training or external fitting.
Modesty: The authors explicitly state that $\alpha \approx 1.43$ is a protocol-dependent estimate of a universality class, not a physical constant. They do not claim to prove classical SOC but rather to characterize a universal critical structure that is robust across a natural parameter range. The significance lies in providing a minimal, generative inference rule that bridges the gap between rational optimization (which fails to capture human dynamics) and the empirical observation of criticality in biological systems.