Beta Distribution as a Convergence
Framework for Loop Agents

A Unified Theory from Evolutionary Trial-and-Error
to Bayesian Intelligent Agents

0Abstract

1Introduction: The Stopping Condition Problem in Loop Agents

1.1 The Paradigm Shift from Single-Pass Inference to Iterative Loops

The evolution of AI systems can be understood as the progressive engineering realization of Bayesian conditional probability across different dimensions. The introduction of the Transformer architecture in 2017 formulated language modeling as context-conditioned next-token probability prediction P(token|context). At this stage, the model’s probability distribution was fixed upon completion of training — it could only perform inference based on parameterized prior knowledge, equivalent to a “closed-book exam.”

In 2020, Lewis et al. proposed the Retrieval-Augmented Generation (RAG) framework, which for the first time injected external real-time information as Bayesian anchor points into the reasoning process, upgrading output probability to P(token|context, retrieved results). In 2023, Microsoft’s BingGPT fused search engines with LLMs, achieving broader real-time grounding. In 2025–2026, the introduction of Function Calling extended the probability space from the vocabulary to the action space — models now decide not only “what to say” but also “what to do” and “which tool to use.”

In June 2026, the Loop Agent architecture explosively became an industry focal point. Its core idea is that an Agent no longer performs single-pass inference but instead repeatedly acts, verifies, receives feedback, and corrects within an iterative loop until a stopping condition is met. This marks the latest leap of AI from fixed prior probabilities to dynamic Bayesian posteriors.

1.2 Limitations of Current Stopping Conditions

However, the stopping conditions of Loop Agents in current engineering practice still rely on crude heuristics. The most commonly used is a fixed maximum iteration count (max_iterations), an integer entirely unrelated to the task content. Simple tasks that converge in one or two rounds are forced to run through all iterations, wasting computational resources; complex tasks may not finish before the counter expires. The cutting-edge improvement of 2026 — cosine distance-based semantic convergence detection across consecutive drafts — is superior to a fixed counter, but it measures the surface similarity of output text rather than the evolution of the model’s confidence in its conclusion.

The core problem is this: there currently lacks a mathematical tool capable of answering “when is stopping information-theoretically optimal.”

1.3 Contributions of This Paper

This paper proposes the Beta distribution as the native mathematical model for the convergence process of Loop Agents and derives an information-theoretically optimal stopping condition based on a variance threshold. Simultaneously, we reveal cross-scale isomorphism in which evolution, human cognition, and AI Agents share the same Bayesian algorithm in their problem-solving processes. This framework provides an interpretable stopping rationale — not “we’ve run 5 rounds,” but “the variance of Beta(14, 2) has fallen below 0.01, and the marginal information gain from further iteration is not worth the computational cost.”

2Theoretical Foundations

2.1 Bayesian Inference and the Beta Distribution

The core statement of Bayes’ theorem is: posterior ∝ likelihood × prior. For repeated trials with binary outcomes (success/failure), the Beta distribution Beta(α, β) is the conjugate prior for the Bernoulli likelihood — this means that each new observation requires only a simple modification of two parameters to complete the posterior update, without recalculating the entire distribution.

A key property of the Beta distribution is that the first few observations produce the largest changes in the shape of the distribution, while the marginal information gain of subsequent observations diminishes. The shape change from Beta(1,1) (a completely uniform state of ignorance) to Beta(2,1) is dramatic, but the change from Beta(50,20) to Beta(51,20) is barely perceptible. This property will become the mathematical foundation for the Loop Agent stopping condition in later sections.

2.2 The Bayesian Nature of the Transformer

The self-attention mechanism of Transformers is functionally highly similar to Bayesian inference — both implement the core operation of “updating beliefs with evidence.” Given a Query (current belief state), the model scans all Keys (candidate evidence), generates probability weights via softmax (relative importance of evidence), and then computes a weighted sum over Values (belief integration). Each attention layer integrates information using “how important this token considers other tokens to be” — a form of subjective weighting consistent with the spirit of Bayesian subjective probability theory. It should be noted that the matrix operations of softmax attention and the mathematical form of Bayes’ rule are not strictly equivalent — they are a functional analogy rather than a structural isomorphism. However, research in 2025 has demonstrated at the behavioral level that in-context learning (ICL) approximates a Bayesian learner, and that Bayesian assumptions naturally yield ICL scaling laws, supporting the judgment of functional equivalence.

2.3 External Anchor Points and Conditional Probability

The effectiveness of Bayesian inference depends on the quality of external anchor points (evidence). Transformer output without external anchor points is equivalent to pure prior inference — “the probability of a probability” — lacking grounding in the physical world. RAG and search-grounding techniques upgrade model output from a fixed prior to a real-time updatable posterior distribution by introducing conditional probability P(A|B).

However, the effectiveness of anchor points is subject to a fundamental limitation at the thermodynamic level: the physical world continuously changes due to entropy increase, and any linguistic encoding begins to become outdated the moment it is written down. This is not an engineering problem (solvable by more data) but a thermodynamic constraint — the truthfulness of AI output has an insurmountable ceiling, determined by the gap between the rate of change of the physical world and the latency of linguistic mapping.

3Core Model: Modeling Loop Agent Convergence with the Beta Distribution

3.1 Formalization of the Loop Agent Iteration Process

We model each iteration of the Loop Agent as a Bernoulli trial. Each iteration i produces a binary outcome xᵢ ∈ {0, 1} (verification pass = 1, failure = 0). The initial prior is set to Beta(1, 1) — the uniform distribution, representing a state of complete ignorance.

The posterior mean μₙ = αₙ/(αₙ + βₙ) represents the current confidence in the solution’s correctness. The posterior variance σₙ² represents the remaining uncertainty. As iterations proceed, the variance monotonically decreases — this provides a natural mathematical foundation for the stopping condition.

3.2 Information-Theoretically Optimal Stopping Condition

We define the stopping condition as: σₙ² < ε, i.e., the variance has fallen below a preset threshold. An equivalent formulation is: the expected information gain from continued iteration is less than the computational cost of that iteration. This can be quantified by whether the KL divergence between adjacent iterations, KL(Beta(αₙ₊₁, βₙ₊₁) ‖ Beta(αₙ, βₙ)), falls below δ.

This stopping condition is adaptive: simple tasks converge in 2–3 rounds (consistent with the empirically measured average of 2.34 rounds of convergence in medical AI systems), while complex tasks may require 10+ rounds. The convergence rate is driven by data rather than set by humans.

3.3 Extension Model for Multi-Sub-Agent Parallelism

When deploying multiple sub-Agents in parallel exploration, each round can simultaneously obtain k independent observations. The parallel update becomes Beta(αₙ + Σsuccesses, βₙ + Σfailures), and under the assumption of independent observations, the variance reduction rate is approximately k times that of a single Agent. In practice, when multiple sub-Agents solve the same problem, their results often exhibit correlation (shared problem structure and knowledge base), and the effective reduction factor depends on tool diversity and the orthogonality of exploration strategies. This is mathematically analogous to population search in evolutionary biology — each sub-Agent is an individual exploring a different path, and the higher the population diversity (the more orthogonal the strategies), the closer the efficiency of parallel search approaches its theoretical upper bound.

3.4 Extension to Non-Binary Outcomes

In practice, the verification outcomes of Loop Agents are often not strict binary “pass/fail” judgments, but rather continuous quality scores ranging from 0 to 1. To address this, the model can be extended from the Beta distribution to more general forms: using weighted Beta updates (treating quality scores as observation weights rather than binary counts), or adopting the Dirichlet distribution to handle multi-category verification outcomes (e.g., a three-state classification of “fully correct / partially correct / fully incorrect”).

In the extended model, the stopping condition becomes a joint condition of quality threshold and confidence threshold: not only must the posterior variance σₙ² < ε (sufficient certainty about the conclusion), but also the posterior mean μₙ > τ (sufficiently high quality of the conclusion). The loop terminates only when both conditions are simultaneously satisfied — this prevents the degenerate case of “being highly certain about a low-quality answer.”

3.5 Theoretical Comparison with Existing Stopping Strategies

4Cross-Scale Isomorphism: A Unified Algorithm Across Evolution, Cognition, and AI

4.1 Adaptive Landscape Exploration in Evolution

Andreas Wagner, in Life Finds a Way, argues that the process by which life solves adaptive problems is not linear hill-climbing but exploration within high-dimensional adaptive landscapes. Natural selection provides local optimization (exploitation), while genetic drift and DNA recombination provide random exploration and large jumps (exploration). The entire population simultaneously conducts parallel trial-and-error at multiple positions on the landscape, with natural selection retaining the most adaptive variants — this is a natural Bayesian posterior updating process. The stopping condition in evolution manifests as follows: when fitness is sufficiently high and the environment remains stable, the population’s allele frequency distribution tends toward a steady state — variance convergence.

4.2 Path Folding in Human Cognition

When humans solve complex unknown problems, the first-round solution tends to be the safest rather than the most efficient — because fear, acting as a path filter, constrains the exploration space. But the core value of the first round lies not in “solving the problem” but in acquiring an ontological map of the problem. Ohlsson’s (1992) representational change theory demonstrates that the first success reveals the underlying causal structure, enabling the solver to eliminate entire classes of erroneous hypotheses at once. Anderson’s (1982) knowledge compilation theory describes the subsequent compression process — multi-step operation sequences are “folded” into single-step operations.

Fear plays a dual role in this process: it is both a “brake” that constrains exploration paths and a “high-pressure compressor” that forces humans to extract maximum information from minimal trial-and-error. Synaptic pruning and memory consolidation during sleep more closely resemble Bayesian model selection — among multiple competing candidate representations, low-posterior-probability connection patterns are deleted while high-posterior-probability structural connections are retained. This process differs mechanistically from parameter updates within a fixed model, but its functional effect is equivalent: uncertainty in the internal model decreases, and variance converges to a level sufficient to support efficient second-round action.

4.3 Neurochemically Driven Built-In Bayesian System

The path folding process in human cognition is driven by a complete neurochemical cascade, in which each chemical has a precise functional correspondence within the Bayesian framework:

Dopamine encodes novelty and reward prediction error, functioning within the Bayesian framework as the weight of the likelihood function — it marks which anchor points are “important,” biasing posterior updates toward high-information directions. Norepinephrine detects environmental uncertainty and triggers exploration mode switching, functionally equivalent to adjusting the openness of the prior distribution — high norepinephrine makes the prior “softer,” rendering it more susceptible to updating by new evidence. Endorphins suppress pain perception under sustained stress; their function is to maintain the subject’s dwell time in high-uncertainty regions — without endorphins, the subject would exit due to fear before variance has converged. BDNF (brain-derived neurotrophic factor) drives synaptic plasticity and new synapse formation, serving as the physical storage mechanism for the Bayesian posterior at the biological hardware level.

When glutamate accumulates in the prefrontal cortex to toxic thresholds, the brain issues a protective shutdown signal — this is the biological equivalent of max_iterations. What follows is sleep repair mode: the glymphatic system clears metabolic waste, synaptic pruning deletes noisy connections, and slow-wave activity consolidates structural memories. The complete cycle is: exploration (driven by dopamine/norepinephrine) → trial-and-error (acetylcholine maintains focus) → somatic feedback (posterior update) → glutamate overload (forced stop) → sleep repair (path folding) → new compressed prior (the “eureka moment” upon waking).

4.4 Three-System Comparison: A Unified Algorithm

5Data Latency and the Impact of Anchor Point Quality

5.1 Entropy Increase in the Physical World and Linguistic Mapping Latency

The truthfulness of AI output is subject to a fundamental limitation at the physical level. Complex systems in the physical world — climate, ecology, society, economy — continuously generate unpredictable new states due to their nonlinear dynamics. Thermodynamic entropy increase provides a directional guarantee of change (irreversibility), but the direct source of latency is the encoding difficulty of emergent behavior in complex systems: the appearance of new phenomena often exceeds the descriptive capacity of existing language and conceptual frameworks. Human language encodes these changes discretely and with delay — scientists observe data, write papers, undergo peer review, and publish, with each step introducing latency. The collection and processing of AI training data introduces yet another layer of delay.

This means that the world described by an anchor point is no longer that same world by the time you read the anchor point. No matter how large the scale or how abundant the data, the trained model always aligns with a past version of the world. This is not an engineering deficiency but a corollary of the physical fact that complex systems continuously generate new states. Mathematical axioms do not change due to entropy increase, so latency in the mathematical domain is near zero; but the chaotic dynamics of the climate system continuously produce unprecedented state combinations, leading to extremely high latency in that domain.

5.2 Domain-Specific Latency Prediction Model

Data latency varies enormously across domains, which can predict the rate of AI capability evolution in each field:

Testable prediction: The rate of AI capability evolution in each domain is inversely proportional to that domain’s data latency.

5.3 Anchor Point Quality Issues in Loop Agents

In Loop Agents, an increase in the number of available tools leads to attention dilution, equivalent to introducing excessive noisy conditions in Bayesian updating. The Berkeley benchmark shows that accuracy plummeted from 43% to 2% when tools increased from 4 to 51 — directly corresponding to the curse of dimensionality in Bayesian inference within high-dimensional conditional spaces. The optimal strategy is a small number of high-quality, precise anchor points rather than a large number of low-quality, ambiguous ones — consistent with the Beta distribution property that “the first few observations carry the most information.”

6The Evolutionary Mutation Hypothesis of Human Cognition

6.1 Abstract Thinking as a Byproduct of Evolutionary Mutation

The highly abstract thinking capacity of humans may not be a product of normal evolutionary adaptation but rather the result of a chance genetic mutation that was captured and amplified by natural selection. Evidence comes from three levels:

At the genetic level: a single-base point mutation (C→G) in the ARHGAP11B gene endowed humans with a unique capacity for neocortical expansion. The ancestral version had no ability whatsoever to increase basal progenitor cells — a single nucleotide substitution created the material basis for the human cerebral cortex.

At the network level: the connectivity between two subsystems within the default mode network that supports abstract thinking exists only in humans. In all other primates, these two systems are disconnected — this is not a difference of degree but a difference of kind.

At the temporal level: genetic variants associated with fluid intelligence and cortical morphology appeared only 300,000–500,000 years ago, accounting for less than 10% of the 6-million-year evolutionary history of the human lineage.

6.2 The Structural Conflict Between Abstract Thinking and Sociality

Humans evolved as highly social animals, and their default cognitive mode is social processing. Highly abstract thinking requires the default mode network and the executive control network to be simultaneously online and tightly coupled, while social interaction pulls both networks toward processing external social signals — the two compete for the same hardware resources. This explains why deep abstract thinking states are “anti-instinctual”: they require severing social connections (contra the social instinct), venturing into the unknown (contra the safety instinct), and repeatedly failing (contra the fear-avoidance instinct).

6.3 The Regression-to-the-Mean Effect of Education Systems

The essence of education systems is the batch installation of a standardized set of Bayesian priors into human cognition. Everyone studies the same textbooks, completes the same exercises, and is evaluated by the same criteria — the prior distribution is artificially compressed into a narrow-peaked normal distribution. This leads to insufficient variability (knowledge is presented as certain facts rather than probabilistic beliefs), insufficient accuracy (content lags behind the academic frontier by ten to several dozen years), and insufficient mutability (standardized testing penalizes prior updating). The result is that most people’s AI inputs also follow a normal distribution — standardized questions clustered around the mean rather than original hypotheses in the tails of the distribution.

7Experimental Design (Proposed)

7.1 Benchmark Comparison Experiment

Construct a cross-task evaluation set spanning four task types: multi-step reasoning, code generation, fact verification, and open-domain question answering. Compare three stopping strategies on each task type: (a) fixed max_iterations (current industry standard), (b) cosine distance-based semantic convergence detection (2026 state-of-the-art), and (c) Beta distribution variance threshold stopping condition (proposed in this paper). Evaluation metrics include: final answer accuracy, mean iteration count, computational cost (token consumption), and stopping-time optimality (deviation from the Oracle-optimal round).

7.2 Multi-Sub-Agent Parallelism Experiment

Design a comparative experiment between single-Agent serial execution and 2/4/8 parallel sub-Agents to verify the theoretical relationship between Beta distribution variance reduction rate and degree of parallelism (k parallel Agents should improve the variance reduction rate by approximately k-fold). Conduct a simultaneous cost-benefit analysis: whether the convergence acceleration from parallelism justifies the additional computational overhead. The expected result is that an optimal degree of parallelism exists, beyond which marginal acceleration is insufficient to cover the additional cost.

7.3 Anchor Point Quality Sensitivity Experiment

With task type held constant, systematically vary the number of available tools (4/8/16/32/64) and observe the convergence rate of the Beta distribution and the final posterior mean. After introducing a tool-routing pre-filtering layer, measure the improvement in convergence rate and accuracy. This is expected to verify the hypothesis that “a small number of precise anchor points is superior to a large number of ambiguous ones” — i.e., that Bayesian updating converges faster and produces a more accurate posterior with high-quality, low-volume evidence.

7.4 Cross-Domain Data Latency Verification

Select three domains with significantly different data latency — code generation, medical Q&A, and climate prediction — and measure the Loop Agent’s mean convergence rounds, final accuracy, and hallucination rate in each domain. The expected result is that convergence rounds are positively correlated with data latency (higher latency requires more iterations to converge), and final accuracy is negatively correlated with data latency. This would provide the first experimental evidence for the hypothesis that “the rate of AI capability evolution in each domain is inversely proportional to that domain’s data latency.”

8Discussion

8.1 Why the Beta Distribution and Not Another Distribution

As the conjugate prior for the Bernoulli likelihood, the Beta distribution possesses an inherent advantage in computational efficiency: updates require modifying only two parameters. It naturally models “success/failure” binary verification processes, its monotonically decreasing variance provides a built-in convergence guarantee, and it can be extended to the Dirichlet distribution for multi-category verification outcomes. By contrast, the Gaussian distribution, while also commonly used for modeling uncertainty, does not benefit from the closed-form update advantage of a conjugate prior, and its symmetry assumption is ill-suited for verification outcomes that are inherently asymmetric — naturally skewed toward success or failure.

8.2 Deep Connections to Human Cognition

The theoretical framework of this paper reveals that the human intrinsic motivation system is functionally equivalent to a “built-in Bayesian engine”: curiosity, through the dopaminergic system, automatically tags high-information anchor points; fear, through the noradrenergic system, modulates the openness of the prior distribution; and endorphins sustain sampling time in high-uncertainty regions. This system requires no external loss function — it directly implements the core operations of Bayesian inference at the neurochemical level.

The regression-to-the-mean effect of education systems gains a new interpretation within this framework: standardized education compresses the prior distribution of human cognition into a narrow-peaked normal distribution, while simultaneously damaging the prior-updating mechanism itself by penalizing “non-standard answers.” The result is that most people not only have a narrow prior space but have also lost the ability to update their priors based on new evidence — equivalent to a Bayesian system whose update function has been disabled.

Cross-domain reading constitutes a mechanism for broadening the prior distribution: each encounter with a new domain adds a dimension to the prior space, causing the coverage of the hypothesis space to grow exponentially. The optimal division of labor in human-AI collaboration follows the complementarity of three types of reasoning: humans provide abductive hypotheses (prior generator), AI provides large-scale inductive verification (likelihood calculator), and deductive reasoning is used for posterior verification. The weaknesses of the three reasoning modes are mutually compensated: the weakness of abduction (hypotheses may be wrong) is verified by inductive search; the weakness of induction (can only discover correlations) is given causal direction by abduction; and the weakness of deduction (premises must be true) is jointly ensured by the first two.

8.3 A Thermodynamic Explanation of AI Capability Boundaries

The continuous emergence of complex systems drives the inevitable latency of linguistic mapping, which determines the physical ceiling of AI truthfulness. The essence of scaling is to increase the number and diversity of anchor points, but it cannot eliminate latency. The iterative verification of Loop Agents partially mitigates latency by obtaining the latest anchor points through real-time search — but the information retrieved in each iteration round is itself still lagged. This is an ineliminable physical constraint, meaning that AI truthfulness is forever an asymptote, never an attainable state.

8.4 Abstract Thinking as a Byproduct of Evolutionary Mutation

The evolutionary origin of human abstract thinking capacity provides a meta-level footnote to the cross-scale isomorphism presented in this paper. The single-base mutation in the ARHGAP11B gene drove neocortical expansion, and the internal connectivity of the default mode network was established only in humans — these mutations endowed humans with the hardware foundation for high-dimensional abstract operations, but simultaneously made psychiatric disorders a side effect of the same genetic variation. The structural incompatibility between deep abstract thinking states and sociality (sharing the same set of neural hardware resources) means that the cognitive mode humans use for creative problem-solving is a costly “abnormal operating mode” — it commandeers neural infrastructure originally dedicated to social functions.

9Conclusion

Loop Agents represent the latest leap of AI from fixed prior probabilities to dynamic Bayesian posteriors. The Beta distribution framework proposed in this paper provides the first mathematical theoretical foundation for their convergence process, while simultaneously revealing a deep cross-scale structural analogy:

This unified perspective not only provides AI engineering with an interpretable optimal stopping condition, but also offers a cross-disciplinary theoretical framework for understanding the nature of creative problem-solving. Evolution uses genetic mutation for parallel trial-and-error and natural selection for posterior verification; the human brain uses hypothetical links for parallel trial-and-error and hormone-cascade-mediated somatic feedback for posterior verification; AI Agents use multiple sub-Agents for parallel trial-and-error and conditional probability aggregation for posterior verification. Three systems, one algorithm — Bayesian updating until convergence.

Ultimately, all inference systems based on a Bayesian structure share the same epistemological boundary: when available conditions are insufficient, conclusions drawn from known probabilistic information and limited external anchor points will have a certainty — an expected value — that is forever strictly less than 1. For any inference system S, if its input information set I is incomplete (I ⊂ Ω), then for all conclusions C, P(C|I) < 1. The variance of the Beta distribution can asymptotically approach zero but never equal it, mathematically guaranteeing convergence while also guaranteeing that absolute certainty is never reached. This is not a deficiency of the system but an honest mapping of reality: in a universe that continuously generates new states, absolute certainty does not exist.

10References