To Use AI as Dice of Possibilities with Timing Computation

Technical Summary: To Use AI as Dice of Possibilities with Timing Computation

Problem Statement
The paper argues that the dominant "noun-based" modeling paradigm in AI—exemplified by Large Language Models (LLMs) and standard machine learning approaches—fundamentally constrains the representation of the future. In this paradigm, models treat objects as stable, well-defined entities (tokens, features, data points) and relationships as functional mappings ($f(x)=y$). While effective for outcome prediction, this approach fails to represent the future as an open temporal dimension. It relies on speculation regarding which variables to include, often substituting dense correlations for causality. Furthermore, existing causal inference methods, while utilizing counterfactuals, still operate under the assumption that relevant variables are closed-ended and exhaustively specified in advance. Consequently, current models cannot adequately represent "choosable futures" or perform genuine counterfactual reasoning regarding when events might occur, as they treat timing as a fixed, observed input rather than a computable variable.

Methodology: The Verb-Based Paradigm
The authors propose a "verb-based" paradigm where the modeling target shifts from a single, determined future to the complete space of possible futures. The core innovation is Timing Computation, defined as the capacity to treat time not merely as a learnable input feature, but as a computable dimension derived from the model's internal logic.

Key methodological components include:

• Possibility Variables: The authors redefine variables from $X \in \mathbb{R}^n$ to possibility variables $X = (X, t) \in \mathbb{R}^{n+1}$, where $t$ is a computable timing dimension.
• Timing Learnable vs. Timing Computable: The paper distinguishes between "timing-learnable" (extracting temporal info from ground truth) and "timing-computable" (predicting timing distributions). Existing methods lack the latter because they lack a formal definition of the Distribution of Possible Timing.
• Neural Network Implementation: The framework utilizes neural networks trained via backpropagation. Unlike forward modeling where $f(x)=y$ is pre-specified, this approach reasons backward from a determined outcome ($y$) to recover the historical significance of past events ($x$).
• Virtual Timeline ($\tau$): The model constructs a latent temporal space where input timestamps are treated as ground truth, but a virtual relative timing $\tau(t)$ is defined by trainable parameters. This allows the model to generate multiple independent instantiations of the same event, creating a distribution of possible timings rather than a single point estimate.
• Timing Attention: The concentration of the timing distribution across a patient cohort is measured via kurtosis (termed "timing attention"). Events with higher timing attention exhibit more concentrated distributions, indicating greater temporal significance.

Key Contributions

1. Theoretical Framework: The paper introduces a verb-based ontology where "verbs" (actions/choices) represent open-ended possibilities, and "timing" is a computable variable essential for causal reasoning. It posits that AI should function as "dice of possibilities"—rolling the space of open futures into view—rather than as an oracle delivering fixed answers.
2. Formal Definition of Timing Computation: The authors provide the first formal definition of timing computation in a neural network context, distinguishing it from standard survival analysis or time-series prediction.
3. Data-Driven Discovery without Prior Knowledge: The framework demonstrates that clinically significant trajectories and counterfactual timing deductions can be achieved purely through data-driven methods, requiring no prior domain knowledge or pre-specified variable sets.

Results
The framework was applied to longitudinal Electronic Health Record (EHR) data from 3,276 breast cancer patients to predict the onset timing of cardiotoxicity-induced heart disease. The study reports two primary empirical demonstrations:

1. Automatic Discovery of Patient Trajectories:

   • The model automatically identified clinically significant patient trajectories without pre-defined rules.
   • Timing Attention ($\kappa$): Events with larger cohort sizes consistently showed higher timing attention.
   • Trajectory Convergence: The model identified that different intermediate pathways could merge at a shared terminal event (e.g., a second radiation therapy), indicating equivalent temporal significance.
   • Critical Turning Points: The model detected specific events (e.g., administration of diuretics) where the positive rate of cardiotoxicity peaked and then declined, contrasting with monotonic trends in other pathways.
   • Efficiency: The approach evaluated events concurrently without logical ordering constraints, avoiding the exponential search complexity of traditional association rule-based methods.

2. Counterfactual Timing Deduction:

   • The model performed counterfactual reasoning by assuming positive patients (those who developed cardiotoxicity) remained negative.
   • Timing Shifts: The resulting distributions showed that for approximately half the cohort, the relative timing of the third targeted therapy (Target_3) shifted markedly toward larger values under the counterfactual assumption.
   • Clinical Implication: This suggests that postponing or canceling the third targeted therapy could substantially reduce the risk of cardiotoxicity onset for these specific patients.

Significance and Claims
The paper claims these results represent the first such demonstrations in the machine learning literature for both automatic trajectory discovery and counterfactual timing deduction using purely data-driven methods.

The authors argue that this work addresses a fundamental gap between the "grammar of thought" (which is future-directed and choice-oriented) and current modeling paradigms. By treating the "factum" (the conceptual content of a causal scenario) as a primitive unit within a possibility space, the verb-based paradigm:

• Dissolves the confounding problem inherent in noun-based inference by absorbing residual causal structure into an open possibility space rather than misattributing it to observed variables.
• Provides a foundation for a new class of generative models ("World Models") that are structurally aligned with human conceptual structures, moving beyond the statistical token associations of current LLMs.

The paper explicitly directs readers to an accompanying technical paper (Li et al., 2026) for implementation details, maintaining that the current work establishes the conceptual and empirical necessity of shifting from noun-based to verb-based modeling.