💻 computer science

Causal-Reasoning World Models: Toward AI’s Subjectivity through Backward Generation

This paper proposes a paradigm shift from noun-based to verb-based causal reasoning in AI, introducing a "possible timing distribution" framework that enables agent-like world modeling and individual-level counterfactual deduction, as demonstrated through breast cancer patient trajectories, thereby laying the theoretical groundwork for AI subjectivity.

Original authors: Jia Li, Vipin Kumar, Rui Zhang

Published 2026-10-05

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jia Li, Vipin Kumar, Rui Zhang

Original paper licensed under CC BY 4.0 (https://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

For decades, the most powerful artificial intelligence systems have been trained to recognize patterns by treating the world as a collection of static objects. In this view, a patient is a set of fixed traits, a disease is a specific label, and a treatment is a distinct event. The computer learns to draw lines between these nouns, asking questions like, "If a person has this symptom, are they likely to have that disease?" This approach has yielded impressive results, but it hits a hard ceiling. It assumes that the world is made of things that can be named and counted, ignoring the fact that life is actually a flow of actions and changes. The problem is that by forcing reality into a box of fixed categories, we lose the very thing that makes cause and effect work: the timing and the process of change itself. We know that a cause must happen before an effect, but standard models often struggle to understand how that timing shapes the outcome for a specific individual, rather than just for a large group of people on average.

A team of researchers at the University of Minnesota is proposing a radical shift in how we build these systems. Instead of asking the computer to memorize relationships between objects, they are teaching it to think in terms of actions and possibilities. They call this a move from "noun-based" modeling to "verb-based" modeling. In this new framework, the AI does not just record that a patient took a drug and then got sick. Instead, it constructs a virtual world where it can run backward from the sickness to imagine all the different ways that outcome could have happened. It asks, "What if the drug had been taken at a different time? What if the patient had a different history?" By generating these alternative scenarios, the AI builds a "world model" that exists inside the computer, independent of the physical world. This allows it to reason about cause and effect in a way that feels more like human intuition, where we understand that the same action can lead to different results depending on when and how it occurs.

The researchers tested this idea using real medical records from 3,276 breast cancer patients. They wanted to see if the AI could discover hidden patterns in how these patients moved through their treatment journeys. In traditional analysis, researchers would look for a single, average path that most patients followed. But the new system worked differently. It started with the final outcome—a patient developing heart disease, a serious complication of cancer treatment—and worked backward to generate thousands of possible timelines that could have led there. It didn't just look at the order of events; it paid close attention to the specific timing of each step. The system found that it could spontaneously identify distinct groups of patients who shared similar, significant paths through their treatment, even though those paths were not obvious when looking at the data as a whole.

One of the most striking findings was the system's ability to spot the importance of when things happened, not just what happened. For example, the AI identified specific moments in a patient's treatment where the timing of a therapy became critical. It noticed that for some patients, a particular drug administered at a certain point in their timeline was a strong predictor of heart trouble, while for others, the same drug at a different time meant nothing. The system could even simulate a "what-if" scenario for a specific group of thirteen patients who shared a complex history of diabetes and high blood pressure before their cancer treatment. By running the simulation backward, the AI deduced that for this specific group, delaying a third round of targeted therapy might have significantly reduced their risk of developing heart disease. This was not a guess based on a general rule, but a deduction derived from the unique timing patterns of that small group.

This work suggests that the future of artificial intelligence in medicine lies in its ability to handle individual stories rather than just population averages. The researchers emphasize that their system is not yet a perfect replica of human consciousness, but it is a necessary first step toward giving AI a form of "subjectivity"—the ability to occupy a position relative to the world it is studying, rather than just observing it from the outside. By constructing these virtual worlds where it can test millions of possibilities without any risk to real patients, the AI can offer hypotheses about cause and effect that traditional statistics might miss. The team found that this approach is particularly powerful for rare diseases or complex cases where patient groups are too small for standard methods to find clear answers. While the system still requires human experts to interpret and validate its findings, it provides a new tool for asking questions that were previously unanswerable, turning the computer from a passive recorder of facts into an active explorer of possibilities.

Technical Summary: Causal-Reasoning World Models: Toward AI's Subjectivity through Backward Generation

1. Problem Statement

The paper identifies a fundamental structural limitation in contemporary AI, particularly Large Language Models (LLMs) and traditional causal inference: the "Noun Trap." Current paradigms are grounded in a noun-centric ontology where modelable objects are treated as stable, well-defined entities (tokens, features, events). Relationships are encoded as functional mappings between these nouns.

This approach creates two critical issues:

Confounding and Information Loss: Representing causal relationships (e.g., $X \to Y$ ) using noun-based observables inevitably introduces residuals. To account for the gap between an observed noun and its true causal role, additional noun variables (confounders) must be added. This leads to a structural regress where increasing the number of noun variables results in a loss of relational information rather than a gain in understanding.
Inability to Model Individual Dynamics: Traditional statistics rely on the Independent and Identically Distributed (i.i.d.) assumption, which operates within a single "observed world." This framework treats time as an input or a static state (Markov assumption) rather than a dynamic, individual-level outcome. Consequently, it cannot capture the "possible timing" of events for specific individuals, limiting AI to population-level correlations rather than individual-level causal reasoning.

The paper argues that interpretability is often conflated with "nameable relations" (nouns), which bounds AI's reasoning to what can be pre-specified and named. True causal reasoning requires a shift to a verb-based paradigm, where AI constructs a world model to reason spontaneously, rather than merely recording pre-defined relations.

2. Methodology: The Verb-Based Paradigm

The proposed solution is a paradigm shift from noun-based to verb-based modeling, characterized by Backward Generation and Possibility Emergence.

Core Concepts:

Possibility Emergence: Unlike forward-directed generation (cause $\to$ effect), the proposed method performs backward generation from a targeted effect to constitute a "hyper-world." This allows AI to generate multiple instantiated scenarios (possible worlds) from a single effect, uncovering hidden causal factors and timing distributions that were not pre-specified.
Hyper-World: A collection of multiple possible worlds generated under the same causal condition (verb). Within this hyper-world, AI acts as an agent to discover significant relational information.
Possible Timing Distribution: A critical new concept that moves beyond the i.i.d. assumption. Instead of predicting a single timestamp, the model infers a distribution of "possible timing" ( $\tau$ ) across unobservable possible worlds. This timing is an individual-level construct, distinct from population-level averages.
Two-Stage Process:
1. Possibility Emergence (AI as Subject): AI captures verb-based relational information from data as latent-space representations, generating a hyper-world.
2. Causal Deduction (AI as Agent): AI operates within the constituted hyper-world to identify informative causal relationships, which can then be interpreted into nameable (noun-based) relations by humans.

Technical Implementation:

Timing Computation: Implemented using a two-level parameter architecture within a neural network.
- Population-level parameters: Frozen after the first training run, encoding the common dynamical pattern and establishing a relative timeline $\tau$ in the latent space.
- Individual-level parameters: Independently updated in subsequent training runs. This generates distinct coefficients for the transformation between observed time $t$ and the fixed timeline $\tau$ for each run, effectively constituting an individual-level timing distribution across multiple possible worlds.
Multi-Verb System: The framework treats events as verbs with their own timing distributions. It allows for inter-verb inference (e.g., $A \to B$ and $A \to C$ ) without reducing the system to a single Markov chain, thereby preserving dynamical information.

3. Key Contributions

Theoretical Framework: The paper formalizes the "Noun Trap" and proposes "Possibility Emergence" as a solution. It distinguishes between Causal Emergence (identifying gaps), Causal Representation (verb-oriented but noun-constrained), and Possibility Emergence (constructing a hyper-world for spontaneous discovery).
New Concept of Timing: It introduces Possible Timing Distribution as a computational dimension that exists across possible worlds, breaking the constraints of the i.i.d. assumption and enabling individual-level causal modeling.
Backward Generation for Causal Deduction: The paper demonstrates that backward generation from an effect allows AI to spontaneously discover causal conditions (e.g., specific timing thresholds) without pre-specifying interventions or thresholds, a capability absent in traditional forward-directed causal inference.
What-If Machine: The constituted hyper-world serves as a virtual experimental ground where AI can perform unlimited trial-and-error to answer "What-If" questions (counterfactuals) before real-world attempts.

4. Results and Applications

The methodology was applied to longitudinal Electronic Health Record (EHR) data from 3,276 breast cancer patients (Fairview Health System, 2012–2024) to predict cardiotoxicity-induced heart disease.

Application I: Multi-Verb Trajectory Discovery:
- The model spontaneously discovered significant patient trajectories involving sequential clinical events (chemotherapy, targeted therapy, radiation, diuretics).
- It identified sub-populations based on timing attention ( $\kappa$ , measured by kurtosis of the timing distribution).
- Result: The model successfully distinguished patient sub-groups and identified critical turning points (e.g., the significance of diuretic medication peaking at a specific trajectory step). It achieved high accuracy in population selection ( $C_T/C_M \approx 1$ ), demonstrating that timing computation can filter cohorts more effectively than traditional association rules.
Application II: Causal Deduction (What-If Analysis):
- The model generated a counterfactual scenario for a specific sub-population (13 patients) by flipping their outcome label (heart disease $\to$ no heart disease) and re-running the emergence process.
- Result: The comparison of factual vs. counterfactual timing distributions revealed that delaying the third targeted therapy ( $Target\_3$ ) significantly shifted the timing distribution for half the cohort, suggesting a potential reduction in heart disease risk. This deduction was derived purely from data without pre-specified hypotheses.

5. Significance and Claims

The paper positions this work as a necessary first step toward AI's subjectivity. It does not claim to have achieved human-level intelligence or a fully realized subjectivity, but rather the constitution of a "causal-reasoning world model" that allows AI to operate as an agent.

Modest Claims: The authors emphasize that the causal relationships deduced (e.g., delaying therapy) are hypotheses generated within a virtual ground, requiring further human validation. They do not claim these are definitive clinical recommendations.
Precision Medicine: The ability to perform individual-level causal deduction and fine-grained subtyping (e.g., identifying 13 specific patients out of 3,000+) is highlighted as particularly promising for rare disease care, where patient subgroups are small and population-level statistics are insufficient.
Efficiency: The paradigm shift allows for the reuse of constituted worlds, avoiding the exponential complexity of re-deriving association rules for every new inference.
Philosophical Grounding: The work draws on the philosophy of Zhao Tingyang to argue that constructing a "world" for AI requires occupying the position of a subject relative to that world, moving beyond mere statistical correlation to causal reasoning.

In summary, the paper proposes a theoretical and computational framework where AI constructs a verb-based, backward-generated world model to overcome the Noun Trap, enabling individual-level causal reasoning and the spontaneous discovery of patient trajectories and counterfactuals in complex clinical data.