Resolving the Reference Class Problem for Individual Probabilities

19 concepts15 papers1 proposal

Imagine a doctor tells you that you have a 70% chance of surviving an upcoming surgery. What does that number actually mean?

From a strict frequentist perspective, probability is defined by long-run relative frequencies across repeated trials. But your surgery is a unique, non-repeatable event: you will either survive or you won't. There are no repeated trials for you. To generate that 70% figure, the doctor must place you into a group of similar patients—a reference class—and look at the survival rate within that group.

Herein lies the dilemma: you belong to an infinite number of potential reference classes. Should the doctor look at the statistics for patients of your age? Patients with your specific medical history? Patients undergoing surgery at this specific hospital? If these overlapping groups yield wildly conflicting frequencies, which one represents the "objective" probability for your life?

This puzzle is known as the Reference Class Problem. For decades, it created a sharp philosophical divide between objective frequentists (who focus on population statistics) and subjective Bayesians (who treat probabilities as personal degrees of belief). However, a rich body of research across logic, information theory, and computer science has sought to bridge this gap, transforming the reference class problem from a philosophical paradox into a solvable computational challenge.

The Core Conflict: Populations vs. Individuals

The tension begins with how we define uncertainty. As Peter C. Cheeseman argued in In Defense of Probability, restricting probability purely to frequentist repetitions severely limits an intelligent agent's ability to make decisions in unique, one-off contexts. Yet, relying entirely on ungrounded subjective opinions feels equally unscientific.

Henry E. Kyburg Jr. challenged this dichotomy in Objective Probabilities, arguing that the divide is artificial. Every unique event belongs to various classes for which empirical data does exist. The challenge is not a lack of data, but rather structural ambiguity: when an individual satisfies the criteria for multiple reference classes, we face a logical conflict among competing frequencies.

Without a formal mechanism to navigate these overlapping classes, automated reasoning systems struggle. For example, as Carsten Lutz et al. noted in Probabilistic Description Logics for Subjective Uncertainty, knowledge bases must explicitly distinguish between general population defaults (e.g., "90% of patients recover") and an agent's subjective confidence regarding a specific individual who might possess unique, complicating traits.

Formalizing the Bridge: From Statistics to Beliefs

To resolve this, researchers first attempted to build rigorous logical frameworks that could systematically map population statistics down to individual degrees of belief.

One prominent approach is Evidential Probability, introduced by Henry E. Kyburg Jr. in Evidential Probability. Instead of picking a reference class arbitrarily, Evidential Probability provides a set of logical rules to select the most specific and informative reference class available, systematically discarding broader or conflicting classes that contain less precise data. Bülent Murtezaoglu et al. later refined this framework in A Modification to Evidential Probability to prevent conflicting evidence from unduly weakening the system's final conclusions.

A parallel breakthrough came from applying the Principle of Indifference to first-order logic. In From Statistics to Beliefs, Fahiem Bacchus et al. introduced the "Random Worlds" method. If all you know is a set of statistical facts about a population, the probability you should assign to an individual is derived by looking at all possible worlds that satisfy those statistics and counting how many of them the individual possesses the trait in question. This provided a mathematically sound inductive step, moving past ad-hoc heuristic rules.

Similarly, Joseph Halpern et al. in First-Order Conditional Logic Revisited and Vaishak Belle et al. in A First-Order Logic of Probability and Only Knowing in Unbounded Domains designed epistemic logics where population-level defaults naturally and fluidly shape an agent's internal, subjective belief states without triggering logical paradoxes.

Information Theory and Cross-Entropy Minimization

What happens when our statistical knowledge about the population is incomplete or vague? How do we avoid making arbitrary assumptions when mapping those statistics to an individual?

Manfred Jaeger addressed this in A Logic for Default Reasoning About Probabilities and Minimum Cross-Entropy Reasoning: A Statistical Justification by leveraging minimum cross-entropy.

[Incomplete Population Statistics] 
               │
               ▼
┌──────────────────────────────┐
│  Cross-Entropy Minimization  │ ◄─── Acts as a "Thought Experiment"
└──────────────────────────────┘
               │
               ▼
[Unbiased Individual Degrees of Belief]

The intuition behind this mechanism is elegant: treat your internal reasoning as a hypothetical "thought experiment." You start with a prior belief structure, and when presented with objective population frequencies, you update your beliefs only as much as necessary to satisfy the new data. Minimizing cross-entropy ensures that you incorporate the population constraints while remaining completely unbiased about everything else, providing a clean semantic bridge between general data and specific instances.

Modern Computational Solutions: Indistinguishability and Scale

In the modern era of machine learning, the reference class problem has re-emerged with high stakes. When an algorithm assigns a "risk score" to an individual for a loan default or medical diagnosis, how do we evaluate if that individual probability is correct?

In Outcome Indistinguishability, Michael P. Kim et al. reframed the problem using computational complexity and cryptography. They argued that while we cannot verify a probability for a single non-repeatable event, we can ensure that the predicted probabilities are indistinguishable from reality to any computationally bounded observer. If no efficient algorithm can look at a subgroup of individuals and successfully detect a systematic bias in the risk scores, the scores are effectively calibrated across all relevant reference classes simultaneously.

Furthermore, recent work demonstrates that while the reference class problem is mathematically unsolvable for an isolated individual, it can be mitigated when looked at across entire populations.

Aaron Roth et al. in Reconciling Individual Probability Forecasts✱ and Alexander Tolbert in Resolving the Reference Class Problem at Scale introduced the concept of model reconciliation. If two different models assign different probabilities to the same person, it means they have chosen different reference classes.

By computationally identifying these points of disagreement across a large population, the authors show that we can constructively falsify inaccurate groupings. Forcing competing models to reconcile their disagreements drives them to converge on identical, highly refined individual forecasts.

Ultimately, this modern synthesis shows why the idea works: we do not need to find a single, mystically "correct" reference class. Instead, by leveraging logical constraints, information theory, and computational indistinguishability, we can bound subjective uncertainty using objective data—giving meaningful, actionable weight to individual probabilities.

Key Foundational Work

Objective Probabilities (1987) — contributed 3 concept elements to this direction.
Objective Probabilities (1987) — contributed 3 concept elements to this direction.
In Defense of Probability (1985) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_reference_class_problem_cfa178f8,
  title = {reference class problem},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/reference-class-problem/rdr_cfa178f8d7704eb392ffdd3b382d9bd5},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_reference_class_problem_cfa178f8_v1,
  title = {reference class problem},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/reference-class-problem/rdr_cfa178f8d7704eb392ffdd3b382d9bd5/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake