When Models with Identical Accuracy Disagree on Individuals

99 concepts61 papers14 proposals

When building a machine learning model, the standard playbook is straightforward: define a loss function, train an architecture, and select the model that maximizes aggregate performance metrics like accuracy, F1-score, or area under the ROC curve.

However, this macro-level optimization conceals a striking micro-level paradox. Multiple models can achieve near-identical aggregate performance on the same dataset yet yield completely conflicting predictions for specific individuals. This phenomenon is known as predictive multiplicity (or the Rashomon effect).

If two models are equally accurate on average, but Model A approves a loan for an applicant while Model B rejects them, the automated decision becomes fundamentally arbitrary. Understanding why this happens, how to measure it, and how to navigate it has become a vital subfield of machine learning reliability and fairness.


The Root Cause: Underspecification and Stochasticity

To understand why predictive multiplicity occurs, consider a dataset where two models both achieve 90% accuracy. Both models make mistakes on 10% of the data. Crucially, standard training objectives only constrain the total number of mistakes, not which specific instances are misclassified. Because the data alone is often insufficient to uniquely pin down a single optimal decision boundary, the problem is statistically underspecified.

This underspecification means that minor, seemingly irrelevant technical choices during the training pipeline act as arbitrary tie-breakers. These include:

  • Random Seeds and Initialization: As Gomez et al. demonstrate in content moderation systems, simply changing a random seed can cause models with identical macro-accuracy to output conflicting labels for individual posts.
  • Label Indeterminacy: In high-stakes domains like healthcare, ground-truth labels are sometimes fundamentally unknowable or proxy-dependent. De-Arteaga et al. show that when labels are indeterminate, models can appear equally accurate on known cases while assigning wildly different risk assessments to the same patient.
  • Algorithmic Regularization: Even interventions designed to protect users can exacerbate multiplicity. For example, Kulynych et al. prove that the random noise injection required for differentially private training significantly expands the pool of conflicting, equi-performing models.

Quantifying the Disagreement: The Rashomon Set

To study this phenomenon rigorously, researchers analyze the Rashomon set—the collection of all models that perform within a small tolerance threshold of the optimal model.

  [ All Possible Models ]
         │
         ▼
  [ Rashomon Set ]  ◄─── Models with near-identical aggregate accuracy
    ├── Model A  ───► Predicts "Approve" for Individual X
    ├── Model B  ───► Predicts "Reject"  for Individual X
    └── Model C  ───► Predicts "Approve" for Individual X

In their foundational work, Marx et al. formalized metrics to quantify multiplicity within this set:

  • Ambiguity: The proportion of individuals in a dataset who receive conflicting predictions from at least two models in the Rashomon set.
  • Discrepancy: The maximum possible difference in predictions between any two equally accurate models for a given subpopulation.

Because training every single model in a Rashomon set to find these discrepancies is computationally prohibitive, recent work focuses on efficient exploration. Detyniecki et al. introduced the DIG (Discrepancy Intervals Generation) algorithm to map out zones of disagreement in tabular data. Taking a different approach, Hsu et al. demonstrated that applying dropout at inference time can approximate the Rashomon set up to 5,000 times faster than retraining models from scratch.


The Fairness Lottery

Predictive multiplicity is not just a statistical curiosity; it introduces severe ethical and practical challenges. When an automated system's decision changes based on a random seed, the system lacks procedural justice.

This "algorithmic lottery" does not affect everyone equally. van de Meent et al. found that standard training objectives fail to uniquely constrain minority subgroups, meaning that random seeds cause massive fluctuations in how fairly marginalized groups are treated.

Furthermore, these technical choices translate directly into implicit policy decisions. In an analysis of statistical profiling for public unemployment services, Kern et al. showed that models with identical accuracy had surprisingly low overlap in the specific populations they flagged for interventions. The choice of model pipeline effectively decided who received government support and who was excluded.


Turning Multiplicity into a Feature

While predictive multiplicity exposes a vulnerability in how we evaluate AI, it also presents an opportunity. If data performance alone cannot uniquely identify a single "best" model, we are free to navigate the Rashomon set to find models that satisfy other desirable properties.

As Cynthia Rudin argues, instead of forcing a single arbitrary choice, practitioners should embrace the Rashomon set to select models that are simpler, more interpretable, or better aligned with human domain knowledge.

This flexibility is particularly powerful for algorithmic fairness. Dai et al. prove that by being intentional during model selection, developers can discover models within the Rashomon set that drastically reduce demographic disparities with negligible loss in accuracy. This insight underpins the framework developed by Laufer et al. for identifying "Less Discriminatory Alternatives" (LDAs) in legal and regulatory compliance.


Strategies for Resolution

When a single, definitive prediction is required, how do we resolve the contradictions inherent in predictive multiplicity? Researchers have proposed several distinct paradigms:

1. Consensus and Abstention

If models disagree, we can use their variance as a signal of uncertainty. Long et al. show that ensembling multiple models from the Rashomon set provably restores individual-level consistency. To protect individuals from arbitrary errors, Lockfisch et al. combine ensembling with an abstention strategy, deferring cases where models disagree to human experts.

2. The Unanimity Principle

For applications requiring extreme precision, Fereshte Khani proposes a strict unanimity rule: a system should only emit a prediction if every valid hypothesis consistent with the training data agrees on the output. If any two valid models disagree, the system outputs an explicit "don't know."

3. Social Choice and Reconciliation

When different stakeholders deploy different models, we can resolve disagreements post-hoc. Xiong et al. demonstrate that voting mechanisms from social choice theory can eliminate the majority of predictive conflicts without harming aggregate accuracy. Alternatively, Roth et al. provide multi-party reconciliation algorithms that computationally force divergent individual probability forecasts to converge.

4. Procedural Randomization

Finally, when allocating scarce resources (like jobs or medical care), hiding behind a single deterministic model can create systemic, patterned exclusion. Jain et al. argue that because model selection is inherently arbitrary, we should replace deterministic top-$k$ cutoffs with explicit, weighted lotteries that respect individual claims, transforming hidden algorithmic arbitrariness into transparent, fair randomization.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_predictive_multiplicity_dd59509e,
  title = {predictive multiplicity},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/predictive-multiplicity/rdr_dd59509e07814181a8e04c10db9d5f2f},
  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_predictive_multiplicity_dd59509e_v1,
  title = {predictive multiplicity},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/predictive-multiplicity/rdr_dd59509e07814181a8e04c10db9d5f2f/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