In high-stakes domains like medicine or autonomous driving, an AI system should not always act alone. Instead, it should operate within a human-AI pipeline, deciding when to provide a prediction and when to hand the task to a human expert. This paradigm, known as Learning to Defer (L2D), shifts the goal from maximizing standalone model accuracy to maximizing the accuracy of the combined system.
The fundamental challenge in L2D is training: how do we mathematically optimize a model to know its own limits relative to a specific expert? The answer lies in the development of consistent surrogate losses.
The Collaborative Objective
In a standard classification task, we minimize the probability of the model being wrong. In L2D, we minimize the probability of the system being wrong. If the model predicts, the system errs if the model is wrong. If the model defers, the system errs if the expert is wrong.
As Hussein Mozannar and David Sontag argue in Consistent Estimators for Learning to Defer to an Expert, this creates a joint optimization problem. We aren't just learning the data distribution; we are simultaneously learning the "error distribution" of the expert.
The "true" loss for this system is a discrete 0-1 loss that counts an error if:
- The model predicts and is incorrect.
- The model defers and the expert is incorrect.
Because this 0-1 loss is non-differentiable and non-convex, we cannot use it directly with gradient descent. We need a surrogate.
The Necessity of Consistency
A surrogate loss is not useful just because it is differentiable; it must be consistent. Consistency is a mathematical guarantee that as we provide more data, minimizing the surrogate loss will lead us to the same optimal decision boundary as the true 0-1 loss.
Without consistency, a model might achieve a low surrogate loss while making fundamentally wrong deferral decisions—for example, deferring to an expert on a task where the model is actually more competent. Nontawat Charoenphakdee and colleagues demonstrated in Classification with Rejection Based on Cost-sensitive Classification that many intuitive "rejection" losses fail this test, leading to suboptimal performance even with infinite data.
Recent research has refined this into two critical types of consistency:
- Bayes-Consistency: The surrogate converges to the best possible theoretical decision rule (the Bayes optimal).
- $\mathcal{H}$-Consistency: Introduced by Anqi Mao, Mehryar Mohri, and Yutao Zhong in Realizable H-Consistent and Bayes-Consistent Loss Functions for Learning to Defer, this ensures the loss remains optimal even when we are restricted to a specific class of models (like neural networks) that might not be able to represent the true Bayes rule perfectly.
Architecting the Deferral Mechanism
How do we actually build these losses? The research has converged on several key architectures:
One-vs-All and Softmax Parameterization
Early L2D models often struggled with calibration, sometimes estimating an expert's accuracy to be over 100%. Eric Nalisnick and Rajeev Verma addressed this in Calibrated Learning to Defer with One-vs-All Classifiers by treating the deferral option as a distinct binary task within a One-vs-All framework.
Alternatively, Yuzhou Cao and Lei Feng showed in In Defense of Softmax Parametrization for Calibrated and Consistent Learning to Defer that the standard softmax can be consistent if we use an Asymmetric Softmax Parameterization (ASM). This asymmetry accounts for the fact that the "cost" of the model's error is different from the "cost" of the expert's error.
Predictor-Rejector Frameworks
In many modern applications, we cannot retrain the base model (e.g., a massive pre-trained LLM). In these cases, we train a separate "rejector" or "router." Christopher Mohri and Anqi Mao developed consistent surrogate losses specifically for this two-stage setup in Learning to Reject with a Fixed Predictor, allowing a small auxiliary model to optimally gate a large, fixed model.
Scaling to Real-World Complexity
The intuition of consistent deferral is now being applied to increasingly complex scenarios:
- Multiple Experts: When choosing between many experts, systems can suffer from "underfitting" where the model fails to learn because it relies too heavily on experts. Shuqi Liu and colleagues proposed the PiCCE loss in When More Experts Hurt to restore classifier performance by focusing on the most reliable expert during training.
- Sequential Tasks: In Learning to Partially Defer for Sequences, Ambuj Tewari and Sahana Rayan extended consistency to sequences, allowing a model to defer specific tokens or steps rather than the entire task.
- Active Learning: To reduce the cost of gathering expert labels, Mohammad-Amin Charusaie and Samira Samadi introduced Disagreement on Disagreements (DoD) in Sample Efficient Learning of Predictors that Complement Humans, an active learning strategy that uses consistent surrogates to identify which human-expert data will most improve the system.
By grounding the decision to defer in the rigorous theory of consistent surrogate losses, researchers are moving away from "black-box" confidence scores toward systems that have a mathematically guaranteed understanding of when a human—or another model—is better equipped to handle the task.