Identity-Free Deferral For Unseen Experts

Joshua Strong, Pramit Saha, Yasin Ibrahim, Cheng Ouyang, Alison Noble

TL;DR

Identity-Free Deferral (IFD) is a new architecture for "Learning to Defer" (L2D) systems that decide when an AI should handle a task or hand it off to a human expert. By architecturally enforcing permutation symmetry—ensuring the system cares about structural competence rather than arbitrary class labels—IFD generalizes significantly better to new, out-of-distribution experts while requiring far fewer expensive human annotations during training.

What Problem Does This Paper Address?

In high-stakes fields like radiology or autonomous driving, AI systems shouldn't just predict; they should know when to defer to a human expert. This is the core of Learning to Defer (L2D).

The "holy grail" of L2D is building a system that can adapt to unseen experts at test time. Current state-of-the-art models (like L2D-Pop) try to "meta-learn" an expert's profile from a few past examples. However, these models have a hidden flaw: they are sensitive to the arbitrary indexing of classes. If a model learns that "Expert A is good at Class 3," it might fail if a new expert is good at "Class 5," even if the underlying specialty pattern is the same.

The authors diagnose this as identity leakage. Because standard neural encoders process class signals in fixed coordinates, they learn "shortcuts" (e.g., "if index 3 is strong, defer") that violate the problem's inherent symmetry. When experts are out-of-distribution (OOD)—perhaps they have a different specialty or work in a hospital with different labeling conventions—these shortcuts break, leading to poor collaboration.

Methods

The authors introduce Identity-Free Deferral (IFD), which replaces high-capacity, identity-prone encoders with a lean, structurally-aware interface.

  1. Bayesian Competence Profiles: Instead of a black-box embedding, IFD builds an explicit Bayesian profile for each expert using Beta-Binomial distributions. It tracks how often an expert is correct for each class, providing both a mean estimate of skill and a measure of uncertainty.
  2. Role-Indexed State: This is the "secret sauce." Instead of showing the rejector a vector of all $K$ classes, IFD only shows it information relative to specific roles. It looks at the model’s top-ranked class and the expert’s estimated best class. The rejector sees a tiny 6-dimensional vector (e.g., model confidence at its top class, expert skill at that same class) that completely obscures absolute class IDs.
  3. Permutation Invariance by Construction: The authors formally prove that IFD is invariant to "coherent relabeling." If you swap the names of "Cat" and "Dog" everywhere in the system, IFD’s decision remains identical, whereas standard models would likely change their minds.
  4. Context-Only Training: IFD uses an uncertainty-aware loss function that supervises the system using only the expert's past performance (context), removing the need for expensive expert labels for every new query during training.

Datasets

The researchers evaluated IFD across several high-stakes and general benchmarks:

  • Medical Imaging: HAM10000 (dermatology), Blood Cells (microscopy), and Liver Tumors (radiology).
  • ImageNet-16H: A subset of ImageNet featuring real human annotations, used to test robustness against varying levels of image noise.
  • Simulated Experts: They introduced a new, more challenging simulation protocol that models "Variable Specialists"—experts whose accuracy fluctuates based on how "prototypical" or difficult a specific image is.

Key Insights and Ideas

  • The "Aha" Moment on Symmetry: The realization that L2D is fundamentally a permutation-symmetric problem. The decision to defer should depend on the relationship between the model's belief and the expert's skill, not the name of the category.
  • Structural vs. Semantic: By moving to a "role-indexed" state, the model is forced to learn the logic of deferral (e.g., "defer if the expert is significantly more confident in their specialty than I am in mine") rather than memorizing specific class-expert pairings.
  • Uncertainty as a Weight: By using a Lower Confidence Bound (LCB) in the loss function, the system naturally learns to trust experts less when their profile is based on very few examples, preventing "over-deferral" to lucky guesses.

Key Results

  • Superior OOD Generalization: IFD consistently outperformed baselines when facing OOD experts. In the Liver Tumor dataset, IFD improved system accuracy (SAC) by up to 10% over the next best method.
  • Annotation Efficiency: Because IFD doesn't require query-time expert labels, it is significantly cheaper to train. In the Blood Cells dataset, it achieved better results while using roughly 100x fewer expert annotations than L2D-Pop.
  • Scaling with Context: Unlike meta-learned encoders that often plateau, IFD’s performance improves monotonically as it sees more of an expert's history, effectively turning more data into better Bayesian priors.

Limitations and Future Work

  • Class-Level Granularity: IFD currently summarizes expert skill at the class level. While this is robust, it might miss "instance-level" nuances (e.g., an expert who is great at identifying tumors in blurry scans specifically). The authors suggest that combining IFD's structural bias with per-instance modeling is a promising next step.
  • Complex OOD Shifts: The current proofs focus on relabeling. Future work could explore more complex shifts, such as experts whose specialties drift independently of the model's posterior.
  • Architectural Remedies: The authors suggest that weight-tying or permutation-equivariant encoders could further strengthen these "identity-free" principles in even larger models.

Summary updated

Abstract

Learning to Defer (L2D) improves AI reliability in decision-critical environments by training AI to either make its own prediction or defer the decision to a human expert. A key challenge is adapting to unseen experts at test time, whose competence can differ from the training population. Current methods for this task, however, can falter when unseen experts are out-of-distribution (OOD) relative to the training population. We identify a core architectural flaw as the cause: they learn identity-conditioned policies by processing class-indexed signals in fixed coordinates, creating shortcuts that violate the problem's inherent permutation symmetry. We introduce Identity-Free Deferral (IFD), an architecture that enforces this symmetry by construction. From a few-shot context, IFD builds a query-independent Bayesian competence profile for each expert. It then supplies the deferral rejector with a low-dimensional, role-indexed state containing only structural information, such as the model's confidence in its top-ranked class and the expert's estimated skill for that same role, which obscures absolute class identities. We train IFD using an uncertainty-aware, context-only objective that removes the need for expensive query-time expert labels. We formally prove the permutation invariance of our approach, contrasting it with the generic non-invariance of standard population encoders. Experiments on medical imaging benchmarks and ImageNet-16H with real human annotators show that IFD consistently improves generalisation to unseen experts, with gains in OOD settings, all while using fewer annotations than alternative methods.
BibTeX
@inproceedings{strong2026identity,
  author = {Joshua Strong and Pramit Saha and Yasin Ibrahim and Cheng Ouyang and Alison Noble},
  title = {{Identity-Free Deferral For Unseen Experts}},
  booktitle = {International Conference on Learning Representations},
  year = {2026},
  url = {https://openreview.net/forum?id=4YG9ufFg58},
}
Figure