Consistent Learning-to-Defer with Expert-Conditional Advice

arXiv:2603.14324 · stat.ML, cs.LG · Submitted 2026-03-15 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Consistent Learning-to-Defer with Expert-Conditional Advice".

Jane: Detailed Research Summary: Consistent Learning-to-Defer with Expert-Conditional Advice This research introduces Learning-to-Defer with advice,

Tom: First, who's behind it and why it matters.

Title and authors: Tom: Alright, let's talk about the title of "Consistent Learning-to-Defer with Expert-Conditional Advice" and who put this research out there because it really sums up the core idea.

Jane: The authors are Yannis Montreuil from the School of Computing at National University of Singapore, Leïna Montreuil from Sorbonne University in France, Axel Carlier from ENAC ISAE-SUPAERO ONERA, and Lai Xing Ng from Agency for Science, Technology and Research Institute for Infocomm Research Singapore.

Lu: It’s a diverse group of researchers coming together to tackle a problem that sits right at the intersection of decision theory and modern machine learning architectures.

Meng: The title itself points directly at the two main components: deferral, which is choosing an expert, and conditional advice, which is deciding what specific information that expert gets.

Lalam: It suggests a more sophisticated AI architecture where the choice of action isn't just one step but a coupled selection of an agent and its necessary context.

The paper's summary: Tom: So, the paper explains that traditional Learning-to-Defer assumes the information available to every expert is static at the moment you choose them, but this paper argues that's not true in real systems.

Jane: They introduce Learning-to-Defer with advice, which means the system jointly selects an expert and an expert-conditional advice action, acknowledging that what that specific person needs can change the whole outcome.

Lu: The core statistical question they are trying to answer is whether we can learn a policy that actually recovers the Bayes-optimal decision, meaning finding the expert whose best advice leads to the smallest expected cost.

Meng: They specifically show that simpler surrogates, like having independent router scores and query heads for each expert, just don't capture this true optimal risk because routing and advice are coupled in a way these simpler models miss.

Lalam: It’s about learning a joint action space over both the expert selection and the advice action, which is much richer than just picking one thing at a time.

The paper's improvements: Tom: The main improvement they propose is using an augmented surrogate loss function that operates over the composite action space of expert selection and advice action, rather than just separate components.

Jane: This loss function decomposes the objective into a baseline cost and a weighted mismatch term that penalizes deviations from the learned combined policy pi.

Lu: Their key technical achievement is establishing an H-consistency bound for this augmented surrogate, which gives us a mathematical guarantee about how well it learns the optimal strategy.

Meng: Furthermore, they proved an excess-risk transfer bound, which suggests that minimizing this augmented surrogate risk in the long run actually leads to recovering the Bayes-optimal deferral-advice risk.

Lalam: This means we can trust that training this complex system will actually result in a policy that is mathematically sound according to the true optimal strategy, even though it’s a lot more complicated to train.

Conclusion: Tom: To wrap things up, the paper shows that Learning-to-Defer with advice moves beyond simple routing by jointly selecting an expert and the specific information they require.

Jane: They demonstrated that this approach achieves H-consistency and bounds for Bayes consistency over the augmented surrogate, meaning it’s theoretically sound for finding the best way to handle expert selection and advice together.

Lu: The implication is that we can build AI systems that are much more adaptive in how they gather external data, tailoring the acquisition strategy based on which expert is being consulted.

Meng: Practically, this means our agents won't just ask for general documents; they will learn to be economical with their resources, only spending time or compute on specific advice when it actually reduces the task error significantly.

Lalam: This capability fundamentally improves AI culture by allowing systems to become more autonomous and resource-aware in their decision-making process, making them incredibly effective tools for complex problem solving.

Yannis Montreuil, Leïna Montreuil, Axel Carlier, Lai Xing Ng, Wei Tsang Ooi

School of Computing, National University of Singapore · Département de Mathématiques, Sorbonne University · Fédération ENAC ISAE-SUPAERO ONERA Université de Toulouse · Agency for Science, Technology and Research Institute for Infocomm Research

stat.ML, cs.LG

Submitted: 2026-03-15

Updated: 2026-10-05

Importance score: 88/100

The gist: This research introduces Learning-to-Defer with advice, a sophisticated framework designed to optimize decision-making in expert systems by jointly selecting both the appropriate expert and the

Key concepts

True Deferral-Advice Loss ($\ell_{ ext{def-adv}}$)
This is a formal measure of the true cost incurred by a decision policy. It accounts for both selecting an expert and the specific piece of advice that expert receives. The goal is to minimize this loss, which represents the real-world penalty for making a wrong choice in an expert system.
Augmented Surrogate
This is the new model structure used for training. Instead of just learning separate scores for experts and queries, this surrogate learns a single composite policy that decides both the expert and the specific advice action simultaneously. This allows it to capture complex interactions where the value of advice depends on which expert gets it.
H-consistency Bound
This is a mathematical guarantee proving that the augmented surrogate model is consistent with the true optimal policy. It ensures that as training progresses, the learned composite policy will eventually converge toward the mathematically best way to defer and advise.
Excess-Risk Transfer Bound
This bound shows how minimizing risk in a simplified, augmented model relates to minimizing risk in the actual Bayes-optimal system. It provides a formal link proving that training this complex surrogate actually leads to achieving the true optimal performance in the limit.

Terminology

Summary

This research introduces Learning-to-Defer with advice, a sophisticated framework designed to optimize decision-making in expert systems by jointly selecting both the appropriate expert and the specific piece of information (advice) that expert should receive. The core motivation stems from recognizing a fundamental limitation in existing Learning-to-Defer methods: they assume that the information available to every selected expert is fixed at the time of selection. In reality, modern systems are dynamic; after an expert is chosen, the system can further decide what specific advice—such as retrieved documents, tool outputs, or escalation context—that expert should process. Crucially, the value of this advice is itself dependent on which expert receives it.

The authors formally define the True Deferral-Advice Loss (def-adv), which captures the true cost incurred by a policy (r, q) given an input (x, a, y) and expert-specific advice e. The Bayes-optimal policy is defined by two sequential selections: first, the Bayes query (q(x, j)), which selects the best piece of advice for each expert j, and second, the Bayes router (r(x)), which selects the best expert based on these optimal queries.

The paper rigorously demonstrates that a broad family of simpler, separated surrogates—specifically, those consisting of independent router scores and query heads per expert—is not Bayes consistent, even in the simplest binary-advice setting. This inconsistency proves that simply optimizing the router and query components separately fails to recover the true optimal deferral-advice risk.

To overcome this, the authors introduce an augmented surrogate that operates over the composite action space = [J] times [K] 0, where J represents expert selection and K 0 represents the set of possible advice actions. The training objective is formulated as:

def-adv(pi; x, a, y, e) = D(x, a, y) + sum i in w(x, a, y) 1 f pi(x) not equal to i

This loss function decomposes the objective into an action-independent offset and a weighted mismatch term that penalizes deviations from the learned composite policy pi.

The key technical contribution is establishing an H-consistency bound for this augmented surrogate. Furthermore, through an excess-risk transfer bound, they prove that minimizing this augmented surrogate risk in the limit recovers the Bayes-optimal deferral-advice risk. This guarantees that while the training process is complex, it converges to the theoretically optimal policy under specific conditions (Theorem 8 Corollary 9).

The empirical validation across diverse benchmarks—including language, tabular data, multi-modal inputs, and synthetic tests designed to instantiate the failure modes of separated surrogates—yielded several critical findings:

  1. Dominance over Standard Deferral: The Bayes query strategy consistently selects the best advice for each expert (Lemma 3), and this leads to a policy that dominates standard deferral strategies, achieving a lower mean true loss (Lemma 5).

  2. Cost-Aware Adaptation: The method successfully adapts its advice-acquisition behavior dynamically according to the cost regime. As the escalation cost increases, the policy shifts away from expensive advice actions (e.g., deeper retrieval or higher security tiers) but does so in a manner that remains expert-dependent, rather than uniformly collapsing to no-advice actions.

  3. Expert-Dependent Trade-offs: Experiments across FEVER (retrieval advice), the sensitive-escalation fraud detection benchmark, and CLIP prompt escalation confirm that the optimal advice action is intrinsically dependent on both the chosen expert and the current cost parameter (lambda). For instance, in tabular settings, different experts might prefer different security tiers (k) based on whether escalation is cheap or expensive.

  4. Superior Performance: The method consistently attains the lowest mean true loss across all tested cost regimes (Tables 11 and 16), showing significant gains over both Learning-to-Defer (L2D) baselines and previous best fixed pair strategies, particularly when escalation costs are low.

The primary impact of this work is the formalization of how a learned system can manage the dual decision process: which expert to consult and what specific information that expert should receive. This capability allows for cost-aware acquisition of sensitive information, enabling systems to strategically trade predictive utility against escalating expert fees.

Improvements for AI systems

Based on the scientific paper Consistent Learning-to-Defer with Expert-Conditional Advice, here are the specific improvements that can be made to AI systems, categorized by technical capability:

)

The system will move beyond simple expert routing to a sophisticated, joint decision-making process involving both expert selection and information acquisition. It can perform complex tasks where the optimal action is not just who should answer? but also what external data or tool output should that specific expert receive?

It can adapt its information-gathering strategy dynamically based on the expected cost and value of different advice sources for a given expert. For instance, it can decide to query a specialized search engine only when the predicted gain in accuracy for a particular expert outweighs the acquisition fee of using that tool.

The system will exhibit cost-aware behavior in its information acquisition. It will learn to be economical with its resources, only spending computational or monetary costs on retrieving external documents, tool outputs, or escalation contexts when the expected reduction in task error justifies the expense.

It can overcome limitations of standard Learning-to-Defer (L2D) by achieving superior performance across diverse domains (language, tabular data, multi-modal). The system will be able to adapt its advice acquisition behavior to different cost regimes—be it a low-cost regime where it queries frequently or a high-cost regime where it defers to no advice.

It can be designed to explicitly handle scenarios where the value of information is highly dependent on the recipient. The system won't just fetch documents; it will select retrieval actions that are specifically tailored to maximize the performance of the chosen expert, even if that retrieval action might be irrelevant or even harmful for a different expert.

The system can overcome inherent structural failures in simpler decision models (like separated router/query surrogates). By learning directly over the composite expert-advice pair space, it will maintain Bayes consistency—meaning its learned policy is provably aligned with the mathematically optimal strategy for minimizing expected long-term cost, even when experts and advice sources are interdependent.

It can achieve provable consistency guarantees (H-consistency bounds) for its learning process. This means we have a mathematical guarantee that the surrogate loss function we use to train the system will converge to the Bayes-optimal policy as training data grows large, providing high confidence in the system's decision-making capability.

The system's performance can be rigorously benchmarked against theoretical failure modes. It can be tested on synthetic benchmarks designed specifically to expose weaknesses in decoupled decision architectures, ensuring it is robust against the exact types of errors that plague simpler AI models.

Sources

Related papers