Surrogate Rationalization in AI

Updated 3 June 2026

Surrogate rationalization is a technique that builds human-understandable approximations of complex AI models through simplified surrogates or input masks.
Recent methodologies integrate norm-based and hybrid objectives to overcome limitations of traditional mutual information–based approaches.
Applications include local surrogate fitting, rule-based modeling, and probing neural network subspaces to enhance model faithfulness and interpretability.

Surrogate rationalization denotes a class of techniques in explainable artificial intelligence and machine learning for producing interpretable rationales—compact, human-understandable explanations—for complex or opaque predictive models. The surrogate may take the form of a local or global interpretable model, a masked or extracted input subset, or an explicit rule set, all selected to rationalize the original model’s behavior while maintaining a quantifiable trade-off between fidelity, interpretability, and coverage. Recent developments challenge conventional mutual information–based rationalization schemes and reveal failure modes such as post-hoc rationalization and faithfulness dilution, propelling new objective criteria and formats for surrogate construction.

1. Fundamental Principles and Notational Framework

Surrogate rationalization formalizes the process of constructing a simpler, interpretable model or input mask (the surrogate) to explain, justify, or mimic the decisions of a more complex black-box function $f$ . Depending on context, this surrogate can be:

A mask $M$ applied to input $X$ that selects a rationale $Z = M \odot X$ such that a predictor $f_P(Z)$ closely mimics the black-box’s output.
An interpretable model $g \in \mathcal{G}$ (e.g., shallow tree, sparse linear model) trained to approximate $f$ over a domain or local region.

Three central axes structure surrogate rationalization:

Fidelity ( $\mathcal{L}_{\mathrm{fid}}(g;f,R)$ ): Closeness of surrogate predictions/rationales to those of the black-box model over the region of interest $R$ .
Complexity ( $\mathcal{C}(g)$ or model size): Interpretability cost, e.g., rule length, tree depth, mask sparsity.
Coverage ( $M$ 0): The spatial or semantic extent covered by the surrogate or its rationales.

A typical objective involves multi-objective or constrained optimization, e.g.

$M$ 1

with constraints on coverage or fidelity, depending on the application (Poyiadzi et al., 2021).

2. Maximum Mutual Information and the Diminishing Returns Problem

A dominant axis in extractive rationalization (e.g., Rationalizing Neural Predictions, RNP) is the use of maximum mutual information (MMI) as the criterion for rationale extraction. Given an input sequence $M$ 2 and binary mask $M$ 3, the goal is to maximize $M$ 4 where $M$ 5 and $M$ 6 is the label:

$M$ 7

This is typically instantiated by sharing gradients between the extractor $M$ 8 and predictor $M$ 9, minimizing classification loss on the masked input. However, empirical and theoretical analysis demonstrates that MMI rationalization exhibits diminishing marginal returns: once an initial set of highly informative tokens is masked in, further inclusion of truly relevant rationale fragments yields minimal incremental gain in mutual information due to subadditivity:

$X$ 0

Softmax or sigmoid output heads in the predictor saturate, further flattening gradients and impeding complete rationale recovery—an effect most pronounced when reconstructing rationales close to full label predictivity (Liu et al., 8 Mar 2025).

3. Surrogate Norm-Based Rationalization: Probing Input Utilization

To overcome the limitations of MMI, recent work introduces surrogate rationalization objectives based on direct probing of the network’s input utilization. The central insight is that neural networks often learn low-rank weight matrices $X$ 1, and only input projections lying in the subspace spanned by $X$ 2 meaningfully influence predictions. Therefore, a candidate rationale $X$ 3 is preferable if its hidden encoding (e.g., penultimate layer output) lies within this subspace, which manifests as a high $X$ 4 norm:

$X$ 5

The surrogate objective for the extractor is thus:

$X$ 6

while the predictor continues to minimize standard cross-entropy. Optionally, a hybrid loss ("MMI+N2R") combines both terms, initially relying on MMI when norms are small, then shifting to the norm objective for more informative gradients. This approach avoids the diminishing returns of MMI and more robustly identifies network-utilized input components (Liu et al., 8 Mar 2025).

4. Surrogate Rationalization in Practice: Algorithms and Empirical Performance

Algorithmic instantiations of surrogate rationalization span local and global surrogates, mask-based extractors, and transparent rule-based models. Methodologies include:

Local surrogate rationalization: Fit simple surrogates such as LASSO or shallow trees over neighborhoods $X$ 7 of a focal instance, enabling favorable Pareto trade-offs among fidelity, complexity, and coverage. Interactive routines allow users to select coverage and interpret results across a spectrum of surrogate granularities (Poyiadzi et al., 2021).
Input mask rationalization: Train a binary mask extractor with norm- or information-based objectives, then evaluate precision/recall of extracted rationales against human-annotation gold standards.
Rule-based surrogate models: Train interpretable surrogates (e.g., trees) to mimic black-box behavior on observed states, with regularization on depth or feature usage for interpretability. Explanations can be rendered via causal attributions through explicit rules or feature attributions (e.g., SHAP) (Gavriilidis et al., 2023).
Cooperative rationalization in OOD settings: Split input graphs into rationales and non-rationales, aligning rationale representations with classifier embeddings via mutual information and knowledge distillation, while inferring environments based on non-rationale clustering. This yields improved OOD generalization and more robust invariant rationales (Yue et al., 2024).

Recent empirical results demonstrate that surrogate norm-based rationalization (N2R) provides consistent gains over MMI and its enhancements across text and graph domains. For instance, on text datasets at fixed 15% token sparsity, N2R raises F1 from ≈66.5 (MMI) to ≈72.7, and gains persist across BERT and GCN architectures. Hybrid criteria (MMI+N2R) yield further improvements (Liu et al., 8 Mar 2025).

Dataset	MMI F1	N2R F1	MMI+N2R F1
Beer-Appearance	66.5	72.7	75.2
Beer-Aroma	65.8	68.7	—
Hotel-Service	46.8	48.6	—
Hotel-Cleanliness	35.6	39.8	—
BA2Motifs (graph)	36.5	40.2	—

5. Faithfulness and Failure Modes: Surrogate Rationalization in Generative Models

Surrogate rationalization is susceptible to faithfulness breakdowns, particularly in generative systems where justifications may serve as post-hoc rationales rather than genuine evidence of reasoning. For masked diffusion LLMs (MDLMs) on fact verification, verdict tokens converge to stable decisions early, before any justification is generated. Enforcement of a “reasoning-first” constraint by delaying verdict unmasking actively degrades accuracy—from 86.2% (no delay) to 71.9% (deliberation on 90% justification tokens). Intervention experiments establish that verdicts are heavily causally dependent on justification quality: with corrupted justifications, accuracy falls to 57.3% (ground truth: 97.1%), and the model rationalizes forced-wrong verdicts in 56% of cases. These phenomena indicate that, within such architectures, surrogate justifications generally constitute retroactively constructed rationales anchored to an already-fixed verdict, not bona fide reasoning chains (Devasier, 1 Mar 2026).

6. Extensions, Limitations, and Open Problems

Advances in surrogate rationalization suggest several extensions and expose distinct limitations:

Extensions: Norm-based objectives are readily applied for task-agnostic explanations (e.g., probing frozen encoders), and can be combined with regularizers for rationale coherence or sparsity. The approach generalizes to other modalities, including images and graphs, by leveraging norms of hidden representations (Liu et al., 8 Mar 2025, Yue et al., 2024).
Limitations: Surrogate rationalization based on network subspace alignment assumes that the learned weight matrix accurately encodes all task-relevant information; misalignment or biased training may result in rationales inconsistent with human expectations. Access to intermediate representations often requires white-box access, limiting applicability in black-box regimes. Hybrid objectives may be necessary to provide adequate gradient signal early in training.
Failure modes in generative systems: Extended justification generation (“deliberation”) in MDLMs can be counterproductive due to refinement drift, with noisy or inconsistent justifications undermining initial correct decisions.

Ongoing research aims to develop faithfulness regularization, disentangled decoder heads, early stopping protocols, and verification modules to increase the reliability of surrogate rationales in modern architectures (Devasier, 1 Mar 2026).

7. Relation to Surrogate Endpoints in Causal Inference

Surrogate rationalization is conceptually related to the selection of surrogate endpoints in causal inference, where a surrogate variable stands in for a primary outcome. However, even with apparent positive causal effects along both treatment-surrogate and surrogate-outcome edges, unmeasured confounding can result in a negative average causal effect on the true outcome—a phenomenon termed the surrogate paradox. Recently developed optimality criteria are testable based solely on observed data and exclude the surrogate paradox only when specific estimable inequalities are satisfied, underscoring the challenge of ensuring that a surrogate variable delivers faithful causal rationalization in clinical and policy settings (Yin et al., 2016).

Surrogate rationalization thus encapsulates a wide spectrum of techniques for producing interpretable approximations of black-box model behavior, from mask-based rationale extraction and local surrogate fitting to explicit rule-based rationalizers and causal surrogate endpoints. Key challenges include navigating the trade-off between fidelity, interpretability, and coverage, ensuring faithfulness in the face of generative rationalization, and accommodating model and domain–specific constraints. Recent norm-based and mutual information–based objectives, together with compositional architectures and causal tests, collectively define the current research frontier.