New AI Framework Distills Thousands of Explanations Into Just Five Models
Researchers have developed an optimization method that compresses redundant artificial intelligence explanations into a tiny, highly accurate proxy set.
Complex artificial intelligence systems often act as opaque black boxes, leaving researchers and engineers struggling to understand why a model makes specific predictions. A newly developed optimization framework can compress thousands of ambiguous local model interpretations into a tiny, highly accurate group of just five or six proxy models without sacrificing predictive fidelity. This breakthrough makes it feasible to grasp the holistic behavior of complex machine learning models used in high-stakes fields such as medicine, atmospheric chemistry, and particle physics.
Many popular interpretation techniques rely on post-hoc local surrogate models, which are simple, human-understandable approximations like linear regressions or decision rules designed to mimic a complex black-box function within a small neighborhood of data. However, these local explanations suffer from two major flaws: ambiguity and redundancy. Because multiple simple models can explain the same data point equally well, researchers face interpretation ambiguity. Furthermore, if a closed-box function is reasonably smooth, a single local model remains valid across a wide region of input space. Generating a unique explanation for every single data point produces massive redundancy.
To turn this redundancy into an advantage, researchers introduced a new computational procedure called ExplainReduce. As described in a study published in the journal Explainable Artificial Intelligence (DOI: 10.1007/978-3-032-45473-7_15), this approach treats explanation reduction as a formal optimization problem, distilling an overwhelming collection of local models into a compact proxy set.
How Optimization Solves Explanation Redundancy
The ExplainReduce framework operates in three distinct phases. First, a large collection of local explanations is generated using pre-existing interpretation methods such as LIME, SHAP, or SLISEMAP. Second, the reduction task is formalized as a submodular optimization problem that balances model fidelity, coverage, and stability. Third, the system solves this problem using efficient greedy algorithms backed by mathematical performance guarantees.
The authors formulated three specific problem variations: maximizing coverage to ensure most data points are explained adequately, minimizing loss to optimize average faithfulness, and balancing both objectives simultaneously. Because these objective functions exhibit submodular properties, greedy heuristic algorithms can solve them efficiently while guaranteeing that the resulting proxy set captures at least sixty-three percent of the optimal improvement.

Empirical Performance Across Diverse Real-World Datasets
The research team tested their framework across twelve different synthetic and real-world datasets involving regression and classification tasks. Their experiments demonstrated that selecting as few as five or six proxy models was sufficient to match the test set fidelity of an ensemble containing five hundred local explanations. Moreover, testing revealed that researchers do not need to train local models for every single item in a dataset, as randomly subsampling around five hundred local models provides a rich enough foundation to find an optimal proxy set.
When compared against existing aggregation methods such as GLocalX, submodular pick, and integer-programming-based spherical coverage, the balanced ExplainReduce algorithm proved significantly faster and frequently outperformed or matched alternative frameworks across fidelity, coverage, and instability metrics.
Practical Insights in Physics and Synthetic Tests
To validate the framework on real-world physics data, the investigators applied ExplainReduce to simulated Large Hadron Collider proton-proton collisions. A random forest classifier was trained to distinguish between quark-generated and gluon-generated particle jets. Using SLISEMAP to generate five hundred explanations, the balanced reduction algorithm successfully distilled them into four proxy models.
The resulting coefficients aligned perfectly with established quantum chromodynamic theory. For instance, wider particle jets with higher girth metrics are physically known to behave more like gluons, and the optimized proxy model correctly assigned high positive coefficients to these specific physical parameters. Similarly, synthetic benchmarking tests confirmed that data points mapped reliably back to their ground-truth clusters.
While the current implementation focuses primarily on global loss and coverage optimization without explicitly mapping the spatial distribution of data manifolds, the authors note that future extensions could integrate manifold learning for enhanced exploratory data analysis and uncertainty quantification.
The research was published in Explainable Artificial Intelligence.
This article has been fact checked for accuracy, with information verified against reputable sources. Learn more about us and our editorial process.
Last reviewed on .
Article history
- Latest version
Reference(s)
- Seppäläinen, Lauri., et al. “ExplainReduce: Generating Global Explanations from Many Local Explanations.” Communications in Computer and Information Science, November 17, 2026, pp. 299-319. Springer Nature Switzerland, doi: 10.1007/978-3-032-45473-7_15. <https://doi.org/10.1007/978-3-032-45473-7_15>.
Cite this page:
- Posted by Bilal Abbasi