Researchers introduce SPEX and ProxySPEX algorithms to identify critical interactions in Large Language Models (LLMs) at scale. These methods overcome the exponential complexity of analyzing feature, data, and component interactions, enabling efficient interpretability via ablation-based attribution across different lenses.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
When features, data points, and model components grow exponentially, how can we efficiently identify the key interactions that truly drive model behavior? The BAIR team's SPEX and ProxySPEX frameworks leverage sparsity, low-order, and hierarchy to transform interaction discovery from exponential search into a tractable sparse recovery problem, achieving comparable results with roughly 10x fewer ablation runs.
The behavior of large language models (LLMs) is typically not determined by isolated components but arises from complex dependencies among features, training data, and internal components. Existing interpretability methods—feature attribution, data attribution, and mechanistic interpretability—all face the same fundamental obstacle: scaling complexity. As the number of features, data points, or components grows, the number of potential interactions grows exponentially, making exhaustive analysis computationally infeasible.
Attribution is fundamentally about ablation: removing input segments, training subsets, or internal components and observing output changes. Each ablation carries a high cost, whether in inference calls or retraining. Therefore, the goal is to compute attributions with the fewest ablations, and SPEX is designed for this purpose.
SPEX (Spectral Explainer) draws on signal processing and coding theory. Its key insight is that while the total number of interactions is vast, the truly influential ones are relatively sparse and typically involve only a few features (low-order). These two properties transform a difficult search problem into a tractable sparse recovery problem.
SPEX strategically selects ablations to combine many candidate interactions into composite signals, then uses efficient decoding algorithms to separate these signals, thereby identifying the specific interactions responsible for model behavior. This method excels in feature attribution, maintaining high faithfulness even as input scales to thousands of features.
Building on SPEX, ProxySPEX identifies another structural property common in complex models: hierarchy—if a high-order interaction is important, its low-order subsets often are too. This observation yields significant computational savings: ProxySPEX achieves performance comparable to SPEX with roughly 10x fewer ablations.
This improvement enables interaction discovery to be applied to larger models and datasets, opening new applications for feature, data, and component attribution.
In sentiment analysis tasks, SPEX matches the faithfulness of existing interaction techniques (Faith-Shap, Faith-Banzhaf) on short inputs, but maintains high performance when contexts expand to thousands of features. In contrast, marginal methods like LIME and Banzhaf, while able to handle this scale, show significantly lower faithfulness because they fail to capture complex interactions.
In a modified trolley problem, GPT-4o mini achieves only 8% accuracy on the altered question. Standard attribution method SHAP identifies a single instance of 'trolley' as the primary factor, but replacing it with a synonym has little effect on predictions. SPEX, however, reveals a high-order synergy between the two 'trolley' instances and 'pulling' and 'lever', a finding that aligns with human intuition. When these four words are replaced with synonyms, the model's error rate drops to near zero.
Data attribution aims to identify the training data points most responsible for a model's predictions. Interaction analysis is equally crucial here: redundant interactions (e.g., semantic duplication) may reinforce erroneous concepts, while synergistic interactions are essential for defining decision boundaries, as individual samples cannot do so alone.
ProxySPEX was applied to a ResNet model trained on CIFAR-10, identifying significant interaction examples for various difficult test points. Synergistic interactions often involve semantically distinct classes jointly defining decision boundaries, such as shared visual features among cars, sports cars, trucks, and red delivery vans; redundant interactions tend to capture visual duplication.
This article's information primarily comes from the BAIR blog, which is a self-description by the research team and lacks independent verification. All performance data (e.g., 10x fewer ablations, 8% accuracy, error rate dropping to near zero) are source claims, not independently confirmed facts.
SPEX and ProxySPEX leverage the sparsity, low-order, and hierarchy of interactions to transform interaction discovery from exponential search into a tractable sparse recovery problem, significantly reducing computational cost and demonstrating superior capability in feature and data attribution compared to traditional methods. However, their practical effectiveness still requires independent validation.
Primary report
Primary source