Identifying Interactions at Scale for LLMs

2026-03-13 · Source: The Berkeley Artificial Intelligence Research Blog · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Explainable AI · Depth: Advanced, medium

Summary

The SPEX and ProxySPEX frameworks are novel algorithms designed to identify influential interactions within complex machine learning systems, including Large Language Models (LLMs), at scales previously unachievable. These methods address the challenge of "complexity at scale" in interpretability research by efficiently capturing how features, training data points, and internal model components interact to drive predictions. SPEX, or Spectral Explainer, leverages signal processing and coding theory, based on observations of sparsity and low-degreeness in influential interactions, to perform sparse recovery. ProxySPEX further enhances efficiency by exploiting the hierarchical nature of interactions, reducing the number of required ablations by approximately 10x while matching SPEX's performance. The frameworks apply to feature attribution, data attribution, and mechanistic interpretability, enabling analysis of long-context inputs, identification of data synergies and redundancies, and understanding of internal model component dependencies.

Key takeaway

For research scientists developing or deploying complex ML models, understanding the underlying decision-making processes is crucial for trustworthiness and safety. You should consider integrating SPEX or ProxySPEX into your interpretability toolkit, especially when dealing with high-dimensional data or large models. These frameworks offer a scalable way to uncover critical feature, data, and model component interactions, enabling more targeted debugging, architectural interventions, and data curation strategies.

Key insights

SPEX and ProxySPEX efficiently identify influential interactions in ML models by exploiting sparsity and hierarchy.

Principles

• Model behavior emerges from complex dependencies.
• Influential interactions are sparse and low-degree.
• Higher-order interactions often imply lower-order subsets.

Method

Ablation measures influence by observing changes when components are removed. SPEX uses signal processing for sparse recovery, while ProxySPEX adds hierarchical insights for 10x fewer ablations.

In practice

• Identify symptoms driving LLM medical diagnoses.
• Prune attention heads for task-specific performance.
• Select training data by preserving synergies.

Topics

• Machine Learning Interpretability
• Large Language Models
• Feature Attribution
• Data Attribution
• Mechanistic Interpretability

Code references

• mmschlk/shapiq

Best for: Research Scientist, AI Researcher, AI Scientist, Machine Learning Engineer

Related on AIssential

• Identifying interactions at scale for LLMs — SPEX and ProxySPEX efficiently identify influential interactions in LLMs by exploiting sparsity and hierarchy.
• The Representational Limit of Scalar Interactions: An Interventional Decomposition — Scalar interaction scores inherently conflate uniqueness, redundancy, and synergy, necessitating advanced decomposition for accurate feature attribution.
• SAILS: Surrogate-based Analysis of Interactions via Local Effect Smooths — SAILS characterizes pairwise feature interactions in black-box models by fitting GAM surrogates to local effects, enabling detection, categorization, and visualization of functional forms.
• Proxy-Based Approximation of Shapley and Banzhaf Interactions — ProxySHAP improves Shapley and Banzhaf interaction estimation by combining tree-based proxies with residual correction for speed and accuracy.
• From Insight to Action: A Novel Framework for Interpretability-Guided Data Selection in Large Language Models — Leveraging LLM internal causal mechanisms to guide data selection significantly boosts fine-tuning efficiency and performance.
• A Gentle Primer on LLM Explainability — LLM explainability requires dynamic evaluation and model-agnostic methods to understand why models generate specific outputs, moving beyond simple correctness.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by The Berkeley Artificial Intelligence Research Blog.