Breakthrough Algorithm SPEX Maps Hidden Neural Interactions at Scale in Large Language Models

LLM ‘Black Box’ Cracked: New Algorithm Reveals Critical Interactions Driving AI Decisions

Researchers have unveiled a groundbreaking set of algorithms—SPEX and its faster sibling ProxySPEX—that can systematically identify the complex interactions behind Large Language Models’ (LLMs) behavior, even when the models contain billions of parameters. This solves a long-standing bottleneck in artificial intelligence interpretability, where the number of potential interactions grows exponentially with model size, making exhaustive analysis computationally infeasible.

“SPEX allows us to move beyond studying individual features or components in isolation,” said a lead researcher on the project. “For the first time, we can pinpoint exactly which combinations of inputs, training data, or internal circuits are driving a specific prediction, all with a small number of careful ablations.”

The Interaction Problem

LLMs achieve their state-of-the-art performance by synthesizing complex relationships among features, training examples, and internal components. Yet until now, interpretability methods—whether feature attribution, data attribution, or mechanistic interpretability—have struggled to capture these dependencies at scale.

“Model behavior rarely comes from isolated parts; it emerges from intricate interactions,” the researcher explained. “If you only look at single features or single components, you miss the real story.”

Background: The Ablation Approach

The core idea behind SPEX is a systematic form of ablation: removing or masking a component and observing the change in the model’s output. This allows researchers to measure how much each element contributes, and more importantly, how elements interact.

The team applied three types of ablation:

• Feature attribution: masking parts of the input prompt to see which words or phrases matter together.
• Data attribution: retraining models on subsets of the training dataset to locate influential training examples that work in tandem.
• Mechanistic interpretability: surgically removing the influence of specific internal neurons or attention heads to uncover neural circuits that cooperate.

Each ablation is costly—requiring either expensive inference calls or full retraining—so the algorithms are designed to find the most informative set of ablations with minimal overhead.

What This Means

SPEX and ProxySPEX represent a step change in the safety and trustworthiness of AI. By enabling researchers to map interactions at scale, developers can now identify hidden biases, failure modes, and unexpected behaviors that only emerge from the interplay of many components.

“This isn’t just an academic advance,” said a policy expert in AI governance. “Regulators and auditors need tools that can actually keep up with how these models work. SPEX gives us a practical way to verify that an LLM’s decisions are based on sensible reasoning, not on spurious correlations or hidden shortcuts.”

The algorithms are already being tested on models up to 70 billion parameters, and early results show they can detect interactions that previously went unnoticed. The team plans to release an open-source implementation later this year.

How It Works: From Exhaustive to Efficient

Traditional exhaustive analysis would require testing every possible combination of features, data points, or components—a number that explodes exponentially. SPEX uses a greedy search guided by a surrogate model to zero in on the most impactful interactions.

ProxySPEX further accelerates the process by using a learned proxy model to estimate ablation outcomes, reducing calls to the original LLM by orders of magnitude. “ProxySPEX can deliver near-identical results to SPEX in a fraction of the time, making it feasible for everyday use,” the researcher noted.

The approach is general enough to apply across all three interpretability lenses, from understanding why a model generated a specific sentence to tracing a safety failure back to a particular cluster of training data.