ABLE framework maps 239 LLMs with attribution-based embeddings

The explosive growth of large language models has created a heterogeneous and poorly documented ecosystem, making systematic model comparison critical for provenance auditing, security analysis, and model selection. Existing representation methods struggle: internal parameter analysis is powerful but fails under structural heterogeneity, while external output-based methods conflate models with similar behaviors and are hard to align across different tokenizers. To bridge this gap, researchers from multiple institutions introduce ABLE (Attribution-Based Large-model Embedding), a framework that leverages the interpretability space to construct model representations. By aggregating gradient-based feature attributions via a tokenizer-agnostic word-level alignment, ABLE captures each model’s unique input-sensitivity patterns rather than just surface-level outputs. Beyond empirical utility, the team provides a stability analysis showing that under standard regularity assumptions for differentiable Transformer-style models, ABLE induces a Lipschitz-continuous parameter-to-embedding map with finite-sample convergence guarantees.

Extensive experiments on 239 open-source LLMs demonstrate that this training-free approach achieves competitive or superior performance in three key tasks: relation prediction (detecting model families or architecture provenance), model routing (matching queries to the best-performing model), and benchmark score prediction (estimating performance on downstream tasks without running full evaluations). Because ABLE requires no additional training or parameter access—just input-output gradient signals—it scales to massive, heterogeneous model collections. The tokenizer-agnostic alignment allows direct comparison of models from different families (e.g., LLaMA, Qwen, Mistral) without modification. This work opens the door for practical tools that help practitioners audit, select, and monitor LLMs in rapidly evolving ecosystems, addressing a growing need for transparency and comparability in AI.