Summary

LXT (Layer-wise Relevance Propagation for Transformers) accelerates Explainable AI (XAI) research for Large Language Models (LLMs) and Vision Transformers (ViTs). It provides fast, faithful input attributions by revealing how each token and neuron contributes to model predictions, overcoming the noise and unreliability of traditional gradient-based methods. This enables deeper understanding and debugging of complex transformer architectures for researchers and power users.

How It Works

LXT employs AttnLRP, an attention-aware backpropagation technique that corrects gradient flow through non-linearities, ensuring accurate relevance scores. This method requires only a single backward pass, offering significant efficiency gains. It outperforms existing gradient, decomposition, and perturbation-based XAI methods, scaling computationally with $O(1)$ complexity and $O(\sqrt{N})$ memory relative to model depth. Beyond input attributions, LXT facilitates latent feature attribution and visualization by analyzing individual neuron importance, allowing for deeper model introspection.

Quick Start & Requirements

• Installation: pip install lxt
• Dependencies: Tested with transformers==5.9.0, torch==2.12.0, python==3.11.
• Documentation: Official Docs
• Examples: Available in the examples/* directory.

Highlighted Details

• Faithful attributions for black-box transformers via AttnLRP.
• Efficient implementation leveraging Input*Gradient and monkey patching.
• Latent feature attribution and neuron visualization capabilities.
• Broad model support including LLaMA 2/3, Gemma 3, Qwen 2, BERT, GPT-2, and Vision Transformers.

Maintenance & Community

Contributions are encouraged. The project is associated with ICML 2024. No specific community channels (e.g., Discord, Slack) or roadmap links are provided in the README.

Licensing & Compatibility

• License: BSD-3 Clause License.
• Restrictions: The README states LRP is a patented technology, usable free of charge only for personal and scientific purposes. This implies significant restrictions for commercial applications.

Limitations & Caveats

• Attribution for Qwen 3 models may be skewed towards the first token.
• GPT-2 models are best used with contrastive explanations.
• The BSD-3 Clause license includes a patent restriction limiting free use to personal and scientific contexts, posing a barrier for commercial adoption.