AttriBoT: A Bag of Tricks for Efficiently Approximating Leave-One-Out Context Attribution

2024-11-22

arXiv.org

AI Summary

Standard leave-one-out (LOO) attribution for large language models (LLMs) incurs prohibitive computational overhead, hindering its practical use in pedagogy and interpretability analysis. Method: We propose Cache-LOO, an efficient LOO approximation framework leveraging activation caching and reuse, hierarchical context attribution, and lightweight surrogate modeling via knowledge distillation. It introduces a novel activation cache reuse mechanism and constructs high-fidelity surrogate models by semantically partitioning context and distilling LLM knowledge. Contribution/Results: Cache-LOO accelerates LOO computation by over 300× compared to standard LOO, reducing per-attribution latency to just 1/30 of typical text generation time. It achieves significantly higher attribution fidelity than existing baselines while maintaining scalability and usability. An open-source implementation enables large-scale, low-barrier deployment of LLM interpretability methods.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Computer Vision: Large Vision Models

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

The influence of contextual input on the behavior of large language models (LLMs) has prompted the development of context attribution methods that aim to quantify each context span's effect on an LLM's generations. The leave-one-out (LOO) error, which measures the change in the likelihood of the LLM's response when a given span of the context is removed, provides a principled way to perform context attribution, but can be prohibitively expensive to compute for large models. In this work, we introduce AttriBoT, a series of novel techniques for efficiently computing an approximation of the LOO error for context attribution. Specifically, AttriBoT uses cached activations to avoid redundant operations, performs hierarchical attribution to reduce computation, and emulates the behavior of large target models with smaller proxy models. Taken together, AttriBoT can provide a>300x speedup while remaining more faithful to a target model's LOO error than prior context attribution methods. This stark increase in performance makes computing context attributions for a given response 30x faster than generating the response itself, empowering real-world applications that require computing attributions at scale. We release a user-friendly and efficient implementation of AttriBoT to enable efficient LLM interpretability as well as encourage future development of efficient context attribution methods.

Problem

• Large Language Models
• Contextual Influence Estimation
• Educational Simplification

Innovation

• Efficient Context Attribution
• Hierarchical Attribution Method
• Approximation with Small Models