Add new paper:

[wyzh0912]

Title

On Mechanistic Circuits for Extractive Question-Answering

Published Date

2025-02-12

Source

arXiv

Head Name

Attribution Head

Summary

• Innovation: The paper introduces a method to extract mechanistic circuits from language models for extractive QA tasks using causal mediation analysis. It demonstrates how these circuits can elucidate the interplay between parametric memory and context usage, leading to practical applications like data attribution and model steering.

• Tasks: The study involves designing a probe dataset and applying causal mediation analysis to identify circuits responsible for context and memory faithfulness in language models. The insights from these circuits are then used to develop ATTNATTRIB, a fast data attribution algorithm, and to steer models towards improved context faithfulness in QA tasks.

• Significant Result: The research finds that a small set of attention heads within the identified circuits performs reliable data attribution by default, allowing for state-of-the-art attribution results in extractive QA benchmarks. Additionally, using these attribution insights, the model can be steered to improve its reliance on context over parametric memory, enhancing context faithfulness by up to 9% in QA datasets.