Saved in:
Bibliographic Details
Main Authors: de Bos, Niels uit, Garriga-Alonso, Adrià
Format: Preprint
Published: 2024
Subjects:
Online Access:https://arxiv.org/abs/2407.15166
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866917729758347264
author de Bos, Niels uit
Garriga-Alonso, Adrià
author_facet de Bos, Niels uit
Garriga-Alonso, Adrià
contents Circuits are supposed to accurately describe how a neural network performs a specific task, but do they really? We evaluate three circuits found in the literature (IOI, greater-than, and docstring) in an adversarial manner, considering inputs where the circuit's behavior maximally diverges from the full model. Concretely, we measure the KL divergence between the full model's output and the circuit's output, calculated through resample ablation, and we analyze the worst-performing inputs. Our results show that the circuits for the IOI and docstring tasks fail to behave similarly to the full model even on completely benign inputs from the original task, indicating that more robust circuits are needed for safety-critical applications.
format Preprint
id arxiv_https___arxiv_org_abs_2407_15166
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Adversarial Circuit Evaluation
de Bos, Niels uit
Garriga-Alonso, Adrià
Machine Learning
Circuits are supposed to accurately describe how a neural network performs a specific task, but do they really? We evaluate three circuits found in the literature (IOI, greater-than, and docstring) in an adversarial manner, considering inputs where the circuit's behavior maximally diverges from the full model. Concretely, we measure the KL divergence between the full model's output and the circuit's output, calculated through resample ablation, and we analyze the worst-performing inputs. Our results show that the circuits for the IOI and docstring tasks fail to behave similarly to the full model even on completely benign inputs from the original task, indicating that more robust circuits are needed for safety-critical applications.
title Adversarial Circuit Evaluation
topic Machine Learning
url https://arxiv.org/abs/2407.15166