Saved in:
Bibliographic Details
Main Authors: Brown, Davis, Balehannina, Prithvi, Jin, Helen, Havaldar, Shreya, Hassani, Hamed, Wong, Eric
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2503.01986
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866911176320876544
author Brown, Davis
Balehannina, Prithvi
Jin, Helen
Havaldar, Shreya
Hassani, Hamed
Wong, Eric
author_facet Brown, Davis
Balehannina, Prithvi
Jin, Helen
Havaldar, Shreya
Hassani, Hamed
Wong, Eric
contents Language model evaluations often fail to characterize consequential failure modes, forcing experts to inspect outputs and build new benchmarks. We introduce task elicitation, a method that automatically builds new evaluations to profile model behavior. Task elicitation finds hundreds of natural-language tasks -- an order of magnitude more than prior work -- where frontier models exhibit systematic failures, in domains ranging from forecasting to online harassment. For example, we find that Sonnet 3.5 over-associates quantum computing and AGI and that o3-mini is prone to hallucination when fabrications are repeated in-context.
format Preprint
id arxiv_https___arxiv_org_abs_2503_01986
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Adaptively profiling models with task elicitation
Brown, Davis
Balehannina, Prithvi
Jin, Helen
Havaldar, Shreya
Hassani, Hamed
Wong, Eric
Computation and Language
Artificial Intelligence
Machine Learning
Language model evaluations often fail to characterize consequential failure modes, forcing experts to inspect outputs and build new benchmarks. We introduce task elicitation, a method that automatically builds new evaluations to profile model behavior. Task elicitation finds hundreds of natural-language tasks -- an order of magnitude more than prior work -- where frontier models exhibit systematic failures, in domains ranging from forecasting to online harassment. For example, we find that Sonnet 3.5 over-associates quantum computing and AGI and that o3-mini is prone to hallucination when fabrications are repeated in-context.
title Adaptively profiling models with task elicitation
topic Computation and Language
Artificial Intelligence
Machine Learning
url https://arxiv.org/abs/2503.01986