Inverse Scaling in Test-Time Compute

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Gema, Aryo Pradipta, Hägele, Alexander, Chen, Runjin, Arditi, Andy, Goldman-Wetzler, Jacob, Fraser-Taliente, Kit, Sleight, Henry, Petrini, Linda, Michael, Julian, Alex, Beatrice, Minervini, Pasquale, Chen, Yanda, Benton, Joe, Perez, Ethan
Format: Preprint
Veröffentlicht: 2025
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866911321295945728
author Gema, Aryo Pradipta
Hägele, Alexander
Chen, Runjin
Arditi, Andy
Goldman-Wetzler, Jacob
Fraser-Taliente, Kit
Sleight, Henry
Petrini, Linda
Michael, Julian
Alex, Beatrice
Minervini, Pasquale
Chen, Yanda
Benton, Joe
Perez, Ethan
author_facet Gema, Aryo Pradipta
Hägele, Alexander
Chen, Runjin
Arditi, Andy
Goldman-Wetzler, Jacob
Fraser-Taliente, Kit
Sleight, Henry
Petrini, Linda
Michael, Julian
Alex, Beatrice
Minervini, Pasquale
Chen, Yanda
Benton, Joe
Perez, Ethan
contents We construct evaluation tasks where extending the reasoning length of Large Reasoning Models (LRMs) deteriorates performance, exhibiting an inverse scaling relationship between test-time compute and accuracy. Our evaluation tasks span four categories: simple counting tasks with distractors, regression tasks with spurious features, deduction tasks with constraint tracking, and advanced AI risks. We identify five distinct failure modes when models reason for longer: 1) Claude models become increasingly distracted by irrelevant information; 2) OpenAI o-series models resist distractors but overfit to problem framings; 3) models shift from reasonable priors to spurious correlations; 4) all models show difficulties in maintaining focus on complex deductive tasks; and 5) extended reasoning may amplify concerning behaviors, with Claude Sonnet 4 showing increased expressions of self-preservation. These findings suggest that while test-time compute scaling remains promising for improving model capabilities, it may inadvertently reinforce problematic reasoning patterns. Our results demonstrate the importance of evaluating models across diverse reasoning lengths to identify and address these failure modes in LRMs.
format Preprint
id arxiv_https___arxiv_org_abs_2507_14417
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Inverse Scaling in Test-Time Compute
Gema, Aryo Pradipta
Hägele, Alexander
Chen, Runjin
Arditi, Andy
Goldman-Wetzler, Jacob
Fraser-Taliente, Kit
Sleight, Henry
Petrini, Linda
Michael, Julian
Alex, Beatrice
Minervini, Pasquale
Chen, Yanda
Benton, Joe
Perez, Ethan
Artificial Intelligence
Computation and Language
We construct evaluation tasks where extending the reasoning length of Large Reasoning Models (LRMs) deteriorates performance, exhibiting an inverse scaling relationship between test-time compute and accuracy. Our evaluation tasks span four categories: simple counting tasks with distractors, regression tasks with spurious features, deduction tasks with constraint tracking, and advanced AI risks. We identify five distinct failure modes when models reason for longer: 1) Claude models become increasingly distracted by irrelevant information; 2) OpenAI o-series models resist distractors but overfit to problem framings; 3) models shift from reasonable priors to spurious correlations; 4) all models show difficulties in maintaining focus on complex deductive tasks; and 5) extended reasoning may amplify concerning behaviors, with Claude Sonnet 4 showing increased expressions of self-preservation. These findings suggest that while test-time compute scaling remains promising for improving model capabilities, it may inadvertently reinforce problematic reasoning patterns. Our results demonstrate the importance of evaluating models across diverse reasoning lengths to identify and address these failure modes in LRMs.
title Inverse Scaling in Test-Time Compute
topic Artificial Intelligence
Computation and Language
url https://arxiv.org/abs/2507.14417