The Intervention Paradox: Accurate Failure Prediction in Agents Does Not Imply Effective Failure Prevention

Rakshith Vasudev, Melisa Russak, Dan Bikel, Waseem Alshikh · 2026

Read The Intervention Paradox: Accurate Failure Prediction in Agents Does Not Imply Effective Failure Prevention by Rakshith Vasudev, Melisa Russak, Dan Bikel, Waseem Alshikh (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-03338.txt.

Description

A critic that predicts agent failure accurately can still make deployment worse, warns a study by Rakshith Vasudev with 3 colleagues. Proactive interventions by language model critics are assumed to improve reliability, yet deployment effects are poorly understood. Their binary critic had strong offline accuracy. At deployment, though, intervention caused a collapse of twenty-six percentage points on one model while barely affecting another, a variability accuracy alone cannot predict for a new system.

The authors identify a disruption-recovery trade-off. Interventions can rescue failing trajectories but also disrupt ones that would have succeeded. Their pre-deployment test uses a small pilot of tasks to estimate whether intervention will help or harm. Across benchmarks the test anticipated outcomes, degrading on high-success tasks and improving modestly on a high-failure benchmark. Its main value is identifying when not to intervene.

Catalog record

CREATORRakshith Vasudev, Melisa Russak, Dan Bikel, Waseem Alshikh
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE53 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-03338

Browse the full library · About the Tower of Babel