Getting out of the Big-Muddy: Escalation of Commitment in LLMs

Emilio Barkett, Olivia Long, Paul Kröger · 2025

Read Getting out of the Big-Muddy: Escalation of Commitment in LLMs by Emilio Barkett, Olivia Long, Paul Kröger (2025) 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-2508-01545.txt.

Description

Emilio Barkett, Olivia Long, and Paul Kröger test whether language models fall for the sunk cost trap, in a 2025 paper. People often keep funding failing projects because they have already spent much. Models trained on human text might inherit that bias. The team ran a two-stage investment task. Over six thousand trials covered four setups. The model acted as investor in one and advisor in another. Other setups used groups of agents or layered pressures.

Acting alone, the models behaved rationally. Escalation stayed minimal in the solo studies. Group settings changed the picture sharply. Equal peer groups escalated in nearly every trial. Unequal groups with a hierarchy escalated about half the time. Compound personal and organizational pressure also drove high escalation. The authors conclude the bias is context-driven, not built in. That matters for deploying agent teams without supervision. Designers should watch group dynamics closely.

Catalog record

CREATOREmilio Barkett, Olivia Long, Paul Kröger
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE38 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-01545

Browse the full library · About the Tower of Babel