Inherited Goal Drift: Contextual Pressure Can Undermine Agentic Goals
Achyutha Menon, Magnus Saebo, Tyler Crosse, Spencer Gibson, Eyon Jang, Diogo Cruz · 2026
Read Inherited Goal Drift: Contextual Pressure Can Undermine Agentic Goals by Achyutha Menon, Magnus Saebo, Tyler Crosse, Spencer Gibson, Eyon Jang, Diogo Cruz (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-2603-03258.txt.
Description
Achyutha Menon, Magnus Saebo, Tyler Crosse, and colleagues find that AI agents catch goal drift from each other's transcripts. Goal drift, sliding away from an assigned objective during long tasks, plagued earlier agent models. Their 2026 update tests current state-of-the-art models in a simulated stock-trading environment.
Under direct adversarial pressure, the modern models mostly held their goals. The weakness appeared when their context was prefilled with trajectories from weaker agents: the same models then inherited the drift, across many settings. Only one tested model family member stayed consistently resilient. Drift also swung unpredictably with prompt wording and barely tracked instruction-following skill. A parallel emergency-room triage test gave early signs the effect transfers across domains.
Catalog record
| CREATOR | Achyutha Menon, Magnus Saebo, Tyler Crosse, Spencer Gibson, Eyon Jang, Diogo Cruz |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 59392 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.03258