Pretrained Embeddings as a Behavior Specification Mechanism

Parv Kapoor, Abigail Hammer, Ashish Kapoor, Karen Leung, Eunsuk Kang · 2025

Read Pretrained Embeddings as a Behavior Specification Mechanism by Parv Kapoor, Abigail Hammer, Ashish Kapoor, Karen Leung, Eunsuk Kang (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-2503-02012.txt.

Description

Parv Kapoor and co-authors wrote this computing paper in 2025 on stating what an AI-driven machine must do. Robots that see the world through a learned model are hard to pin with classic rules, since their input is messy sense data. The team's key move is to treat embeddings, the math form a model gives a concept, as first-class parts of a specification language. Rules then speak of gaps between the ideal embedding and the one observed.

They build a new logic for this, Embedding Temporal Logic, which can state a wider range of traits over time than past tools. A test with foundation-model robots on planning tasks showed the specs could steer the machines toward good conduct. The authors offer the logic as a bridge between formal proof and learned perception.

Catalog record

CREATORParv Kapoor, Abigail Hammer, Ashish Kapoor, Karen Leung, Eunsuk Kang
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE41 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-02012

Browse the full library · About the Tower of Babel