Construct · 1 fields · 1 citations

reward model overoptimization: a standalone construct

One name. One construct. A landmark on the map.

● anchor · a standalone construct Nearest construct by meaning·not indexed

A single standalone construct; nearby constructs are located against it in the Collider. Tap a field for its definition, unit, and measures.

AI safety

optimizing a policy too hard against a learned reward model raises the proxy score while the true preference score eventually falls.

In AI safety, reward model overoptimization refers to optimizing a policy too hard against a learned reward model raises the proxy score while the true preference score eventually falls. This meaning comes from Gao (2022) in arXiv.[1]

Unit
Recovery means
Measured by

Run it through the Collider, or take the citations with you.

See how reward model overoptimization sits against neighbouring constructs in the Collider, or export every reference on this page in one click.

Open in Collider Export BibTeX Report this entry

Nearby constructs

model3 fields · 3 citationsjingle · one word, three constructs policy2 fields · 4 citationsjingle · one word, two constructs preference4 fields · 7 citationsjingle · one word, four constructs

References

Gao, L., Schulman, J., & Hilton, J. (2022). Scaling Laws for Reward Model Overoptimization. arXiv. https://doi.org/10.48550/arxiv.2210.10760

Cited on this page?

Confirm your construct, or correct us.

If your work is cited behind one of these entries, you know better than we do whether we captured the meaning right. Confirm it, suggest an edit, or dispute it, and your name goes on the entry as an author who has checked it, linked to your ORCID.

Claim your construct