Work

The Frame Ladder: How the phrasing of a request changes reward-hacking rates

Year: 2026

Type: preprint

Abstract: Working note reading Figures 12 and 13 of Qi et al., Training a Misaligned Reward Seeker (Anthropic Alignment Science, August 2026), with secondary reference to OpenAI's Hugging Face incident report. All figures quoted are the original authors'; the readings and the labels are mine. Reports the effect of four prompt framings on reward-hacking rates in an impossible-task evaluation, and two open measurement questions arising from it.

Source: Zenodo (CERN European Organization for Nuclear Research)

DOI: 10.5281/zenodo.22236246

Author: Feshter, Yanush

Language: English

Topic: Ethics and Social Impacts of AI

Subfield: Safety Research

Field: Social Sciences

Domain: Social Sciences

Open Access status: green

OpenAlex ID: W7204869627

Cites:

Cited by:

Related to: