Harm-Conditioned Computational Friction as a Diagnostic of Alignment Robustness: A Critical Review and Evaluation Framework

Abstract excerpt

<title>Abstract</title> <p>Robust alignment requires that AI models maintain safety-relevant behavior under distribution shift, adversarial prompting, and optimization pressure. Current evaluation methods often rely on surface compliance metrics—such as refusal rates or policy-template adherence—that may fail to detect fragile safety generalization, reward hacking, or prompt-contingent refusal policies. This pape...

Topics

Open a Topic to create a Post that cites this publication.

• Brain
• Brain Mapping
• Cerebral Cortex
• Environment
• Individuality
• Intelligence
• Phenotype
• Socioeconomic Factors

Identifiers and source

Literature Corpus work

0295915d-62b9-5d9a-9666-d33b83ec1f33

DOI

10.21203/rs.3.rs-8327468/v1