Beyond a Global Norm: Personalizing Toxicity Sensitivity in Language Models Without Retraining

2026-07-25

AI Summary

This work addresses the challenge of aligning language models with users’ subjective sensitivities to harmful content without relying on global alignment standards. It systematically evaluates training-free, inference-stage interventions across three phases—pre-decoding, during decoding, and post-decoding—to enable personalized control over toxicity sensitivity. The study presents the first comprehensive comparison of diverse training-free techniques, including prompt modulation, token/logit/representation manipulation, and candidate re-ranking, in the context of personalized toxicity alignment. Through this analysis, it reveals inherent trade-offs among alignment accuracy, degree of personalization, and linguistic quality. Experimental results on the PRISM dataset demonstrate that the proposed approaches reduce alignment error by 28%–47%, establishing the feasibility and effectiveness of achieving personalized alignment without model retraining.

Abstract

Reducing toxicity is often framed as a global alignment problem, yet perceptions of harmful language are subjective and context-dependent. We present the first comparative evaluation of training-free methods for aligning language generation to user-specific toxicity sensitivities across three inference-time intervention stages: pre-decoding (prompt conditioning and rewriting), in-decoding (token, logit, and representation steering), and post-decoding (candidate re-ranking). Evaluated against toxicity sensitivity targets derived from the PRISM dataset, all methods reduce alignment error by 28-47%. However, the results reveal a fundamental trade-off between alignment effectiveness, personalization, and general language quality, showing how toxicity sensitivity alignment is an inherently multi-objective problem.

Problem

• toxicity sensitivity
• personalization
• language models
• alignment
• subjective harm

Innovation

• training-free alignment
• toxicity sensitivity personalization
• inference-time intervention
• multi-objective alignment
• language model safety