Skip to content
conftrace_
2026 EMNLP EMNLP 2026

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

AbstractReducing toxicity is often framed as a global alignment problem, yet perceptions of harmful language are subjective and context-dependent. We present a 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 classifier-based toxicity targets derived from the PRISM dataset, all methods reduce alignment error by 28-47%. However, the results show trade-offs between alignment effectiveness, personalization, and general language quality.

More from EMNLP 2026

All EMNLP 2026 papers →