Derailing Non-Answers via Logit Suppression at Output Subspace Boundaries in RLHF-Aligned Language Models

Documenti analoghi

• OrthAlign: Orthogonal Subspace Decomposition for Non-Interfering Multi-Objective Alignment
  di: Lin, Liang, et al.
  Pubblicazione: (2025)
• Knowledge-Aware Conversation Derailment Forecasting Using Graph Convolutional Networks
  di: Altarawneh, Enas, et al.
  Pubblicazione: (2024)
• Multi-Stream LLMs: Unblocking Language Models with Parallel Streams of Thoughts, Inputs and Outputs
  di: Su, Guinan, et al.
  Pubblicazione: (2026)
• Promote, Suppress, Iterate: How Language Models Answer One-to-Many Factual Queries
  di: Yan, Tianyi Lorena, et al.
  Pubblicazione: (2025)
• Resolving Conflicts in Lifelong Learning via Aligning Updates in Subspaces
  di: Zhou, Yueer, et al.
  Pubblicazione: (2025)