ResearchTrend.AI
  • Papers
  • Communities
  • Events
  • Blog
  • Pricing
Papers
Communities
Social Events
Terms and Conditions
Pricing
Parameter LabParameter LabTwitterGitHubLinkedInBlueskyYoutube

© 2025 ResearchTrend.AI, All rights reserved.

  1. Home
  2. Papers
  3. 2501.15758
32
1

Risk-Aware Distributional Intervention Policies for Language Models

28 January 2025
Bao Nguyen
Binh Nguyen
Duy Nguyen
V. Nguyen
ArXivPDFHTML
Abstract

Language models are prone to occasionally undesirable generations, such as harmful or toxic content, despite their impressive capability to produce texts that appear accurate and coherent. This paper presents a new two-stage approach to detect and mitigate undesirable content generations by rectifying activations. First, we train an ensemble of layerwise classifiers to detect undesirable content using activations by minimizing a smooth surrogate of the risk-aware score. Then, for contents that are detected as undesirable, we propose layerwise distributional intervention policies that perturb the attention heads minimally while guaranteeing probabilistically the effectiveness of the intervention. Benchmarks on several language models and datasets show that our method outperforms baselines in reducing the generation of undesirable output.

View on arXiv
Comments on this paper