Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Dynamic Evaluation for Oversensitivity in LLMs

Sophia Xiao Pu, Sitao Cheng, Xin Eric Wang, William Yang Wang

TL;DR

Oversensitivity hampers user interactions by causing harmless prompts to be reflexively refused. The paper formalizes oversensitivity and introduces a dynamic evaluation framework that combines a proxy detector, frequency-adjusted feature attribution, and adversarial prompt generation to produce model-specific, benign prompts that trigger defensive responses. This approach yields OverBench, a large-scale 450{,}000-prompt benchmark across 25 models, plus OverBench-Hard for high-conflict samples, enabling continual monitoring of defensive triggers as models evolve. The work demonstrates cross-family patterns in defensiveness, showcases substantial variation with model families and sizes, and provides a practically impactful resource for evaluating and mitigating oversensitivity over time.

Abstract

Oversensitivity occurs when language models defensively reject prompts that are actually benign. This behavior not only disrupts user interactions but also obscures the boundary between harmful and harmless content. Existing benchmarks rely on static datasets that degrade overtime as models evolve, leading to data contamination and diminished evaluative power. To address this, we develop a framework that dynamically generates model-specific challenging datasets, capturing emerging defensive patterns and aligning with each model's unique behavior. Building on this approach, we construct OVERBENCH, a benchmark that aggregates these datasets across diverse LLM families, encompassing 450,000 samples from 25 models. OVERBENCH provides a dynamic and evolving perspective on oversensitivity, allowing for continuous monitoring of defensive triggers as models advance, highlighting vulnerabilities that static datasets overlook.

Dynamic Evaluation for Oversensitivity in LLMs

TL;DR

Oversensitivity hampers user interactions by causing harmless prompts to be reflexively refused. The paper formalizes oversensitivity and introduces a dynamic evaluation framework that combines a proxy detector, frequency-adjusted feature attribution, and adversarial prompt generation to produce model-specific, benign prompts that trigger defensive responses. This approach yields OverBench, a large-scale 450{,}000-prompt benchmark across 25 models, plus OverBench-Hard for high-conflict samples, enabling continual monitoring of defensive triggers as models evolve. The work demonstrates cross-family patterns in defensiveness, showcases substantial variation with model families and sizes, and provides a practically impactful resource for evaluating and mitigating oversensitivity over time.

Abstract

Oversensitivity occurs when language models defensively reject prompts that are actually benign. This behavior not only disrupts user interactions but also obscures the boundary between harmful and harmless content. Existing benchmarks rely on static datasets that degrade overtime as models evolve, leading to data contamination and diminished evaluative power. To address this, we develop a framework that dynamically generates model-specific challenging datasets, capturing emerging defensive patterns and aligning with each model's unique behavior. Building on this approach, we construct OVERBENCH, a benchmark that aggregates these datasets across diverse LLM families, encompassing 450,000 samples from 25 models. OVERBENCH provides a dynamic and evolving perspective on oversensitivity, allowing for continuous monitoring of defensive triggers as models advance, highlighting vulnerabilities that static datasets overlook.
Paper Structure (18 sections, 6 equations, 4 figures, 1 algorithm)

This paper contains 18 sections, 6 equations, 4 figures, 1 algorithm.

Table of Contents

  1. Introduction
  2. Problem Formalization
  3. Methodology
  4. Proxy Modeling
  5. Feature Attribution
  6. Adversarial Generation
  7. OverBench
  8. Dataset Construction
  9. Prompt Categorization
  10. Benchmark Aggregation and Evaluation
  11. Feature Attribution Analysis
  12. Related Work
  13. Static Benchmarks.
  14. Diagnostic Analyses.
  15. Conclusion
  16. ...and 3 more sections

Figures (4)

  • Figure 1: An overview of our dynamic OverBench construction framework.
  • Figure 2: Oversensitivity rates of various LLMs evaluated on OverBench and OverBench-Hard.
  • Figure 3: Percentile-based heatmap of the top 20 global features across models. Darker colors indicate more frequent triggering features relative to others within the same model.
  • Figure 4: Prompt Distribution