Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Evaluating Proactive Risk Awareness of Large Language Models

Xuan Luo, Yubin Chen, Zhiyu Hou, Linpu Yu, Geng Tu, Jing Li, Ruifeng Xu

TL;DR

A proactive risk awareness evaluation framework that measures whether LLMs can anticipate potential harms and provide warnings before damage occurs is introduced, and the Butterfly dataset is constructed to instantiate this framework in the environmental and ecological domain.

Abstract

As large language models (LLMs) are increasingly embedded in everyday decision-making, their safety responsibilities extend beyond reacting to explicit harmful intent toward anticipating unintended but consequential risks. In this work, we introduce a proactive risk awareness evaluation framework that measures whether LLMs can anticipate potential harms and provide warnings before damage occurs. We construct the Butterfly dataset to instantiate this framework in the environmental and ecological domain. It contains 1,094 queries that simulate ordinary solution-seeking activities whose responses may induce latent ecological impact. Through experiments across five widely used LLMs, we analyze the effects of response length, languages, and modality. Experimental results reveal consistent, significant declines in proactive awareness under length-restricted responses, cross-lingual similarities, and persistent blind spots in (multimodal) species protection. These findings highlight a critical gap between current safety alignment and the requirements of real-world ecological responsibility, underscoring the need for proactive safeguards in LLM deployment.

Evaluating Proactive Risk Awareness of Large Language Models

TL;DR

A proactive risk awareness evaluation framework that measures whether LLMs can anticipate potential harms and provide warnings before damage occurs is introduced, and the Butterfly dataset is constructed to instantiate this framework in the environmental and ecological domain.

Abstract

As large language models (LLMs) are increasingly embedded in everyday decision-making, their safety responsibilities extend beyond reacting to explicit harmful intent toward anticipating unintended but consequential risks. In this work, we introduce a proactive risk awareness evaluation framework that measures whether LLMs can anticipate potential harms and provide warnings before damage occurs. We construct the Butterfly dataset to instantiate this framework in the environmental and ecological domain. It contains 1,094 queries that simulate ordinary solution-seeking activities whose responses may induce latent ecological impact. Through experiments across five widely used LLMs, we analyze the effects of response length, languages, and modality. Experimental results reveal consistent, significant declines in proactive awareness under length-restricted responses, cross-lingual similarities, and persistent blind spots in (multimodal) species protection. These findings highlight a critical gap between current safety alignment and the requirements of real-world ecological responsibility, underscoring the need for proactive safeguards in LLM deployment.
Paper Structure (39 sections, 4 equations, 12 figures, 14 tables)

This paper contains 39 sections, 4 equations, 12 figures, 14 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Safety and Proactive Risk Evaluation
  4. Environmental Knowledge Evaluation
  5. Biodiversity and Species Evaluation
  6. Dataset Construction
  7. Legal Grounding of Environmental Harm
  8. Scenario and Query Generation
  9. Quality Verification
  10. Protected Species Subset
  11. Proactive Environmental Intelligence
  12. Definition
  13. Query Space.
  14. Response Categories.
  15. Metrics
  16. ...and 24 more sections

Figures (12)

  • Figure 1: Safety evaluation paradigms.
  • Figure 2: Illustrations of our data construction process for systematic evaluation of proactive environmental intelligence. A neutral User Query is generated from a realistic real-world scenario that does not explicitly mention environmental harm. Each query is mapped to a legally grounded harmful Behavior, which has potential Environmental Impact. Models' Response to User Query are evaluated along two related dimensions: whether the responses adopt the harmful behavior (Behavior Adoption) and whether they issue environmental reminders aligned with the ecological impact (Aligned Reminders).
  • Figure 3: Data examples of protected species. Two evaluation modes: Image + Text Query and Text-only Query.
  • Figure 4: Illustration of metrics ProR($\uparrow$), GR($\downarrow$), HAR($\downarrow$), and BR($\downarrow$). Statistics in Appendix \ref{['sec:appendix:main-results']} Table \ref{['tab:main-results-en']} and \ref{['tab:main-results-zh']}.
  • Figure 5: The proportion within the proactive risk awareness {SafeAlt $\cup$ WarnIntel} and within the harmful behavior adoption {WarnIntel $\cup$ WarnGeneral $\cup$ Blind}. Statistics in Appendix \ref{['sec:appendix:main-results']} Table \ref{['tab:main-results-en']} and \ref{['tab:main-results-zh']}.
  • ...and 7 more figures