Beyond Reactive Safety: Risk-Aware LLM Alignment via Long-Horizon Simulation

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Sun, Chenkai, Zhang, Denghui, Zhai, ChengXiang, Ji, Heng
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866909660281307136
author Sun, Chenkai
Zhang, Denghui
Zhai, ChengXiang
Ji, Heng
author_facet Sun, Chenkai
Zhang, Denghui
Zhai, ChengXiang
Ji, Heng
contents Given the growing influence of language model-based agents on high-stakes societal decisions, from public policy to healthcare, ensuring their beneficial impact requires understanding the far-reaching implications of their suggestions. We propose a proof-of-concept framework that projects how model-generated advice could propagate through societal systems on a macroscopic scale over time, enabling more robust alignment. To assess the long-term safety awareness of language models, we also introduce a dataset of 100 indirect harm scenarios, testing models' ability to foresee adverse, non-obvious outcomes from seemingly harmless user prompts. Our approach achieves not only over 20% improvement on the new dataset but also an average win rate exceeding 70% against strong baselines on existing safety benchmarks (AdvBench, SafeRLHF, WildGuardMix), suggesting a promising direction for safer agents.
format Preprint
id arxiv_https___arxiv_org_abs_2506_20949
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Beyond Reactive Safety: Risk-Aware LLM Alignment via Long-Horizon Simulation
Sun, Chenkai
Zhang, Denghui
Zhai, ChengXiang
Ji, Heng
Artificial Intelligence
Computation and Language
Given the growing influence of language model-based agents on high-stakes societal decisions, from public policy to healthcare, ensuring their beneficial impact requires understanding the far-reaching implications of their suggestions. We propose a proof-of-concept framework that projects how model-generated advice could propagate through societal systems on a macroscopic scale over time, enabling more robust alignment. To assess the long-term safety awareness of language models, we also introduce a dataset of 100 indirect harm scenarios, testing models' ability to foresee adverse, non-obvious outcomes from seemingly harmless user prompts. Our approach achieves not only over 20% improvement on the new dataset but also an average win rate exceeding 70% against strong baselines on existing safety benchmarks (AdvBench, SafeRLHF, WildGuardMix), suggesting a promising direction for safer agents.
title Beyond Reactive Safety: Risk-Aware LLM Alignment via Long-Horizon Simulation
topic Artificial Intelligence
Computation and Language
url https://arxiv.org/abs/2506.20949