aiqbench analysis
This research demonstrates a critical vulnerability in LLM alignment, showing that Group Relative Policy Optimization (GRPO) can induce systematic bias with just a single biased training example. The findings highlight how easily post-training guardrails can be overridden and reveal varying susceptibility across different models based on their initial output likelihoods. For the AI safety community, this underscores the fragility of current fine-tuning methods against targeted adversarial inputs designed to bypass ethical constraints.
It Takes One to Bias Them All: Breaking Bad with One-Shot GRPO
arXiv:2606.10931v2 Announce Type: replace Abstract: Warning: This paper contains several toxic and offensive statements. Modern large language models (LLMs) are typically aligned through large-scale post-training to ensure fair and reliable behavior. In this work, we investigate how easily such guardrails can be broken by Group Relative Policy Optimization (GRPO). We show that one-shot GRPO training on a single biased example is sufficient to induce systematic bias, with stereotype-driven reasoning generalizing across attributes, categories, and benchmarks. We further find that models differ in their susceptibility based on the initial likelihood of producing biased outputs. Our results reveal a critical vulnerability in post-training: alignment can be overridden by a single example.
Abstract:Warning: This paper contains several toxic and offensive statements. Modern large language models (LLMs) are typically aligned through large-scale post-training to ensure fair and reliable behavior. In this work, we investigate how easily such guardrails can be broken by Group Relative Policy Optimization (GRPO). We show that one-shot GRPO training on a single biased example is sufficient to induce systematic bias, with stereotype-driven reasoning generalizing across attributes, categories, and benchmarks. We further find that models differ in their susceptibility based on the initial likelihood of producing biased outputs. Our results reveal a critical vulnerability in post-training: alignment can be overridden by a single example.
Computation and Language (cs.CL)
arXiv:2606.10931 [cs.CL]
(or arXiv:2606.10931v2 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2606.10931 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 197 words
