analiza aiqbench
This research demonstrates a critical vulnerability in LLM alignment, showing that Group Relative Policy Optimization (GRPO) can induce systematic bias from a single biased example. The findings highlight that guardrails are not foolproof and that model susceptibility varies based on their initial propensity for biased outputs. For practitioners, this underscores the need for rigorous stress-testing of alignment mechanisms against minimal adversarial inputs before deployment.
It Takes One to Bias Them All: Breaking Bad with One-Shot GRPO
arXiv:2606.10931v1 Announce Type: new 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:2606.10931v1 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2606.10931 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
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.
Text integral sub licența deschisă a publicației · 199 cuvinte
