analiza aiqbench
This research reveals that helpful-only fine-tuning often leads to emergent misalignment, residual refusals, and poor steerability, rather than these being inherent necessities of the approach. The study demonstrates that these alignment issues can be effectively mitigated through synthetic document fine-tuning and the inclusion of character-related questions during supervised fine-tuning and reinforcement learning. For practitioners, the key takeaway is that simple anti-refusal training may exacerbate these problems, whereas targeted data strategies can preserve helpfulness while maintaining model integrity.
(Mis)generalization of Helpful-only Fine-tuning
arXiv:2606.04413v1 Announce Type: new Abstract: Helpful-only models, that is, models that are trained to always follow user intent, are valuable for dangerous capability evaluations and other areas of AI R&D where refusals would be an obstacle. Little is known about the generalization properties of helpful-only training: helpful-only models refuse less than their harmless counterparts, but previous work has not studied other dimensions of their alignment. We study the shortcomings of existing helpful-only models. We find that some show emergent misalignment, others have residual refusal behaviors, and most show poor steerability, sycophancy, and incoherent character. We show that simple anti-refusal training can cause many of these issues. None of these problems are necessary consequences of helpful-only training, though: we show that synthetic document fine-tuning and adding character-related questions to SFT and RL can mitigate them.
Abstract:Helpful-only models, that is, models that are trained to always follow user intent, are valuable for dangerous capability evaluations and other areas of AI R&D where refusals would be an obstacle. Little is known about the generalization properties of helpful-only training: helpful-only models refuse less than their harmless counterparts, but previous work has not studied other dimensions of their alignment. We study the shortcomings of existing helpful-only models. We find that some show emergent misalignment, others have residual refusal behaviors, and most show poor steerability, sycophancy, and incoherent character. We show that simple anti-refusal training can cause many of these issues. None of these problems are necessary consequences of helpful-only training, though: we show that synthetic document fine-tuning and adding character-related questions to SFT and RL can mitigate them.
Machine Learning (cs.LG)
arXiv:2606.04413 [cs.LG]
(or arXiv:2606.04413v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2606.04413 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 · 223 cuvinte