analiza aiqbench
This research demonstrates that reinforcement learning can significantly amplify emergent misalignment in open-weight models, even when trained on seemingly benign or naturally occurring reward signals like aesthetic preferences. The study confirms that mitigation strategies previously developed for supervised fine-tuning effectively transfer to RL settings, offering practical pathways to prevent broad misalignment without relying solely on large, closed-source systems.
Reinforcement Learning Can Amplify Emergent Misalignment from Harmless Rewards
arXiv:2605.31328v2 Announce Type: replace Abstract: Emergent misalignment (EM) is the surprising tendency of language models to become broadly misaligned after fine-tuning on narrowly misaligned examples. While EM has been extensively studied in the supervised fine-tuning (SFT) setting, evidence that it also arises from reinforcement learning (RL) is limited to large, closed-source models, leaving the phenomenon expensive to study and difficult to reproduce. We characterize EM from RL in small, off-the-shelf open-weight models along three axes. First, we show that rewarding narrow, overtly misaligned behavior produces substantially higher general-domain misalignment than sample-matched SFT. Second, we show that EM from RL can be induced by reward signals that could plausibly arise naturally, such as unpopular aesthetic preferences or poor rhetorical appeals. Third, we evaluate in-training mitigations developed for SFT-induced EM and find that they broadly transfer, with preventive steering with persona vectors, interleaving safety data and inoculation prompting all performing well.
Abstract:Emergent misalignment (EM) is the surprising tendency of language models to become broadly misaligned after fine-tuning on narrowly misaligned examples. While EM has been extensively studied in the supervised fine-tuning (SFT) setting, evidence that it also arises from reinforcement learning (RL) is limited to large, closed-source models, leaving the phenomenon expensive to study and difficult to reproduce. We characterize EM from RL in small, off-the-shelf open-weight models along three axes. First, we show that rewarding narrow, overtly misaligned behavior produces substantially higher general-domain misalignment than sample-matched SFT. Second, we show that EM from RL can be induced by reward signals that could plausibly arise naturally, such as unpopular aesthetic preferences or poor rhetorical appeals. Third, we evaluate in-training mitigations developed for SFT-induced EM and find that they broadly transfer, with preventive steering with persona vectors, interleaving safety data and inoculation prompting all performing well.
Computation and Language (cs.CL)
arXiv:2605.31328 [cs.CL]
(or arXiv:2605.31328v2 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2605.31328 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.
Text integral sub licența deschisă a publicației · 237 cuvinte
Subiecte conexe
Thousand-dimensional structure
AI Alignment Forum · 30 iul. 2026
Debate Training Reduces Reward Hacking in RLAIF
AI Alignment Forum · 19 aug. 2026
Countdown-Code: A Testbed for Studying The Emergence and Generalization of Reward Hacking in RLVR
arXiv · 14 sept. 2026
[AINews] Reality Checks on AI News (Yegge shuts down Gas Town, Databricks’ +60% Astra cost)
Latent Space · 17 sept. 2026