aiqbench analysis
This research demonstrates a critical vulnerability in knowledge distillation where student models inadvertently inherit the safety-pragmatism dispositions of their teachers, even when trained on non-safety tasks. The study reveals that behavioral tendencies are not merely incidental but can be intentionally imprinted, significantly altering downstream decision-making without explicit task supervision. These findings highlight the need for rigorous auditing of teacher-student pairs to prevent unintended propagation of bias or safety profiles in deployed AI systems.
Imprinting: Knowledge Distillation Can Shift Student Behavior
arXiv:2610.18517v1 Announce Type: new Abstract: Knowledge distillation trains student models on teacher-generated responses to transfer task capability, but those responses may also carry behavioral tendencies of their source. We investigate whether the source teacher can shape student behavior beyond the capability that distillation is intended to transfer. We construct harm-accepting and harm-avoiding variants of the same base teacher, use them to generate responses to ordinary reasoning and knowledge questions unrelated to safety, and distill those responses into students with different base models. Students generally acquire task capability while their downstream decisions shift toward the source teacher's safety-pragmatism disposition, a phenomenon we call imprinting. Distillation from the harm-accepting teacher reduces harm avoidance, whereas distillation from the harm-avoiding teacher increases it. Imprinting varies across teacher-student pairs and does not extend uniformly across different forms of safety behavior. Excluding a sparse set of divergence tokens reduces imprinting, whereas response rewriting preserves the source-dependent effect and can introduce additional rewriter-dependent shifts. Post-distillation preference optimization raises harm avoidance but does not eliminate the source-dependent difference. These results show that behavioral tendencies can be intentionally imprinted through the source teacher, even with ordinary task supervision.
Abstract:Knowledge distillation trains student models on teacher-generated responses to transfer task capability, but those responses may also carry behavioral tendencies of their source. We investigate whether the source teacher can shape student behavior beyond the capability that distillation is intended to transfer. We construct harm-accepting and harm-avoiding variants of the same base teacher, use them to generate responses to ordinary reasoning and knowledge questions unrelated to safety, and distill those responses into students with different base models. Students generally acquire task capability while their downstream decisions shift toward the source teacher's safety-pragmatism disposition, a phenomenon we call imprinting. Distillation from the harm-accepting teacher reduces harm avoidance, whereas distillation from the harm-avoiding teacher increases it. Imprinting varies across teacher-student pairs and does not extend uniformly across different forms of safety behavior. Excluding a sparse set of divergence tokens reduces imprinting, whereas response rewriting preserves the source-dependent effect and can introduce additional rewriter-dependent shifts. Post-distillation preference optimization raises harm avoidance but does not eliminate the source-dependent difference. These results show that behavioral tendencies can be intentionally imprinted through the source teacher, even with ordinary task supervision.
Computation and Language (cs.CL)
arXiv:2610.18517 [cs.CL]
(or arXiv:2610.18517v1 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2610.18517 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.
Full text under the publisher's open licence · 279 words
Related coverage
Exploring Hierarchical Interest Representation For Meta Ads Deep Funnel Optimization
Meta Engineering · Jul 15, 2026
[AINews] not much happened today
Latent Space · Oct 3, 2026
Frontier post-training recipe review with Finbarr Timbers
Interconnects · Jun 16, 2026
Dyna-2: General Purpose Physical Agents with Hour-Long Autonomy — Jason Ma, Dyna Robotics
Latent Space · Oct 13, 2026