analiza aiqbench
This systematic study reveals a critical trade-off in LLM conditioning where efficient activation steering methods often severely degrade generation fluency, particularly on instruction-tuned models compared to base versions. The research highlights that while simple prompting and supervised fine-tuning work well for injecting concepts, they struggle with removal tasks, whereas textual metrics offer a cost-effective proxy for expensive evaluation scores. These findings provide essential guidance for researchers aiming to balance output control with quality in reliable AI deployment.
On The Effectiveness-Fluency Trade-Off In LLM Conditioning: A Systematic Study
arXiv:2606.12234v2 Announce Type: replace Abstract: Controlling the output of Large Language Models (LLMs) is a central challenge for their reliable deployment, yet a clear understanding of the involved trade-offs remains elusive. Current approaches to conditioning are often evaluated with a narrow focus on their effectiveness at injecting or removing a target concept, neglecting generation quality. We systematically investigate a range of conditioning methods in both injection and removal scenarios. We find that efficient steering methods frequently achieve conditioning at a steep cost to fluency. Furthermore, we identify a critical yet previously overlooked interaction with the training paradigm: activation steering methods are far less effective on instruction-tuned models than on their base counterparts. Simple prompting and full-fledged supervised fine-tuning, on the other hand, are viable options for concept injection, but are not as good at concept removal. Finally, cheaply computed textual metrics highly correlate to costly LLM-as-judge scores, and provide insights on the behavior of conditioning methods.
Abstract:Controlling the output of Large Language Models (LLMs) is a central challenge for their reliable deployment, yet a clear understanding of the involved trade-offs remains elusive. Current approaches to conditioning are often evaluated with a narrow focus on their effectiveness at injecting or removing a target concept, neglecting generation quality. We systematically investigate a range of conditioning methods in both injection and removal scenarios. We find that efficient steering methods frequently achieve conditioning at a steep cost to fluency. Furthermore, we identify a critical yet previously overlooked interaction with the training paradigm: activation steering methods are far less effective on instruction-tuned models than on their base counterparts. Simple prompting and full-fledged supervised fine-tuning, on the other hand, are viable options for concept injection, but are not as good at concept removal. Finally, cheaply computed textual metrics highly correlate to costly LLM-as-judge scores, and provide insights on the behavior of conditioning methods.
Computation and Language (cs.CL)
arXiv:2606.12234 [cs.CL]
(or arXiv:2606.12234v2 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2606.12234 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 · 244 cuvinte
Subiecte conexe
On The Effectiveness-Fluency Trade-Off In LLM Conditioning: A Systematic Study
arXiv · 11 iun. 2026
On the Effectiveness-Fluency Trade-Off in LLM Conditioning: A Systematic Study
Apple · 30 sept. 2026
Angles of attack for continual learning safety
LessWrong · 16 iun. 2026
Why do models task game?
AI Alignment Forum · 6 aug. 2026