On the Effectiveness-Fluency Trade-Off in LLM Conditioning: A Systematic Study

A paper titled "On the Effectiveness-Fluency Trade-Off in LLM Conditioning: A Systematic Study" has been published by Apple Machine Learning Research and accepted at EMNLP. Its authors are Iuri Macocco, Pau Rodríguez Lopez, Arno Blaas, Luca Zappella, Marco Baroni, and Xavier Suau Cuadros. Macocco and Baroni are affiliated with Universitat Pompeu Fabra; Baroni and Suau Cuadros are noted as equal contributors.

The work addresses the challenge of controlling Large Language Model output for reliable deployment, stating that a clear understanding of the involved trade-offs remains elusive. According to the paper, current conditioning approaches are often evaluated with a narrow focus on how effectively they inject or remove a target concept, while neglecting generation quality.

The authors systematically investigate a range of conditioning methods in both injection and removal scenarios. They report that efficient steering methods frequently achieve conditioning at a steep cost to fluency.

The study also identifies an interaction with the training paradigm that it describes as critical and previously overlooked: activation steering methods are far less effective on instruction-tuned models than on their base counterparts.

Simple prompting and full-fledged supervised fine-tuning are described as viable options for concept injection, but as not as good at concept removal.

Finally, the authors report that cheaply computed textual metrics highly correlate with costly LLM-as-judge scores and provide insights into the behavior of conditioning methods. The paper is listed under the research areas Methods and Algorithms and Speech and Natural Language Processing, with a publication date of September 2026.