Language models struggle to adapt to changing user preferences in conversation

Researchers have developed a new benchmark to assess how well language models adapt to changing human preferences in conversational AI. The benchmark, which includes 250 personas across five languages and 25,000 conversational turns, revealed a significant gap between models' ability to recognize appropriate strategies and their ability to generate them. While models can identify correct strategies when given options, they tend to default to exploiting existing preferences in open-ended generation, showing a 35-46% drop in strategy calibration. Targeted training on state-conditional human preferences can improve model performance, narrowing the user acceptability gap by one-third for out-of-distribution personas and conversational styles. AI

IMPACT Highlights a key limitation in current LLMs regarding dynamic user preference adaptation, suggesting a need for improved training methodologies.

RANK_REASON Research paper detailing a new benchmark and findings on language model behavior. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• alphaXiv
• arXiv
• CatalyzeX Code Finder for Papers
• Connected Papers
• CORE Recommender
• DagsHub
• Gotit.pub
• Hugging Face
• Influence Flower
• Litmaps
• ScienceCast
• scite Smart Citations

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →