Presenting users with diverse responses from foundation models is crucial for enhancing user experience and accommodating varying preferences. However, generating multiple high-quality and diverse responses without sacrificing accuracy remains a challenge, especially when using greedy sampling. In this work, we propose a novel framework, Synthesize-Partition-Adapt (SP A), that leverages the abundant synthetic data available in many domains to elicit diverse responses from foundation models. By leveraging signal provided by data attribution methods such as influence function, SP A partitions data into subsets, each targeting unique aspects of the data, and trains multiple model adaptations optimized for these subsets. Experimental results demonstrate the effectiveness of our approach in diversifying foundation model responses while maintaining high quality, showcased through the HumanEval and MBPP tasks in the code generation domain and several tasks in the natural language understanding domain, highlighting its potential to enrich user experience across various applications.

Our framework is the infinite-horizon discounted Markov Decision Process (MDP).

Multimodal interleaved datasets featuring free-form interleaved sequences of images and text are crucial for training frontier large multimodal models (LMMs).

MIRcontains25uniquelabels,andweremoved the label "night" as it is not in the label set of any MLAPIs. For each instance in those datasets, we have evaluated the prediction from the mainstream ML APIs from 2020 to 2022. HAPI was collected from 2020 to 2022. For classification tasks, the predictions/annotations of each API were collected in the spring of 2020, 2021, and 2022. Theoriginal IMDB dataset hasbeenpartitioned into training and testing splits, and thus we used its testing split, including 25,000 textparagraphs.