Fine-tuning large language models usually starts with a large pool of data drawn from many distinct domains, while downstream evaluation spans a wide range of tasks. That creates a complication: a single sample may be useful for one task and less useful for another. The authors of this paper argue that data selection methods should account for this task-dependent value rather than treating each sample in isolation.
The paper's central claim is that the right unit of selection is not the sample but the mixture. Instead of ranking individual examples and keeping the top-scoring ones, the authors propose a combinatorial approach that chooses a combination of data—essentially a mixture across domains—that best serves the full evaluation suite. This shifts the question from “which samples are best?” to “what proportions of domains make the best overall mix?”
The abstract does not include experimental results or algorithmic details, so the practical gains of this approach are not yet described in the source. But the framing is notable because it matches a known reality of LLM training: data value is contextual, and a good mixture may matter more than any individual sample.