TESS Framework Improves LLM Data Selection via PVM Objective
October 2, 2026
The TESS framework uses a Pointwise Value Matching (PVM) objective to replace per-sample weights with a selection network for training data. This method avoids the optimization instability and poor generalization seen in existing meta-learning approaches when selecting data for instruction tuning and safety.
HOW THIS AFFECTS YOU
●
builderThis provides a more scalable way to select high-quality training data for specialized instruction tuning.
●
researcherYou can leverage PVM to improve the transferability of data selection methods to unseen datasets.