# Data Selection

Offline reinforcement learning with verifiable rewards usually supervises every token of every response in a fixed sampled pool, although responses and tokens are not equally informative. GIST splits the bidirectional KL mismatch between the policy and the Gibbs-optimal target of KL-regularized RLVR into a forward and a reverse term, then pairs correct and incorrect responses by sampling temperature and supervises only low-probability tokens of correct responses and high-probability tokens of incorrect ones. On Qwen3-0.6B/GSM8K with all methods choosing from the same candidates, it supervises 66.6% fewer tokens than training on all candidates and reaches 45.79 strict and 49.73 flexible accuracy, above seven selection baselines.

Offline reinforcement learning with verifiable rewards usually supervises every token of every response in a fixed sampled pool, although responses and tokens are not equally informative. GIST splits the bidirectional KL mismatch between the policy and the Gibbs-optimal target of KL-regularized RLVR into a forward and a reverse term, then pairs correct and incorrect responses by sampling temperature and supervises only low-probability tokens of correct responses and high-probability tokens of incorrect ones. On Qwen3-0.6B/GSM8K with all methods choosing from the same candidates, it supervises 66.6% fewer tokens than training on all candidates and reaches 45.79 strict and 49.73 flexible accuracy, above seven selection baselines.

Agents use experimental feedback to revise data-engineering strategies, and across four tasks—math and code data selection, math reward design, and math curriculum learning—the selected strategies gain up to 13.34% absolute over generic baselines on AIME25 Pass@32 under fixed training conditions.

Agents use experimental feedback to revise data-engineering strategies, and across four tasks—math and code data selection, math reward design, and math curriculum learning—the selected strategies gain up to 13.34% absolute over generic baselines on AIME25 Pass@32 under fixed training conditions.

The work reports that directly incorporating a selection network into existing meta-learning for training-data selection (MTS) objectives causes unstable optimization and poor generalization due to weight suppression and persistent reliance on easy-to-learn features, and proposes Transferable Example Scoring and Selection (TESS), a scalable data-selection framework built on a Pointwise Value Matching (PVM) objective, which in experiments on LLM safety and targeted instruction tuning demonstrates strong transfer across datasets, from subsets to full corpora, and from smaller to larger models.

The work reports that directly incorporating a selection network into existing meta-learning for training-data selection (MTS) objectives causes unstable optimization and poor generalization due to weight suppression and persistent reliance on easy-to-learn features, and proposes Transferable Example Scoring and Selection (TESS), a scalable data-selection framework built on a Pointwise Value Matching (PVM) objective, which in experiments on LLM safety and targeted instruction tuning demonstrates strong transfer across datasets, from subsets to full corpora, and from smaller to larger models.