data-selection

ICCLHerd 提出一种基于 herding 的 LLM 训练数据选择框架，利用 von Mises–Fisher 核平衡 mode-seeking 与 mode-averaging，并以 in-context learning 作为验证代理进行超参数搜索，实验表明其优于现有数据选择算法。

该论文研究了 LLM 预训练数据采集中的爬虫预算分配问题，发现文档质量在主机层面存在显著聚集性，通过先采样少量页面再按质量深入爬取，可比统一深度或基于中心度的分配多收集数倍高质量页面。

论文提出 TQS（Trajectory Quality Score），一种基于轨迹结构信号（信息聚焦、意图-行为对齐、策略动态）的线性时间评分框架，用于筛选高质量编码智能体轨迹进行监督微调。实验表明在 TQS 排名靠前的轨迹上训练的模型在 SWE-Bench Verified 上优于随机或低质量子集训练的结果。

This paper introduces FAER, an auditable full-trajectory replay framework for language model post-training that formalizes the gap between cache-level selection feedback and downstream learner utility, showing that a learner-aware selector (FAER-UTILITY) outperforms format-feedback and uniform baselines on GSM8K with Qwen2.5-1.5B-Instruct.

该论文通过冻结选中子集并只更换训练学习器的受控实验，证明样本价值（以及 easy-first 与几何覆盖策略的 crossover budget）是相对于目标学习器定义的，而非数据的固有属性。

This paper shows that when selecting math training data for LLM fine-tuning, sharing a reasoning approach (e.g., invariant reasoning or double counting) yields significantly greater transfer to held-out targets than sharing a mathematical topic, with advantages of 8–16 percentage points across five base models in counterbalanced 2x2 designs.

The paper introduces GapFT, a fine-tuning method that selects training examples based on a model's Pass@K–Pass@1 gap, showing on LogiQA 2.0 and ReClor that it raises single-sample accuracy dramatically while using far less data than uniform verified RFT.

AutoData introduces an AI agent that automates pre-training data selection for large language models by searching over executable algorithms, outperforming human-designed curation pipelines and improving downstream metrics.

The paper presents a data selection method for terminology-aware translation that trains only on hard examples where the model's output contradicts the glossary, achieving improved term accuracy, and describes the BSC system submission to the WMT26 Terminology Shared Task.

The paper proposes RMS-RSP, a perturbation-sensitive method for selecting medical questions to receive rationale supervision, improving robust accuracy and semantic consistency in QA systems.

DICS introduces a self-scoring metric for visual instruction data selection that enhances vision-language model performance with significantly less data by ensuring intra-sample consistency, outperforming state-of-the-art methods.

The paper introduces MASS, a hierarchical data selection framework for LLM post-training that uses manifold and sparse feature coverage to select high-value data subsets, outperforming existing baselines in experiments.

Data-DPO is a target model-oriented data selection method for LLM supervised fine-tuning that learns data preferences through one-step probing and combines them with quality scores and diversity, outperforming baselines on Vision-Flan and LLaVA-CoT datasets.

This research revisits likelihood-based data selection for reasoning fine-tuning, showing that its effectiveness depends on model capacity and training duration, with a capacity-dependent 'Fast-Fit/Slow-Gain' pattern observed.

Introduces CODS, an iterative critic-guided data selection method for offline reinforcement learning that retains task performance at low data budgets by selecting high-residual transitions over multiple rounds.

This paper proposes a generator-agnostic post-generation curation method that selects informative subsets of synthetic images by splitting real classes into canonical homogeneous and non-redundant heterogeneous subsets, and scoring synthetic images via a fidelity-diversity criterion. It consistently outperforms existing data-selection baselines and matches real-data performance with up to 40% fewer synthetic samples.

SIEVE is a structure-aware data selection method for vision-language-action imitation learning that identifies reusable visuo-motor primitives and transition interfaces, outperforming full-data training with only 50% of demonstrations and training steps.

Proposes CodeBlock, a structure-aware sparse supervision framework for supervised fine-tuning of code LLMs. It selects high-quality instruction-response pairs and partitions code responses into syntactically coherent coding items, applying loss only to selected items to achieve stronger pass@1 rates using only 1.9% of supervised response tokens.

This paper introduces Spokes, a probabilistic diversification framework using the G-Vendi score to optimize diversity in pretraining data selection, achieving significant improvements in downstream task performance on FineWeb and DCLM by jointly optimizing quality and diversity.

This paper demonstrates that data selection in low-resource verification regimes, where verifiers only have access to fragmented and biased slices of the target distribution, can paradoxically accelerate model collapse by pruning globally relevant tail modes. The authors provide theoretical proof and propose a collaborative proxy reference mechanism as a mitigation strategy.