模型感知的数据选择：内外部信息交互
Model-Aware Data Selection from In-and-Out Information Interplay
浏览论文内容
中文总结 AI 辅助
提出模型感知的数据选择方法CAP，利用模型生成与参考响应在早期和后期层表示的发散差距判断数据可访问性，在数学、代码和科学领域平均改进35.4%，仅用10%数据即可匹配全池训练。
中文摘要 AI 辅助
大语言模型（LLMs）是有效的表示，在预训练期间吸收了海量知识，但模型需要经过后训练才能可靠地访问这些知识并“知道它们知道什么”。我们观察到权重中存储的知识与流经模型的数据流之间存在一个有趣的秩平衡。在所有模型层中，我们发现隐藏状态（数据流）遵循U形模式，在早期层中表现出显著压缩，并在后期层解码阶段急剧上升。相比之下，权重秩遵循倒U形模式，在早期和后期层中秩非常低，而在中间层秩较高。我们将此解释为一种内外部信息交互：中间激活不需要携带权重稍后可以提供的內容，因此它们主要保留权重无法提供的内容。受此观察启发，我们提出了一种模型感知的数据选择方法，即反事实同化剖面（CAP），该方法通过利用模型生成与参考响应在早期和后期层表示之间的发散差距，来确定数据候选是否包含当前模型可访问的信息。在数学、代码和科学领域，CAP在不同选择预算下相比最强基线，相对于基础模型平均改进了35.4%。仅使用数据池的10%，CAP在数学和科学上就超过或匹配全池训练。我们进一步表明，CAP可迁移到多模态数据选择，并且对响应长度和噪声具有鲁棒性。
英文摘要
LLMs are effective representations that assimilate vast amounts of knowledge during pretraining, but post-training is necessary for models to reliably access this knowledge and "know what they know." We observe an interesting rank equilibrium between knowledge stored in the weights and the data stream passing through the model. Across all model layers, we find that the hidden states (data stream) follow a U-shaped pattern, showing substantial compression in early layers and a steep rise during the late-layer decoding phase. In contrast, the weight rank follows an inverted U-shaped pattern, with very low rank in the early and late layers and high rank in the middle. We interpret this as an in-and-out information interplay: intermediate activations do not need to carry content that the weights can supply later, so they primarily preserve what the weights cannot provide. Motivated by this observation, we propose a model-aware data selection method, CAP (Counterfactual Assimilation Profile), which can determine whether a data candidate contains information accessible to the current model by utilizing the divergence gap in early- and late-layer representations between model-generated and reference responses. Across math, code, and science domains, CAP delivers 35.4% greater average improvement over the base model than the strongest baseline under different selection budgets. With only 10% of the data pool, CAP surpasses or matches full-pool training on math and science. We further show that CAP transfers to multimodal data selection and is robust to response horizon and noise.
