Search papers, labs, and topics across Lattice.
The paper introduces Neuron-Activated Graph Ranking (NAG-based Ranking), a training-free method for selecting pretraining data relevant to specific target tasks. NAG identifies high-impact neurons in pre-trained LLMs for target inputs, constructs a graph based on these neurons, and ranks pretraining data by similarity to the target NAG. Experiments across six benchmarks demonstrate that NAG-based data selection improves target-oriented pretraining by 4.9% on average compared to random sampling and outperforms state-of-the-art baselines.
Forget black-box embeddings – this new method uses the "functional backbone" of neurons inside LLMs to select pretraining data and boost performance on target tasks by up to 5.3%.
Everyday tasks come with a target, and pretraining models around this target is what turns them into experts. In this paper, we study target-oriented language model (LM) pretraining by introducing Neuron-Activated Graph Ranking (NAG-based Ranking), a training-free and interpretable framework for target pretraining data selection. Rather than using black-box representations, our approach directly characterizes each target input by a sparse set of high-impact neurons in any off-the-shelf LLMs. Concretely, we quantify neuron impact and select the most influential neurons across layers into a compact Neuron-Activated Graph (NAG), and rank candidate data by NAG similarity to target examples. We conduct experiments across six benchmarks, where our NAG-based Ranking improves target-oriented pretraining by 4.9% on average over random sampling, and also outperforms state-of-the-art baselines by 5.3% accuracy on HellaSwag. It also remains effective under a more applicable multi-target setting, where our best setup surpasses two baselines by 1.1% and 4.1%, respectively. Furthermore, we provide a comprehensive analysis on why and how our NAG works, e.g., deactivating NAG-selected neurons (only 0.12% of all) causes a 23.5% performance collapse, and restricting NAG to the final layer incurs a 4.1% average drop, indicating that NAG captures a sparse"functional backbone"for learning target features. We release the code at https://github.com/asillycat/NAG.
