Daily AI Paper

Training a large language model is expensive, and after pretraining, every extra bit of data you choose for fine-tuning or reinforcement learning can affect quality. This paper tackles a practical bottleneck: how do you pick the most useful post-training examples without paying the cost of full backpropagation on every candidate? The key idea is surprisingly simple. Instead of computing gradients for all model parameters, the authors show that gradients from just the output layer are often enough to rank data effectively. That means you can estimate which samples will best improve the model using only a forward pass, which is much cheaper than a full backward pass. Their method, called LESSER, plugs into existing data-selection pipelines and cuts feature-extraction cost by nearly 10 times for supervised fine-tuning and about 3 times for reinforcement learning benchmarks, while preserving the performance of full-gradient selection. Why it matters is straightforward: as models and data pools keep growing, cheaper training decisions can translate directly into faster iteration, lower cost, and more practical post-training workflows.