Researchers at the University of Pennsylvania and the University of Texas at Austin have developed LESSER, a way to select examples for fine-tuning large language models. Instead of calculating a full-model gradient for every candidate, it estimates each example’s usefulness from a normal forward pass; the authors report lower selection compute with similar downstream results in their tests.
When adapting a language model to a task, a lab must choose which examples from a much larger pool to use. Gradient-based selectors estimate how each example would change the model’s parameters, then compare that signal with a small set of reference examples drawn from the target task. The paper calls these “query” examples. Calculating a gradient through the whole model for every candidate requires an extra backward pass, which becomes costly across a large pool.
LESSER keeps the existing selection rule but replaces the full gradient with a signal from the model’s output layer, the part that turns its final internal representation into probabilities for possible next tokens. It builds that signal from information produced during the ordinary forward pass, then ranks candidates against the task’s reference examples. That avoids backpropagating through every model layer just to score data.
In the authors’ supervised fine-tuning tests, feature extraction used 9.7 times fewer floating-point operations (FLOPs); their reinforcement-learning setup used three times fewer FLOPs to score candidate problems. In a separate online test that selected Countdown arithmetic problems for reinforcement learning, LESSER and the full-gradient method GradAlign reached the same final test accuracy.
These savings apply to feature extraction and scoring, not the full cost of model training. The supervised fine-tuning comparison also changes more than the feature: LESSER uses the base model, while the LESS baseline uses gradients collected at four warm-up checkpoints. So the experiment does not isolate how much of the result comes from replacing full gradients. The October 2 paper is a preprint, and its benchmarks cover only a limited set of models and tasks; it shows a promising way to reduce data-selection work, not yet a general reduction in training cost.
Sources
- LESSER paper and abstract (arXiv, version 1, October 2, 2026)
- Full paper in HTML
- Lead illustration: Figure 1 from the paper, reproduced under its CC BY 4.0 license.