Paper Radar

ELI5When you train a language model by having it generate its own examples and learn from a teacher model's feedback, the weight updates are surprisingly small and scattered—mostly in certain network parts—rather than dense rewrites. You can even skip 90% of updates and keep performance.

Problem solvedPost-training recipes blend on-policy learning with teacher guidance, but it was unclear what actually changes in the model. This work reveals the sparse update structure, letting practitioners identify which parameters to train and why dense optimizers matter more than sparsity tricks here.

ELI5A system that predicts how wildfires will spread across terrain, then automatically designs a plan for where and when aerial water/retardant drops should happen to minimize damage. It combines a neural network with fire simulation to test strategies against different weather and uncertainty scenarios.

Problem solvedWildfire suppression crews need to decide where to drop water/retardant in real-time under uncertainty about fire behavior. This tool automates strategy design by modeling fire spread and optimizing drop locations, helping operators make faster, data-driven decisions instead of relying on intuition.

ELI5Brain-inspired learning using random feedback weights doesn't work well in deep networks because error signals get squashed into low-dimensional spaces. Adding techniques that spread updates across more dimensions fixes this and makes the learning work much better.

Problem solvedFeedback alignment is a biologically plausible alternative to backprop, but it fails in deep networks. This identifies why (rank collapse) and shows how to fix it, making brain-inspired learning practical for real architectures.