Machine Learning
International Conference on Learning Representations
DEADLINES
PAPERS
Studies how compactly Transformers represent formal languages. The proofs compare their representation size with automata and temporal logic, while exposing the computational difficulty of model verification.
Examines failures when instructions arrive over multiple conversation turns. Simulated conversations separate task aptitude from reliability and trace how premature assumptions affect later responses.
Computes the matrix polar decomposition used by Muon. Minimax-designed polynomial iterations use GPU-friendly matrix multiplications, with analysis of approximation error and finite-precision behavior.
Investigates safety alignment concentrated in the first output tokens. Case studies and regularized fine-tuning objectives explore deeper alignment and its persistence under subsequent fine-tuning.
Attributes training data contributions to a particular model run. It accumulates Shapley values of gradient updates rather than repeatedly retraining models on different data subsets.
Extends promptable segmentation from images to video. A Transformer with streaming memory works with an interactive data engine to maintain prompted object masks across frames.