Researchers introduce CONE, a hybrid transformer encoder that embeds numbers, ranges, and gaussians while preserving unit and variable semantics. The model achieves 87.28% F1 on the DROP benchmark, a 9.37% improvement over prior state-of-the-art systems.
Researchers have introduced Mixture of Universal Experts (MoUE), a generalization of Mixture-of-Experts architectures that adds a new scaling dimension called virtual width. The approach reuses a shared expert pool across layers while maintaining fixed per-token computation, achieving up to 1.3% improvements over standard MoE baselines and enabling 4.2% gains when converting existing MoE checkpoints.
Researchers propose AlignVAR, a visual autoregressive framework for image super-resolution that addresses critical consistency problems in existing VAR models. The approach combines spatial consistency autoregression and hierarchical consistency constraints to achieve 10x faster inference with 50% fewer parameters than leading diffusion-based methods.
Researchers propose TTSR, a test-time training framework where a single LLM alternates between Student and Teacher roles to improve its own reasoning. The method generates targeted variant questions based on analyzed failure patterns, showing consistent improvements across mathematical reasoning benchmarks without relying on unreliable pseudo-labels.
Researchers have identified a critical gap in how large language models learn mathematics: they can solve problems but often don't understand the underlying concepts. A new reinforcement learning framework called CORE addresses this by using explicit concept definitions as training signals, rather than just reinforcing correct final answers.