C-voting: Boost Test-Time Accuracy Without Energy Functions
Discover C-voting, a confidence-based test-time voting method that improves neural model accuracy without explicit energy functions in reasoning tasks.
Self-Supervised & Predictive Representation Learning Guide
Explore self-supervised learning and predictive representation techniques to enhance AI models with robust, predictive capabilities from unlabeled data.
Overcoming Surrogate Hacking in Multi-Timescale PPO
Discover how Target Decoupling architecture improves multi-timescale PPO by preventing surrogate hacking and enhancing reinforcement learning performance.
Optimizing SFT-GRPO Data Overlap for Autoformalization
Discover how controlling SFT-GRPO data overlap boosts autoformalization accuracy and model performance in post-training AI optimization.
Chain of Uncertain Rewards with LLMs for Reinforcement Learning
Discover CoUR, a novel framework using large language models to optimize reward design in reinforcement learning, boosting efficiency and performance.