Mage: Evaluating LLM-Generated Game Scenes Beyond Compile Rate
Discover Mage, a multi-axis framework assessing LLM-generated game scenes on runtime success, structure, and mechanism adherence beyond compile-pass rates.
Cumulative Token Importance Sampling for LLM Policy Optimization
Discover how cumulative token importance sampling improves LLM policy optimization by reducing variance and bias for stable, efficient reinforcement learni...
SparseRL-Sync: Efficient Weight Sync with 100x Less Data
SparseRL-Sync reduces RL weight synchronization communication by 100x while maintaining lossless updates, boosting scalability and performance in bandwidth...
CSR Framework: Real-Time AI Policies with Massive State Caches
Discover how the CSR framework and ASR algorithm reduce latency for real-time AI policies using massive cached state representations in robotics.
Detecting Backdoors in SAE Architectures: Diff-SAE vs Crosscoders
Explore how Diff-SAE outperforms Crosscoders in isolating backdoors in language models, enhancing AI safety through effective detection methods.