LBLLM: Efficient Lightweight Binarization for Large LLMs
Discover LBLLM, a three-stage distillation method for lightweight binarization of large language models, boosting efficiency and accuracy in low-resource s...
How Answer Tokens Read Reasoning Traces in LLMs
Explore how answer tokens interpret reasoning traces in LLMs, improving accuracy in quantitative reasoning with self-reading patterns and SRQ scores.
Nexusformer: Efficient Nonlinear Transformer Scaling Method
Discover Nexusformer, a novel nonlinear attention method enabling stable, efficient Transformer scaling with 41.5% less compute and preserved knowledge.
ST-Prune: Efficient Token Pruning for Autonomous Driving AI
Discover ST-Prune, a training-free spatio-temporal token pruning method that boosts Vision-Language Models' efficiency in autonomous driving.
Verbal Tics in Large Language Models: Systematic Study
Explore the rise of verbal tics in top LLMs, their impact on naturalness, and the need for improved alignment to enhance AI-human communication.