SAERL leverages model internals for enhanced LLM reinforcement learning. By focusing on intrinsic data properties, it achieves better accuracy with fewer training steps.
The Latent Recurrent Transformer (LRT) offers a streamlined approach to language modeling by reusing hidden states for improved efficiency and performance, enhancing both language-modeling loss and in-context learning.
HiSpec leverages early-exit models to significantly speed up speculative decoding in LLMs, boasting up to 2.01x throughput improvement without compromising accuracy.
