Researchers have proposed a new framework called Latent Reward Steering (LRS) to improve the performance of large language models in reasoning tasks. LRS works by optimizing the latent states of these models to promote good cognitive behaviors during generation. This is done through an adaptive inference-time framework that estimates the quality of intermediate latent states based on their correctness. The framework has been tested on multiple benchmarks and shown to consistently outperform existing methods, with post-hoc analysis indicating that it implicitly promotes good cognitive behaviors. --- Why it matters: This matters because large language models often struggle with reasoning tasks due to their inability to effectively deploy cognitive behaviors during generation. LRS provides a potential solution by adapting to the specific needs of each task and model, making it an important contribution to the field of natural language processing. Source: https://arxiv.org/abs/2606.00726