Abstract
Strong reasoning depends not only on model knowledge but also on how effectively cognitive behaviors are deployed during generation. Existing methods often rely on explicit behavior-level control, making them insufficiently adaptive when failures and required corrections vary across reasoning states, tasks, and models. To this end, we propose Latent Reward Steering (LRS), an adaptive inference-time framework that promotes cognitive behaviors by optimizing the sparse autoencoder (SAE) latent states that implicitly carry them.
Rather than relying on predefined cognitive behaviors or steering directions derived from them, LRS trains a latent reward model on reasoning traces by final answer correctness to estimate the quality of intermediate latent states. During inference, reward gradients provide state-specific correction directions for fragile latent states, while a reward and confidence gate restricts intervention to states the reward signal flags as fragile.
Experiments on multiple reasoning LLM backbones and benchmarks show that LRS consistently improves performance over various baselines, and post-hoc analyses further indicate that LRS implicitly promotes good cognitive behaviors that fix the original reasoning errors. Code is available at GitHub.
Blogger's Review: The proposed Latent Reward Steering method offers a novel adaptive strategy for reasoning LLMs, significantly enhancing model performance across diverse tasks, especially in terms of flexibility and accuracy when handling complex reasoning. By optimizing latent states instead of fixed behaviors, it strengthens the model's self-correction capabilities, making it a promising avenue for further exploration and application.