Researchers propose AIRL-S, a framework that combines reinforcement learning and search-based methods to scale test-time computation for large language models. The approach uses adversarial inverse reinforcement learning to infer a dense reward model from reference trajectories, eliminating the need for labeled process data. This method is evaluated on eight benchmarks in mathematics, science, and code generation, showing an average improvement of 9% over the base model. --- Why it matters: This matters because it addresses the challenges of scaling test-time computation for large language models, which is crucial for complex reasoning tasks like math, science, and coding. The proposed framework could enable more efficient and effective use of these models in real-world applications. Source: https://arxiv.org/abs/2508.14313