A new arXiv paper introduces Decoupled Marginal Sharpening, a training-free approach to inference-time scaling for large language models. The method aims to improve LLM reasoning without updating model weights or relying on external reward models, addressing a key limitation of current scaling techniques.
The work builds on recent reward-free power-sampling methods, which define a sequence-level power target of the form $\pi_\alpha(y\mid x)\propto$ ... (the abstract is truncated at this point). The proposed decoupled marginal sharpening appears to refine this target, though the full details are not available in the source.
Because the abstract is cut off, the exact mechanism and experimental results are not described in the provided text. The paper is listed as arXiv:2610.20483v1, with announcement type "new", indicating it is a recent submission.