ExpGSI cuts estimated computation for reward-guided language-model inference

Researchers introduced ExpBoN, a preprint method that adds exponential noise when selecting the best response from multiple language-model samples, allowing finer control over reward and deviation from the model's original output distribution. They integrated it with guided speculative inference to create ExpGSI, which uses reward signals while reducing estimated computation. Tests on MATH500, MMLU-STEM and Minerva Math found that ExpGSI maintained comparable accuracy while cutting estimated computation by 14%-39% across candidate budgets for Qwen2.5-Math and by up to 45% at 16 candidates for Qwen3. The findings are limited to the reported models, benchmarks and estimated compute measure.