Adaptive Generate-Rank-Verify: Inference-Time Search with Costly Verification
Researchers have developed a new algorithm called ADAP for optimizing inference-time pipelines in language models. This method is designed for scenarios where a cheap reward signal is used alongside a more expensive verification process, such as checking mathematical solutions or executing code. ADAP adaptively increases the number of sampled responses and verifications to find a positive example efficiently, outperforming fixed or difficulty-adaptive baselines in experiments. AI
IMPACT Optimizes inference efficiency for complex language model tasks like code generation and mathematical reasoning.