Low-Cost Candidate Exploration for Higher-Quality Large Language Model Outputs: A Unified Framework and 300 Proposed Methods

Large language models (LLMs) can produce high-quality responses, but obtaining a particularly strong output may require substantial inference-time computation. Generating multiple complete responses using a large model and evaluating every response with an expensive verifier can be computationally inefficient. This paper proposes Low-Cost Candidate Exploration for Large Language Models (LLM-LCCE), a unified framework that investigates whether higher-quality outputs can be obtained by first exploring a set of relatively inexpensive candidate responses and subsequently allocating additional computation to the most promising candidates. Rather than generating every candidate with the same model, token budget, reasoning depth, and computational configuration, LLM-LCCE organizes candidate exploration into multiple fidelity levels. Initial candidates may be produced using smaller models, shorter reasoning traces, constrained token budgets, reduced sampling costs, partial generation, or alternative prompting configurations. Lightweight evaluators then estimate candidate potential, diversity, uncertainty, and task relevance. A selection policy determines which candidates should receive additional computation, such as longer reasoning, stronger-model evaluation, independent verification, or selective refinement. The final output is selected according to task-specific criteria and an explicitly defined computational budget. The framework distinguishes candidate generation, candidate evaluation, candidate selection, and final-output verification. It also emphasizes the difference between method-only computational cost and end-to-end cost, including shared candidate-pool preparation. This distinction is important because a method that appears inexpensive after candidate generation may become inefficient when the cost of generating and evaluating the candidate pool is included. This paper presents a general architecture, mathematical formulation, evaluation protocol, and 300 author-proposed methods under the LLM-LCCE framework. The proposed methods cover predictive candidate ranking, low-cost candidate generation, multi-fidelity evaluation, adaptive computation allocation, uncertainty estimation, diversity management, verification, selective refinement, reasoning, coding, mathematical problem solving, question answering, summarization, and agentic tasks. The proposed framework is conceptual and methodological; the 300 methods are author-proposed formulations rather than experimentally validated algorithms. Their novelty, feasibility, and effectiveness require individual literature comparison and empirical investigation.