Making Advanced AI Reasoning More Efficient, by Youkang Wang

EDPS is pleased to recognise Youkang Wang, a final-year BSc (Honours) Computer Science student at The Hong Kong Polytechnic University, as one of the joint highest-scoring awardees of the EDPS Innovation Scholarship 2025/26.

His project, Efficient Large Language Model Inference-time Scaling and Training, addresses a growing challenge in artificial intelligence: advanced large language models can improve their answers by generating and evaluating more responses, but this requires substantial computing resources, energy and cost.

Youkang developed three complementary frameworks—OptScale, MarkovScale and OptPO—to help AI models determine how much computation is genuinely needed. Instead of relying on fixed budgets or trial-and-error rules, the frameworks apply probability theory, Markov-chain analysis and Bayesian sequential testing to decide when a model should continue reasoning and when it should stop.

OptScale focuses on situations in which a model generates multiple answers and selects the best one. It estimates the number of samples required to reach a target level of confidence, reducing unnecessary generation. On the DeepSeek-R1-Distill-Qwen-7B model, OptScale reduced token consumption by up to 56% while matching or exceeding the accuracy of established approaches.

MarkovScale examines multi-step reasoning and estimates whether another round of refinement is likely to improve or weaken an answer. OptPO applies a Bayesian stopping method to test-time policy optimisation, ending the generation of training rollouts once there is sufficient evidence to support a leading answer. Across the reported benchmarks, OptPO achieved token savings of 30% to 50% while maintaining reasoning accuracy.

The project has produced three first-author research papers. OptScale was published at the AAAI Conference on Artificial Intelligence in 2026, while MarkovScale and OptPO were submitted to NeurIPS 2026, according to the scholarship proposal.

By reducing unnecessary token usage, Youkang’s research could lower the cost and environmental impact of advanced AI. It may also make high-performance reasoning tools more accessible to universities, startups, public-sector organisations and smaller enterprises.

His work reflects the purpose of the EDPS Innovation Scholarship: recognising rigorous technical research that can create meaningful academic, commercial and societal value.