More about HKUST
Effective and Efficient Reasoning in Large Language Models: A Survey of Test-Time Reasoning Strategies and Optimization
PhD Qualifying Examination
Title: "Effective and Efficient Reasoning in Large Language Models: A
Survey of Test-Time Reasoning Strategies and Optimization"
by
Miss Eunseo JUNG
Abstract:
The reasoning capabilities of large language models (LLMs) have improved
substantially through increases in model scale, training data, and
post-training optimization. However, further gains through training-time
scaling require substantial computational, financial, and engineering
resources, limiting the practical viability of this approach. This constraint
has shifted attention toward test-time reasoning, where additional
inference-time computation is used to construct intermediate reasoning steps,
explore alternative solutions, or interact with external tools.
However, the growing reliance on inference-time computation introduces a
second scaling problem where additional reasoning computation is not
necessarily used effectively. Models may over-reason on simple problems or
invoke tools and verifiers whose overhead yields little improvement.
Conversely, difficult problems may receive insufficient computation under
fixed inference budgets. Reasoning effectiveness and inference efficiency are
therefore tightly coupled. Yet their methods are often studied separately,
which makes it difficult to obtain a coherent view of the trade-off between
reasoning performance and the computational resources required to achieve it.
A unified perspective is needed to connect these complementary research
directions.
This survey addresses this need by examining test-time reasoning through
complementary perspectives on effectiveness and efficiency. We develop a
unified taxonomy that distinguishes methods for improving reasoning quality
from those for optimizing the use of inference-time computation, while
recognizing that these roles can overlap. Using this framework, we compare
representative approaches in terms of their mechanisms, empirical benefits,
resource requirements, assumptions, and limitations, and identify open
challenges in adaptive computation, reasoning reliability, and cost-aware
evaluation.
Date: Friday, 21 August 2026
Time: 2:00pm - 4:00pm
Venue: Room 3494
Lift 25/26
Committee Members: Prof. Lei Chen (Supervisor)
Prof. Xiaofang Zhou (Chairperson)
Prof. Ke Yi