Effective test-time scaling for large language model reasoning

Determine principled and effective strategies for test-time scaling—i.e., increasing inference-time compute during generation—to improve reasoning performance in large language models across tasks.

Background

The paper discusses recent advances such as OpenAI’s o1 models, which improved reasoning by increasing the length of Chain-of-Thought at inference time (inference-time or test-time scaling). Despite these gains, the authors note that identifying broadly effective and reliable methods for test-time scaling remains unresolved.

They survey related approaches (process-based reward models, reinforcement learning, and search methods like Monte Carlo Tree Search and Beam Search) and report that none yet achieve the general reasoning performance of the o1 series, motivating their reinforcement-learning-based exploration with DeepSeek-R1 and DeepSeek-R1-Zero.

References

However, the challenge of effective test-time scaling remains an open question for the research community.

— DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning  (2501.12948 - DeepSeek-AI et al., 22 Jan 2025) in Section 1, Introduction

In computer science, while Agent_H was discovered autonomously on single-turn benchmark rubrics, how well it generalizes to other clinical settings (such as multi-turn medical dialogues) remains to be assessed.

— Accelerating Scientific Research with Gemini in the Real-World  (2608.26701 - Schmidgall et al., 27 Aug 2026) in Limitations and failure modes, Section 2.3, paragraph “Generalization boundaries”

While dynamic scaling strategies offer partial relief, striking a favorable balance between exploration breadth and computational efficiency remains an open challenge.

— Construting Reverse Thinking: Developing Large Language Models' Reverse Thingking Ability  (2609.24760 - Liu et al., 21 Sep 2026) in Section 1, subsection “Long Chain-of-Thought (Long CoT) reasoning,” paragraph on Extensive exploration

Second, our test-time scaling analysis only evaluates up to five trajectories ($k\leq5$), which means we cannot yet characterize behavior under larger compute budgets, including whether performance continues to improve, plateaus, or degrades beyond this range.

— RefCon: Iterative Refinement and Contrastive Memory Extraction for Context-Evolving Agent  (2609.39143 - Prathama et al., 30 Sep 2026) in Limitations section