In April, we saw around 20% decrease in research publications compared to March 2025. However, there is a similar trend between April and March last year, with a 12% decrease in volumes. On the other hand, compared to the same time last year, there is a 50% increase in research papers.

This month, we saw researchers focusing on refining existing reasoning capabilities, alongside significant efforts in enhancing model efficiency, exploring multimodal architectures, and developing better evaluation methods as most benchmarks.

Reasoning was a central theme for this month. Many papers explored how to improve, control, and understand the reasoning processes within LLMs, often using Reinforcement Learning.

Several studies applied RL to boost reasoning performance. Open-Reasoner-Zero demonstrated scalable RL training for reasoning on base models, while VAPO introduced a value-based RL framework addressing specific challenges in long Chain-of-Thought (CoT) tasks.