May 2025 Research Trend Report

Table of Contents

• 1. Improving Existing Capabilities with RL Instead …

• 2. Scaling Laws for Transformers

• 3. The Multimodal Convergence: Unified Architectur …

• 4. Search within and outside of Reasoning

• 5. Autonomous Scientific Discovery and Evolution

• 5. Diffusion Language Models

• 7. New Updates on Evaluation and Benchmarks

• 8. Interesting Approaches That I Didn’t Have A Cha …

• My Final Thoughts

In May, we saw around 56% increase in research publications compared to April 2025, which is a very large jump compared to the same time last year, with only 5% increase in volume. Compared to last year, there is a 236% increase in research papers published in May.

The past month has witnessed a pivotal shift in our understanding of reinforcement learning's role in large language models. A growing evidence suggests that RL, rather than teaching models fundamentally new capabilities, primarily acts as a sophisticated excavation tool, unearthing and refining abilities already latent within pretrained models. This research trend report will examine this central discourse alongside other developments in scaling, multimodal integration, computational efficiency, and architectural improvements.

1. Improving Existing Capabilities with RL Instead of Creating New Ones

The May 2025 papers present a fascinating convergence around a central insight: reinforcement learning for large language models may not be teaching them fundamentally new skills, but rather surfacing and optimizing latent capabilities already embedded during pretraining. This revelation challenges our understanding of what RL actually accomplishes and has profound implications for how we approach model training and improvement.

The most striking evidence for RL's role as a capability elicitor comes from the "Spurious Rewards" paper. Their work on Reinforcement Learning with Verifiable Rewards (RLVR) demonstrates that even completely incorrect reward signals can lead to substantial performance improvements.

Subscribe to our premium insights to read more