The State of LLMs in 2025: Reasoning, Inference Scaling, and Open Problems

Summary

Sebastian Raschka’s year-end review argues that large language model progress in 2025 came from several interacting levers rather than one decisive breakthrough. The year’s central development was the spread of reasoning models after DeepSeek R1 demonstrated that reinforcement learning with verifiable rewards (RLVR), combined with GRPO, could improve problem-solving behavior during post-training. Because math and code provide deterministic correctness signals, RLVR can scale post-training without relying entirely on costly human-written answers or preference labels, although headline training-cost estimates omit salaries, experimentation, and other development expenses. The author describes GRPO as the year’s main research focus, with modifications such as active sampling, token-level losses, altered KL treatment, and importance sampling improving training stability and results in models including OLMo 3 and DeepSeek V3.2. He expects RLVR to expand beyond math and code, potentially using another language model to score explanations, while inference-time scaling will trade latency and cost for higher accuracy on difficult tasks. Continual learning is another expected direction, but catastrophic forgetting remains unresolved. Architecturally, decoder Transformers remain dominant, while mixture-of-experts, grouped or sliding-window attention, latent attention, Gated DeltaNet, and Mamba-style layers target memory and serving costs. He also highlights tool use as a way to reduce hallucinations and says local models will increasingly need secure tool-use infrastructure. The review cautions that benchmark optimization can produce misleading scores, citing Llama 4 as an example, so practical testing and continually refreshed evaluations remain necessary. In coding, writing, and research, the author sees LLMs as productivity tools that work best with human expertise rather than replacements for it; excessive delegation may weaken learning, motivation, and job satisfaction. He identifies proprietary domain data as a future competitive advantage and predicts more specialized in-house models. His 2026 outlook includes diffusion language models for low-latency inference, local agentic tool use, broader RLVR, less routine reliance on classical RAG, and more gains from inference systems and tooling than from model training alone.