State of LLMs in 2025: Progress, Challenges, and Predictions
DeepSeek
OpenAI
Alibaba/Qwen
Moonshot AI
NVIDIA
Google/DeepMind
In 2025, LLM development was dominated by reasoning models using reinforcement learning with verifiable rewards (RLVR) and the GRPO algorithm, sparked by DeepSeek R1. Key trends include a focus on inference-time scaling, MoE architectures, and efficiency tweaks like Gated DeltaNets. Academic research highlighted GRPO variants with improvements such as zero gradient filtering and token-level loss.
As 2025 concludes, the year was dominated by reasoning models using RLVR and GRPO, following DeepSeek R1's January release. DeepSeek R1 showed that reasoning behavior could be developed with reinforcement learning, and its open-weight model performed comparably to proprietary ones. The paper also fueled debate over training costs, with estimates around $5 million for the final run, though excluding researcher salaries. RLVR uses verifiable rewards (e.g., math, code) for post-training, reducing dependency on expensive human labels. Every major LLM developer released a reasoning variant. Other focus points: 2022 RLHF+PPO, 2023 LoRA SFT, 2024 mid-training. Predictions for 2026-2027 include RLVR extensions, inference-time scaling, and continual learning. GRPO became a research darling, with many mathematical improvements like zero gradient filtering, active sampling, token-level loss, and no KL loss, which significantly improve training stability. Architecturally, state-of-the-art models still use decoder-style transformers with MoE layers and efficiency-tweaked attention (grouped-query, sliding-window, multi-head latent attention). Newer efficiency tweaks include Gated DeltaNets in Qwen3-Next and Kimi Linear, and Mamba-2 layers in NVIDIA's Nemotron 3.
- Сокращения
- LLM = Large Language Model — Большая языковая модель
- RLVR = Reinforcement Learning with Verifiable Rewards — Обучение с подкреплением с верифицируемыми наградами
- GRPO = Group Relative Policy Optimization — Групповая относительная оптимизация политики
- RLHF = Reinforcement Learning from Human Feedback — Обучение с подкреплением на основе человеческой обратной связи
- SFT = Supervised Fine-Tuning — Контролируемая донастройка
- MoE = Mixture of Experts — Смесь экспертов
- PPO = Proximal Policy Optimization — Оптимизация проксимальной политики
- LoRA = Low-Rank Adaptation — Низкоранговая адаптация
- KL = Kullback–Leibler divergence — Дивергенция Кульбака-Лейблера
- DPO = Direct Preference Optimization — Прямая оптимизация предпочтений
- PRM = Process Reward Model — Модель награды за процесс
Source: Sebastian Raschka —
original