The Three-Stage Trajectory of Large Language Models: 2022 to 2026

Abstract

This literature review puts forward a three-stage model of Large Language Models (LLMs), delineates the research goals of LLM in three stages and predicts future trends. Against the rapid, fragmented advancement of large-language-model technologies, systematic sorting of their evolutionary trajectories helps researchers grasp core contradictions and shifting research priorities. It also offers actionable reference for subsequent technical exploration, practical deployment and policy formulation in the LLM domain. In Phase 1 (2022 to 2023), LLMs have improved reasoning and general capabilities, while facing conflicts between capability and reliability. In Phase 2 (2023 to 2025), LLMs enhance evaluation and robustness, while filling the gap between benchmarks and real-world performance. In Phase 3 (2025 to 2026), LLMs improve autonomous action and safe deployment, while balancing accessibility and security. This dilemma catalyzed Phase 4 - "World Model Era", emphasizing physical simulation and causal reasoning. This work synthesizes 12 milestone papers and sorts out the development process and future trend of LLM.