The LLM Conjuncture: Genres in Flux and Progenitors of the Next Wave

The rapid evolution of large language models (LLMs) has reached a critical juncture, where the dominant paradigm of scaling Transformer-based architectures through computational expansion faces growing competition from more resource-efficient algorithm-enhanced hybrid models. A basic question arises: Can algorithmic innovation rival the performance gains achieved by parameter proliferation, or is computational scaling as an inescapable prerequisite for advancing model capabilities? Meanwhile, considering that various structures created in industry have become focal points of academic innovation as well, a comprehensive, genre-spanning synthesis of these innovations is now urgently required to illuminate technical directions amid the rapidly expanding LLM landscape. Motivated by this, we systematically surveyed extensive literature harvested from Web of Science, Scopus, arXiv, ACM, IEEE, Springer and Elsevier. Furthermore, we introduce three genres: compute-driven generalists, algorithm-enhanced hybrids, and domain-specialised verticals Guided by the marginal-gain principle, we map every major LLM family to the driver that chiefly produces its next performance increment. We analyze their core technical features in detail, provide comprehensive computational comparisons across genres, and logically derive future technology pathways based on these abundant analyses. Our analysis reveals that recent advances in algorithm innovations and formulation of hybrid models will creat a new paradigm for achieving competitive performance that is more efficient compared to the monolithic transformer architecture. Hopefully, this work provides a roadmap for researchers to navigate LLMs’ crossroads, balancing scalability with sustainability by framing these innovations within the broader context of historical paradigm shifts.