Over the past few years, progress in large language models has often been associated with a single word: bigger.
Larger parameter counts. Larger training datasets. Larger computational budgets. And, with each new generation, stronger performance across language, coding, reasoning, and instruction-following tasks.
From GPT-3 and ChatGPT to a growing number of open and closed models, LLMs have moved far beyond their origins as specialized tools for natural language processing. They are now becoming part of the technical infrastructure for research, software development, education, content creation, and industrial applications.
But as these models continue to improve, a different set of questions is becoming harder to ignore.
The next stage of LLM competition may not be determined simply by who builds the largest model. It may depend on who can obtain reliable high-quality data, who can evaluate model capabilities more accurately, and who can make models safer, more controllable, and more useful in real-world settings.
A recent review published in Frontiers of Computer Science offers a systematic account of this shift. The article treats large language models not as a single algorithm or product, but as systems that develop through an entire lifecycle. Pre-training builds broad linguistic, factual, and reasoning foundations. Post-training adapts models to downstream tasks and aligns them with human preferences. Utilization methods, including prompting, retrieval augmentation, and agentic frameworks, determine how capabilities are activated in practice. Evaluation then measures whether models are truly capable, reliable, safe, and robust.
This lifecycle view points to an important change in the field. LLM progress is no longer only a question of scale. Model performance depends on the interaction of data, architecture, training strategy, alignment methods, inference-time techniques, deployment constraints, and evaluation standards. A weakness in any one of these stages can limit how well a model performs outside controlled benchmarks.
One of the first bottlenecks is data.
During pre-training, LLMs learn from massive text corpora, including web pages, books, code, scientific literature, mathematical materials, dialogue data, and, increasingly, synthetic data generated by other models. According to the review, several representative open models are now trained on datasets reaching the trillion-token scale, while large closed models may rely on even larger or more carefully curated resources.
The problem is that high-quality data is not unlimited.
As models scale up, publicly available data that is clean, useful, diverse, and suitable for training becomes more difficult to obtain. Low-quality web text, duplicated passages, noisy content, biased information, private data, and copyrighted materials can all affect model performance and deployment risk. Data cleaning, deduplication, filtering, privacy protection, tokenization, and data mixture design are no longer minor technical details. They are becoming central to model development.
Synthetic data may help fill part of the gap. Existing models can generate mathematical reasoning traces, code examples, question-answer pairs, and long-chain reasoning data. Such data can expand training resources and improve models on specific tasks. But synthetic data also introduces new risks. If models are repeatedly trained on content generated by other models, without enough new information or reliable external feedback, their capabilities may stagnate or degrade. In the next phase, the key question will not be how to produce more data, but how to produce data that is novel, reliable, verifiable, and sustainable.
A second bottleneck lies in evaluation.
Determining whether an LLM is truly stronger is not as simple as looking at a few benchmark scores. As models improve, traditional benchmarks are increasingly saturated. Some tests also face data contamination, leakage, and declining ability to distinguish between models. A model may perform well on standard benchmarks while still behaving unreliably in complex real-world tasks.
The review treats evaluation as a major stage in the LLM lifecycle. It covers basic language understanding, knowledge, reasoning, alignment, safety, agentic tasks, and expert-level capabilities. This reflects a broader shift in the field: evaluation is moving from asking whether a model can answer questions correctly toward asking whether it can operate reliably, safely, and transparently in realistic settings.
This is especially important for complex reasoning, long-context processing, tool use, and agentic tasks. In these scenarios, a model may need to plan over multiple steps, call external tools, process feedback, revise its strategy, and recover from errors. Such abilities cannot be fully captured by static multiple-choice questions or short-answer tests. Future evaluation systems will need to be closer to real applications, while also addressing contamination, fairness, reproducibility, and safety.
A third bottleneck is alignment and safety.
Pre-training gives models broad knowledge and generative ability, but it does not automatically ensure that their outputs are helpful, truthful, or safe. Modern LLM development therefore depends heavily on post-training. Supervised fine-tuning, reinforcement learning from human feedback, reinforcement learning with verifiable rewards, and methods designed for long-chain reasoning and agentic behavior are now central to making models usable.
These techniques have helped transform LLMs into more capable assistants. Yet they also bring new challenges. Models can hallucinate. They can produce confident-sounding answers when they are uncertain. They may reveal sensitive information or generate unsafe content under certain prompts. As LLMs are integrated into search engines, office software, education, healthcare, research workflows, and software development, alignment and safety are no longer abstract research topics. They are prerequisites for deployment.
The review also emphasizes the growing importance of agentic capability. LLMs are increasingly used not only to answer questions, but also to plan, use tools, interact with external environments, and execute multi-step tasks. Retrieval-augmented generation, long-context management, self-improvement, multi-agent collaboration, and efficient deployment techniques are expanding the ways LLMs can be used.
But stronger agentic capability also raises the stakes. A model that can act over long task chains needs more reliable error control, clearer safety boundaries, and more appropriate evaluation methods. Without these, greater autonomy could amplify mistakes rather than simply improve productivity.
For this reason, the future of LLMs is unlikely to be a matter of simply making models larger. Larger parameter counts, longer contexts, and more complex reasoning processes may still bring important gains. But these gains must be considered together with data quality, training efficiency, alignment strategy, utilization methods, and evaluation standards.
LLM research is entering a full-lifecycle engineering stage.
This also changes how progress should be understood. In the past, public attention often focused on whether a new model could top a leaderboard, solve difficult math problems, write better code, or generate more fluent text. Those questions still matter. But researchers now face deeper questions: Is the training data reliable? Do benchmark results reflect real capabilities? Can the model recognize uncertainty? Can it avoid compounding errors in complex tasks? If it acts as an agent, does it operate within safe and controllable boundaries?
These bottlenecks do not mean that LLM progress is slowing to a stop. Rather, they suggest that the field is entering a more mature phase. Capability remains important, but reliability, controllability, and evaluability are becoming just as central.
The next phase of competition in large language models may therefore be less about scale alone and more about data governance, training methods, alignment and safety, agentic capability, and trustworthy evaluation. LLMs will likely continue to become stronger. Whether they can reliably serve science, industry, and society will depend on how well researchers address these emerging bottlenecks.