Large language models will continue to get stronger, but data, evaluation, and safety are becoming new bottlenecks.

In the past few years, advances in large language models (LLMs) have been almost synonymous with "bigger": larger parameter counts, larger datasets, larger compute. However, the latest survey research shows that the next phase of LLM competition may no longer be determined by whose model is largest, but by who has high-quality data, who can evaluate models more accurately, and who can make models safer and more controllable.

Industry Background: From a Scale Race to Full Lifecycle Engineering

The survey "A Survey of Large Language Models," recently published in Frontiers of Computer Science, systematically reviews the development of LLMs. The article treats LLMs as a system with a complete lifecycle: pretraining builds the foundation for language, facts, and reasoning; post-training adapts the model to downstream tasks and aligns it with human preferences; usage methods (such as prompting, retrieval augmentation, and agent frameworks) determine how capabilities are activated; and evaluation tests whether models are truly reliable, safe, and robust.

This perspective implies that LLM progress is no longer simply a matter of scale on a single dimension. Model performance depends on the interplay of data, architecture, training strategies, alignment methods, inference-time techniques, deployment constraints, and evaluation criteria. A shortcoming in any one of these areas can limit a model's performance in real-world scenarios.

Market Impact: Reliability and Controllability Become Key to Enterprise Adoption

For enterprises, the value of LLMs lies not only in leaderboard scores, but also in whether they can run stably in complex business environments. Research shows that high-quality data is not unlimited. As model scale grows, clean, useful public data is becoming increasingly difficult to obtain. Low-quality web text, duplicate content, noise, bias, privacy and copyrighted material can all affect model performance and compliance risk. Data cleaning, deduplication, filtering, privacy protection, tokenization, and data mixture design are shifting from technical details to the core of model development.

Synthetic data is seen as one way to fill the data gap. Existing models can generate mathematical reasoning, code examples, question-answer pairs, and more. However, if a model is repeatedly trained on data generated by other models, lacking new information or reliable feedback, its capabilities may stagnate or even regress. Therefore, the future key is not producing more data, but producing novel, reliable, verifiable, and sustainable data.

Evaluation is another bottleneck. Traditional benchmarks are becoming saturated, and problems such as data contamination, leakage, and declining discriminative power make scores increasingly unable to reflect true model capability. Evaluation is shifting from "can it answer questions correctly" to "can it operate reliably, safely, and transparently in real-world environments." For complex reasoning, long contexts, tool invocation, and agent tasks, static tests are far from sufficient.

Alignment and safety are likewise becoming prerequisites for deployment. Models may hallucinate, may give confident answers when uncertain, and may leak sensitive information. As LLMs are integrated into search engines, office software, education, healthcare, scientific research, and software development, alignment and safety are no longer abstract research topics, but necessary conditions for commercialization.## Competitive Landscape: Who Benefits, Who Feels the Squeeze?

New bottlenecks are reshaping the competitive landscape. Enterprises with strong data governance capabilities, proprietary data sources, and mature evaluation systems will gain long-term advantages. For example, companies that hold high-quality industry data can train more specialized and more reliable models; companies that can build dynamic evaluation environments can identify model weaknesses faster and act on them.

In contrast, model companies that rely solely on public data and benchmark score-chasing may hit a growth ceiling. While the open-source model community has lowered the barrier to entry, differences in investment in data quality and safety alignment will widen the gap. In addition, the growing capability of AI Agents brings higher risks: when models execute tasks autonomously, they need more reliable error control and clear safety boundaries. This is creating new markets for security tools, evaluation platforms, and compliance services.

Enterprise Takeaways: What to Focus On

For enterprises currently deploying AI, the following points deserve attention:

1. Do not look at benchmark scores alone; evaluate the model's real performance in your own business scenarios and build internal evaluation sets. 2. Prioritize data compliance and privacy protection to ensure that training and inference data sources are legal and traceable. 3. Focus on model safety and controllability, and establish human review and emergency response mechanisms. 4. When exploring Agent applications, ensure sufficient error recovery and safety boundary mechanisms are in place to keep autonomy from amplifying errors.

Outlook: Future Directions

This survey suggests that LLM research is entering the full-lifecycle engineering stage. Over the next 12 months, we are likely to see more tools and services emerge around data governance, evaluation automation, and safety alignment. Within 24 months, enterprise AI procurement criteria will shift from "stronger models" to "more reliable systems." Within three years, data ownership, evaluation standards, and regulatory compliance could become the most important competitive moats in the AI industry.

Models will continue to get more powerful, but being powerful is no longer an isolated technical metric—it is the combined outcome of data, safety, evaluation, and engineering capabilities. For enterprises, investors, and policymakers, understanding this shift is the key to seizing the next phase of the AI industry.