Post-Training Scaling: A New Paradigm for Unlocking Large Model Capabilities
A controlled model iteration shows how scaling long-horizon environments, automated verifiers, and reinforcement learning infrastructure can significantly improve performance without changing the base model, parameter count, or architecture. The results highlight where post-training scaling delivers the greatest value: verifiable, long-horizon, tool-intensive tasks such as terminal control, software engineering, and cybersecurity.
Introduction: A Differentiated Path Beyond Parameter Scaling
Over the past two years, competition in the large language model industry has largely centered on parameter counts. Models have expanded from hundreds of billions to trillions of parameters, and pre-training Scaling Laws have become widely accepted. However, the marginal benefits of additional pre-training are beginning to decline. The availability of high-quality data, training costs, and inference costs all impose significant constraints.
Against this backdrop, a different technical path is attracting attention: moving the focus of “scaling” from pre-training to the post-training stage.
A recent version update of a coding model provides a useful case study. The vendor explicitly stated that the new version uses exactly the same base model as its predecessor. The total parameter count, activated parameters, and architectural configuration were unchanged. The only variable was approximately one additional month of large-scale training and reinforcement learning in long-horizon environments.
Methodologically, this is close to a controlled experiment. It offers a relatively clean way to answer an important question: how much improvement can post-training scaling actually deliver?
Experimental Design: Why a Controlled Comparison Matters
1. The Base Configuration Remained Unchanged
According to the publicly disclosed information, the model continues to use a mixture-of-experts architecture with approximately 753 billion total parameters and around 40 billion activated parameters per token. It also uses sparse attention and related optimization techniques, supports million-token context windows, and enables extremely long continuous outputs.
These configurations are identical to those of the previous version. In principle, this means that any capability differences observed in the evaluation cannot be attributed to increased model capacity or architectural changes.
2. The Only Variable Was Post-Training Scaling
The improvements in the new version focused on three areas:
- Scaling long-horizon environments: The interactive environments used for reinforcement learning—including terminal simulations, code repositories, and cybersecurity ranges—were expanded by tens of times. This allowed the model to repeatedly explore and learn from a wider range of tasks that more closely resemble real-world scenarios.
- Automated verifier construction: A task verifier acts as the scoring mechanism for reinforcement learning. Automated methods were used to generate large numbers of verifiable tasks and evaluation rules, reducing the cost of manual annotation while improving the coverage and consistency of reward signals.
- Reinforcement learning infrastructure: Engineering improvements in throughput, stability, and sample reuse enabled the training system to complete substantially more computation than the previous version within the same one-month time window.
Together, these changes effectively upgraded reinforcement learning from a fine-tuning step into a large-scale training process. This can be understood as a practical implementation of Scaling Laws on the post-training side.
Performance Gains: Where the Capability Leap Appeared
Public evaluation results indicate that the improvements were not evenly distributed across all capabilities. Instead, they were concentrated in tasks that are verifiable, long-horizon, and heavily dependent on tool use.
Terminal control: On benchmarks that evaluate a model’s ability to perform multi-step operations in real terminal environments, the score increased from 4.6 to 28.3—an improvement of approximately six times. These tasks require the model to execute commands continuously, observe feedback, and adjust its strategy. Errors can accumulate across a long chain of actions, making this exactly the type of scenario where scaled reinforcement learning environments can have the greatest impact.
Real-world software engineering: Performance on tasks simulating defect resolution in real software repositories improved by more than 20 points. These tasks are more representative of everyday engineering work than simple code completion because they require issue localization, code modification, and validation.
Cybersecurity: The model achieved a leading score of 84.5% on an offensive and defensive security benchmark, while its performance on vulnerability exploitation tasks more than doubled. Security tasks naturally provide verifiable outcomes: whether an exploit succeeds is a clear training signal. This makes them particularly suitable for reinforcement learning during post-training.
This distribution supports an important conclusion: the benefits of post-training scaling do not primarily appear as an increase in static knowledge, which is largely determined by pre-training. Instead, they appear in strategic, long-horizon, and verifiable behavior.
Technical Principles: Why Post-Training RL Can Produce Large Gains
1. Reward Signal Density Determines Learning Efficiency
Supervised fine-tuning teaches a model to imitate good answers. Reinforcement learning, by contrast, allows the model to learn through trial, error, and feedback.
When the number and diversity of environments increase by tens of times, the model is exposed to a much broader task distribution. Reward signals become more widely available, and the model’s ability to generalize long-horizon strategies can improve significantly.
2. Verifiable Tasks Have a Special Role
Not every task is well suited to reinforcement learning. The reward for open-ended writing can be difficult to quantify, while tasks such as determining whether code compiles, whether a vulnerability can be exploited, or whether a command achieves its goal provide binary or multi-level outcomes that can be evaluated directly.
These relatively clean signals help explain why the largest gains appeared in terminal control, software engineering, and cybersecurity.
3. Automated Verifiers Enable Scaling
Manually creating high-quality tasks and environments is extremely expensive. Automated verifier construction makes it possible to continuously expand the number of environments in much the same way that training data can be expanded.
This is the engineering foundation that makes post-training scaling practical.
A simplified workflow is shown below:
Large-scale task environment pool↓Multi-turn agent interaction sampling(terminals, code, security ranges)↓Automated verifier scoring↓Policy gradient updates(reinforcement learning infrastructure)↓Iterative training → capability evaluation → release
Implications for the Industry and Developers
Parameter competition is not the only path. When base models have similar capabilities, post-training quality may become the main source of differentiation. Smaller teams could potentially achieve meaningful returns by investing relatively limited computing resources in reinforcement learning environments and verifier systems.
Verifiability will become a central task-design principle. The upper limit of future agent products may depend heavily on whether real business problems can be converted into tasks that can be evaluated automatically.
Developers should evaluate long-horizon performance. When selecting a model, developers should pay more attention to measured performance on long-horizon, tool-intensive tasks rather than relying only on parameter counts or static knowledge benchmarks.
Open weights require a balance with safety evaluation. The model’s complete weights were released only after a safety evaluation conducted approximately two weeks after the initial release. This reflects the trade-off between frontier capabilities and misuse risks, and this release cadence may become more common across the industry.
Conclusion
This model iteration provides near-experimental evidence that substantial gains in real-world engineering and security tasks can be achieved without changing the base model, parameter count, or architecture. The combination of scaled long-horizon environments, automated verifiers, and optimized reinforcement learning infrastructure produced improvements of several times in some capabilities.
Post-training scaling is emerging as an independent and reusable path alongside the pre-training-dominated Scaling Law framework. For the industry as a whole, this may signal a shift in the source of competitive advantage: the next generation of progress may depend not only on building larger models, but also on creating better training environments and more trustworthy feedback signals.
