How Post-Training Scaling Is Driving Major Capability Gains in Large Language Models
A controlled model iteration demonstrates how scaling long-horizon reinforcement learning after pretraining can significantly improve terminal control, software engineering, and cybersecurity performance without changing the base model, architecture, or parameter count. The case highlights environment scale, verifier quality, and infrastructure throughput as the three core levers of post-training scaling.
1. Case Background: A True Controlled-Variable Experiment
In recent years, the dominant narrative in the large language model industry has been a race for scale: trillion-parameter models, larger clusters, and longer pretraining cycles have become defining features of flagship systems. However, a recent iteration of a 753B-parameter MoE base model took a very different path.
The model retained the previous generation’s base model, architecture, total parameter count, and active parameters—approximately 40B active parameters per token across 78 layers, using MLA combined with a shared sparse-attention mechanism. It was not pretrained again. The only change was the addition of roughly one month of long-horizon reinforcement learning during post-training.
From a methodological perspective, this setup is highly valuable. Under standard experimental design principles in deep learning, when all other variables are held constant, performance gains can be attributed with relatively high confidence to post-training scaling itself. The technical team explicitly described the iteration as a “controlled-variable experiment.”
Such experiments are uncommon among leading models with frequent release cycles. They also provide a valuable opportunity to reconsider the boundary conditions of scaling laws.
2. Why Post-Training? Rethinking Scaling Laws
2.1 The Reality of Diminishing Pretraining Returns
Pretraining scaling laws describe how loss decreases according to power-law relationships involving model size, data volume, and compute. However, as high-quality training data becomes increasingly scarce and compute costs continue to rise, the marginal returns from simply adding more pretraining compute are declining rapidly.
The industry has broadly observed that pretrained models are approaching saturation in terms of knowledge capacity. Yet substantial room for improvement remains in the execution of capabilities, including:
- Multi-step tool use
- Long-horizon task planning
- Complex interaction with external environments
- Reliable completion of extended workflows
2.2 Post-Training as a New Scaling Dimension
Post-training—including SFT, RLHF, and RLVR—is fundamentally about teaching a model how to use the knowledge it already possesses. Reinforcement-learning-driven post-training has three critical scaling levers:
- Environment scale: The number and diversity of interactive training environments—such as terminal simulators, code repositories, and cybersecurity ranges—can be expanded substantially. In this case, the environment scale reportedly increased by dozens of times.
- Verifier capability: The coverage and accuracy of automated reward verifiers determine the upper bound of reinforcement-learning signal quality.
- Infrastructure throughput: RL training requires large-scale rollouts and experience replay. The engineering efficiency of the training framework directly limits the number of training iterations that can be completed.
Together, these form a three-part scaling system for post-training. None of the three can be ignored.
3. Observed Capability Improvements
3.1 Terminal Control
On a benchmark measuring real terminal operation capabilities, the model’s score increased from 4.6 for the previous generation to 28.3—an approximately sixfold improvement.
The benchmark measures success on long chains of multi-turn shell tasks, including software installation, debugging, troubleshooting, and automated script development. Terminal environments are particularly suitable for large-scale RL training because their outcomes can be verified automatically: the result of command execution can serve directly as a reward signal.
3.2 Real-World Software Engineering
On a benchmark involving defect fixing and feature development in simulated real-world repositories, the model improved by more than 20 points.
Software engineering tasks are difficult because their feedback is sparse and verification can be expensive. Determining whether a fix is correct may require running an entire test suite. In this case, an automated build and executable test-verification pipeline allowed RL training to scale across a large number of real repositories—something that pure SFT is unlikely to achieve as effectively.
3.3 Cybersecurity
On public evaluations involving vulnerability discovery and exploitation, the model reached the top of global rankings, while its score on offensive and defensive security benchmarks more than doubled.
Cybersecurity tasks are highly compatible with reinforcement learning because they are adversarial, have clearly defined verification criteria, and can be reproduced in controlled environments. The results also support the team’s stated positioning around preparing AI systems for cyber defense.
4. The Critical Role of Technical Infrastructure
Supporting long-horizon environment scaling requires solving several engineering challenges:
- Automatic environment generation: Template-based and parameterized methods are needed to create large numbers of training environments with built-in verifiers, avoiding a bottleneck caused by manual labeling.
- Heterogeneous compute scheduling: RL training alternates between inference for rollouts and model updates. Efficient two-stage resource scheduling and GPU memory management are therefore essential.
- Long-context optimization: A single interaction in a long-horizon task can reach hundreds of thousands of tokens. KV-cache management, sparse attention, and context caching become necessary. This is also the engineering significance of the model’s support for a 1M-token context window and 128K-token continuous output.
A simplified RL training loop is shown below:
while training:tasks = env_sampler.sample(batch_size) # Sample from the expanded environment pooltrajectories = policy.rollout(tasks) # Generate multi-turn interaction trajectoriesrewards = verifier.batch_check(trajectories) # Verify trajectories automaticallypolicy.update(trajectories, rewards) # Update the policy with policy gradients
Every optimization in this loop directly increases the scale of environments that can be trained.
5. Methodological Implications for the Industry
First, model iteration does not have to remain tied to pretraining. Once a base model has developed strong fundamental capabilities, post-training can serve as an independent and cost-effective path for iteration. This is especially useful when responding quickly to specialized application requirements.
Second, evaluation tasks can also function as training resources. Domains with automatically verifiable outcomes—such as terminals, software development, and cybersecurity—are highly productive areas for RL-based post-training. Tasks with strong subjectivity, by contrast, still require preference optimization and follow a different scaling path.
Third, the release strategy for open model weights is worth noting. The team has committed to releasing the complete weights only after completing safety evaluations and hardening. This reflects a responsible trend in which stronger capabilities are accompanied by earlier and more extensive safety assessment, providing a useful example for the broader ecosystem.
Overall, this iteration—achieving gains through post-training rather than parameter expansion—demonstrates under controlled conditions that post-training scaling has become a distinct, reproducible, and engineerable capability-growth curve alongside pretraining.
For developers and enterprises, understanding this shift in methodology may be more valuable over the long term than simply following the leaderboard scores of individual models.
