The AI industry is moving beyond measuring raw model size and shifting focus to how well large language models improve through post-training, a process that has driven recent breakthroughs in reasoning and coding abilities. Two major developments released in September 2026 reveal that benchmarking post-training workloads is becoming as important as benchmarking the initial model training itself.

What Is Post-Training and Why Does It Matter Now?

Post-training is the process that happens after a base model learns from massive amounts of text data. Instead of just absorbing general knowledge, post-training refines the model to excel at specific tasks like writing code, solving math problems, or acting as an AI agent that can use tools. Since the second half of 2025, significant advances in large language model (LLM) performance have come almost entirely from scaling post-training rather than scaling the initial training phase.

Examples of this shift are striking. Small models like Qwen 3.8 27B are now approaching frontier-level performance, and teams have achieved generational improvements in models like GLM-5.3 by scaling post-training alone, without changing the base model at all. This represents a fundamental change in how AI labs prioritize their compute resources.

How Are Researchers Standardizing Post-Training Benchmarks?

MLPerf, a widely respected AI benchmarking consortium, released its first standardized post-training benchmark in September 2026, beginning with version 6.1 submissions in October 2026. The benchmark uses a technique called Reinforcement Learning with Verifiable Rewards (RLVR), which is particularly well suited to software engineering tasks where success or failure can be objectively measured.

Here is how the RLVR process works in practice: an AI-powered coding agent attempts to solve software engineering problems in a sandboxed environment. Each attempt is called a rollout. The system samples multiple rollouts for each problem and assigns a simple binary pass or fail reward. These traces are then fed into an optimization algorithm called Group Relative Policy Optimization (GRPO), which compares rewards within each group and updates the model to make higher-reward solutions more likely. The updated model weights are then used in the next iteration of rollouts.

The MLPerf benchmark selected Qwen 3.5 397B, the largest model in the Qwen 3.5 family released in February 2026, as its test subject. This is a mixture-of-experts (MoE) model with 397 billion total parameters but only 17 billion active parameters per token, making it efficient despite its size. The model is available as an open-weight model under the Apache 2.0 license, allowing any research team to run the benchmark.

What Makes This Benchmark Technically Complex?

Developing this benchmark required the most complex system-level integration in MLPerf Training history. The system must deploy three different components simultaneously: training, inference, and agent environments. At runtime, hundreds of concurrent rollouts compete for CPU time, GPU time, memory, and storage, and resources must be carefully balanced between them.

The benchmark uses the R2E-Gym dataset, which contains 700 training problems and 251 validation problems derived from real GitHub commits in Python projects including NumPy, Pandas, and Tornado. The agent is given a Linux container environment with a Git repository, all dependencies pre-installed, and an issue description. The task is to generate a patch that solves the problem.

To prevent a common problem called "reward hacking," where models learn to game the reward system rather than solve the underlying task, the evaluation uses test files that the agent cannot see or modify during execution. For example, a model might try to remove failing tests rather than fix the bugs causing the failures. Hidden test files prevent this kind of shortcut.

How Are Open-Source Teams Documenting Their Post-Training Recipes?

Beyond MLPerf's standardized benchmark, open-source teams are publishing detailed, reproducible post-training recipes. Rufus-Air, released in September 2026, is an open post-training recipe built on GLM-4.5-Air-Base, a 106-billion-parameter mixture-of-experts model with 12 billion active parameters.

The Rufus-Air pipeline demonstrates a carefully ordered sequence of post-training stages, each targeting different capabilities and using different reward types. The stages progress from basic to advanced capabilities and from hard, verifiable rewards to softer judge-based signals.

Steps to Understanding the Post-Training Pipeline Order

• Supervised Fine-Tuning (SFT): Establishes a strong capability foundation by training the model on high-quality examples. This stage is not just a warm-up; it builds broad capabilities that subsequent reinforcement learning stages refine rather than construct from scratch.
• Reasoning and Coding Reinforcement Learning: Applies RLVR to improve reasoning and coding abilities under low-noise verifiers. These stages run first because their rewards are least vulnerable to hacking, protecting the model from learning shortcuts early in training.
• Instruction-Following and Agentic Stages: Improve constraint following and add tool use in specialized environments. These stages use rule-based assertions and execution tests as rewards, running before the final stage to limit exposure to reward hacking.
• Reinforcement Learning from Human Feedback (RLHF): Shapes open-ended quality where no hard verifier exists. This stage runs last because its judge-based reward is most vulnerable to gaming, and the team wanted to minimize the time the model spent optimizing a gameable signal.

The ordering principle is not strictly by reward type but by exposure to reward hacking. Instruction following, which is close to what the model already does, runs before agentic stages because its judge has little room to be gamed. The preference reward of RLHF, a judge on an open-ended objective, is where the real risk exists, so that stage runs last.

Rufus-Air improves over the official GLM-4.5-Air post-trained release and is competitive with similarly sized open models like INTELLECT-3 and Nemotron-3-Super. Importantly, the recipe is reproducible for teams outside frontier labs because it builds entirely on open-source components and has a compute footprint small enough for academic and mid-size research teams.

Why Does Infrastructure Matter as Much as Algorithm Choice?

Both the MLPerf benchmark and Rufus-Air emphasize that infrastructure and engineering choices are not just implementation details but core parts of the recipe. The specific details that made post-training stages work include token-in and token-out rollouts for on-policy multi-turn reinforcement learning, consistent chat template handling from supervised fine-tuning through agentic stages, a sandbox service reliable for long agentic runs, and large batches with a technique called Rollout Routing Replay for stable reinforcement learning.

This shift toward documenting infrastructure reflects a broader recognition in the AI community that reproducibility requires more than publishing a paper with results. It requires sharing the engineering decisions, data handling practices, and system configurations that made those results possible.

The emergence of standardized post-training benchmarks and open recipes signals a maturation in how the AI industry measures progress. Rather than focusing solely on the size of models or the scale of initial training, the field is now asking harder questions about how efficiently models can improve through targeted reinforcement learning and how those improvements can be reliably measured and reproduced across different hardware and teams.