How Post-Training Unlocks the Potential of Ultra-Large-Scale Models

Today, as the parameter scale of large pre-trained models continues to climb, a frequently debated question arises: when a foundation model already possesses hundreds of billions or even trillions of parameters, where is the leverage point for further capability improvement? Increasing practices indicate that the answer is not always to “train an even larger foundation model,” but rather lies in the refined design of the post-training phase. This article explores this topic from four dimensions: performance, cost structure, technical principles, and engineering practices, analyzing how post-training unleashes the latent capabilities of ultra-large-scale parameter models.

1. Performance Leap: Visible Improvements from Post-Training

The most direct way to measure whether a model has become “stronger” is to look at its performance on standardized benchmarks. Taking a commonly used comprehensive intelligence benchmark in the industry as an example, after upgrading its post-training strategy, a certain ultra-large-scale model saw its overall intelligence score increase from 44 in its previous generation to 55—an 11-point jump—propelling its ranking to the top tier globally. The noteworthy aspect of this shift is that the base parameter scale did not undergo an order-of-magnitude change; the improvement primarily originated from the post-training phase.

Breaking down the evaluation dimensions, these types of improvements typically manifest in the following areas:

• Reasoning capability: Significant accuracy improvements in multi-step logical deduction, mathematical calculation, and causal analysis tasks.
• Instruction following: More stable comprehension and execution of complex, deeply nested instructions.
• Knowledge retrieval: More precise extraction of learned knowledge in open-domain question answering, leading to a reduced hallucination rate.
• Format control: Enhanced standardization in structured output, code generation, and multi-language switching.

The common characteristic of these capabilities is that they do not solely depend on “how much” the base model remembers, but rather on “how” the model organizes and utilizes its acquired knowledge. This is exactly the core space where post-training plays its role.

2. Cost Structure: The Inverse Relationship Between Capability and Invocation Cost

For developers and enterprise users, model capability is only half of the decision-making equation; the other half is cost. A common phenomenon brought about by post-training is that as capabilities improve, the unit cost of invocation actually decreases.

This may seem contradictory, but it has an internal logic. Post-training (especially instruction tuning and preference alignment) enables the model to complete the same tasks with shorter outputs, reducing ineffective reasoning steps. Simultaneously, an aligned model has a higher probability of providing a usable result in its first response, thereby lowering the overhead of retries and multi-turn corrections. From an engineering perspective, this means:

1. Reduced inference token consumption: More precise instruction understanding reduces redundant output.
2. Lower failure retry rate: Improved output stability reduces fallback mechanisms and secondary API calls.
3. Improved caching and reuse efficiency: Structured outputs are easier for downstream systems to consume directly.

Therefore, when evaluating the economics of a model, one cannot merely look at the per-token price; instead, one should comprehensively consider the “total cost to complete a task.” Post-training optimization often yields more significant returns on the latter.

3. Technical Principles: Why Post-Training Unlocks Parameter Potential

To understand the role of post-training, a premise must be clarified: pre-training endows the model with “latent capabilities,” while post-training is responsible for “capability activation.”

During the pre-training phase, ultra-large-scale parameter models absorb massive amounts of corpus data, theoretically possessing broad knowledge and pattern recognition capabilities. However, in their raw state, these capabilities are “diffuse”—the model does not know which capabilities should be prioritized in which scenarios. Post-training completes this activation through the following mechanisms:

1. Instruction Tuning: Establishing Task-to-Capability Mapping

Through supervised fine-tuning on diverse instruction-response pairs, the model learns “when seeing this type of problem, which type of knowledge should be retrieved, and in what format to answer.” This is equivalent to paving clear pathways for common tasks within the vast parameter space.

2. Preference Alignment: Optimizing the Output Distribution

Relying solely on supervised fine-tuning, a model might learn to generate replies that “look like an answer” but are not “good enough.” Preference alignment (such as optimization methods based on human or AI feedback) further adjusts the output distribution, making the model lean towards more accurate, helpful, and safe responses. This step is particularly crucial for ultra-large-scale models because more parameters mean greater variance in the output space, making alignment necessary to narrow down the effective distribution.

3. Reasoning Enhancement: Enabling the Model to “Think Before Answering”

In recent years, an increasingly important component of post-training is the specific reinforcement of reasoning capabilities. By introducing mechanisms such as chain-of-thought, process supervision, or outcome-based rewards, the model is encouraged to perform explicit intermediate deductions before providing a final answer. For a base model with a sufficiently large parameter scale, this “slow thinking” capability can be effectively triggered, allowing it to perform on par with or even outperform larger models that have not been adequately post-trained on complex tasks.

4. Data Composition: Determining the Direction of Capability Release

The quality and mixture ratio of post-training data directly determine which capabilities are prioritized for release. A common practical principle is:

• General instruction data ensures baseline following capabilities;
• Domain-specific data (e.g., code, math, multi-language) directionally enhances specific skills;
• Negative examples and boundary data improve robustness and safety;
• Long-context and multi-turn data augment conversation and document processing abilities.

The mixing ratio of these data types often impacts the final outcome more than merely increasing the data volume.

4. Engineering Practices: Building a Reusable Post-Training Loop

From an engineering implementation perspective, post-training is not a one-time training task but a continuously iterating loop. A reusable workflow typically includes the following stages:

1. Baseline Evaluation → Identify current capability shortcomings
2. Data Construction → Create instruction/preference data targeting shortcomings
3. Training Execution → Supervised fine-tuning + Preference optimization + Reasoning enhancement
4. Regression Evaluation → Verify improvements without degrading existing capabilities
5. Cost Assessment → Measure actual inference overhead per task
6. Canary Deployment → Validate online performance with a small percentage of traffic
7. Feedback Loop → Convert online issues into new training data

There are two easily overlooked yet critically important points within this loop:

First, the necessity of regression evaluation. While enhancing a certain capability during post-training, it might inadvertently weaken others. Therefore, each iteration must include regression tests on existing capabilities to avoid a “whack-a-mole” effect where fixing one issue causes another.

Second, the normalization of cost assessment. Many teams focus on metrics during the training phase but only realize post-launch that inference costs exceed expectations. Treating cost as an evaluation dimension on par with accuracy ensures that capability improvements remain “sustainable.”

5. Implications for Developers and Enterprises

In summary, the pathway of unlocking ultra-large-scale parameter potential through post-training offers the following insights for technical teams:

• Do not blindly pursue a larger base model: Building upon an existing large-scale foundation, the ROI of post-training might be significantly higher.
• Align evaluations with real-world tasks: Public benchmark scores serve as a reference, but custom evaluation sets tailored to business scenarios better reflect actual value.
• Optimize cost and capability simultaneously: Use the “total cost to complete a task” as the core metric, rather than the per-token price.
• Treat post-training as continuous engineering: Establishing a closed loop of data, training, evaluation, and feedback is the only way to ensure model capabilities evolve alongside the business.

The value of an ultra-large-scale parameter model lies not just in “how big” it is, but in whether we can efficiently, stably, and economically unleash its latent capabilities through post-training. For developers and enterprises, this is both a technical challenge and one of the most certain pathways to capability enhancement at the current stage.