Post-Training Advances in Chinese LLMs: From Code Generation to 3D Modeling

Introduction: The Rise of the Post-Training Paradigm

At the current stage of LLM evolution, the marginal returns from scaling pre-training are diminishing, while post-training — the fine-grained polishing of model capabilities — has become a key factor that separates models in real-world performance. Post-training generally refers to the complete process of targeted capability enhancement after a base model has finished large-scale pre-training, using supervised fine-tuning, reinforcement learning, preference alignment, domain data injection, and similar techniques.

This article uses the latest version of a Chinese LLM as a technical analysis sample to systematically review how post-training is applied in code generation, 3D modeling, financial analysis, and hallucination suppression, along with the capability boundaries involved. Note that this article focuses on the technical approaches themselves and does not constitute any product recommendation or commercial promotion.

1. Coding Tasks: From Terminal Operations to Complex Engineering Reasoning

1.1 Why Terminal-Based Coding Tasks Are Hard

Terminal-based coding tasks are an important yardstick for measuring a model’s engineering ability. Unlike traditional function-level code completion, terminal tasks require the model to understand the full command-line interaction context, including file system operations, dependency management, build processes, and error diagnosis and repair. The core difficulties are:

• State awareness: the model must maintain an accurate understanding of the current working directory, environment variables, and command history;
• Multi-step reasoning: a complete terminal task often involves orchestrating dozens of commands, and an error at any step can break the entire chain;
• Error recovery: when facing compilation errors, runtime exceptions, or permission issues, the model needs to diagnose and roll back on its own.

1.2 Capability Gains from Post-Training

After systematic post-training optimization, the model’s performance on terminal coding benchmarks improved significantly. According to publicly available technical information, its score jumped from a relatively low level in the previous generation to around 65.7, leading among models of the same tier and even surpassing some leading international models on similar tasks.

Behind this improvement are typically the following post-training strategies:

1. Trajectory-level supervised fine-tuning: collecting large numbers of real terminal operation trajectories so the model learns complete task execution sequences;
2. Process reward modeling: rewarding not only final results but also providing fine-grained evaluation of intermediate steps;
3. Interactive reinforcement learning: letting the model repeatedly explore in simulated terminal environments and optimize its policy through environmental feedback.

1.3 Engineering Wisdom in Cost Control

Notably, alongside the performance gains, the model also performs well in cost control. Among frontier models, its unit inference cost on terminal tasks sits in the lowest tier. This usually comes down to:

• Inference path optimization: post-training teaches the model shorter, more direct solution paths, reducing wasted token consumption;
• Caching and reuse mechanisms: efficient retrieval and reuse strategies for common operation patterns;
• Dynamic compute allocation: adaptively adjusting inference resources based on task complexity.

For enterprise users, this means that in scenarios such as batch code review and automated operations, higher-quality output can be obtained within an acceptable cost budget.

2. 3D Modeling and Spatial Reasoning: From Text to Geometry

2.1 New Capability Requirements for 3D Creation

3D modeling tasks demand far more from a language model than text generation does. The model must understand professional concepts such as geometric topology, material properties, and spatial transformations, and translate that understanding into executable modeling scripts. Specifically:

• Spatial reasoning: understanding position, rotation, scaling, and Boolean operations in 3D space;
• Parametric modeling: converting natural language descriptions into precise geometric parameters and constraints;
• Tool script generation: writing executable scripts for professional modeling tools (parametric CAD, polygon modeling software, etc.).

2.2 Benchmark Performance

On CAD-related benchmarks, the model scored 92.7 with tool assistance, narrowing the gap with top international models to single digits — the highest level among current open-source models in this area. This result indicates that the systematic injection of 3D domain data during post-training played a key role.

2.3 Real-World Application Scenarios

In actual modeling scenarios, the model demonstrated the ability to drive professional modeling tools directly through scripts, completing the full workflow from scene assembly to detail refinement without relying on an intermediate protocol layer such as MCP. For example:

# Illustrative pseudocode: generate a modeling script from a natural language instruction
# User instruction: create an airport indoor forest scene with 20 trees, evenly distributed
import bpy
import random
# Clear the default scene
bpy.ops.object.select_all(action='SELECT')
bpy.ops.object.delete()
# Create the ground plane
bpy.ops.mesh.primitive_plane_add(size=50, location=(0, 0, 0))
# Batch-create trees
for i in range(20):
    x = random.uniform(-20, 20)
    y = random.uniform(-20, 20)
    # Trunk
    bpy.ops.mesh.primitive_cylinder_add(radius=0.3, depth=3, location=(x, y, 1.5))
    # Canopy
    bpy.ops.mesh.primitive_uv_sphere_add(radius=1.5, location=(x, y, 4))
# Adjust materials and lighting
# ...

Given a natural language description, the model can automatically generate structured modeling scripts like the one above, and supports subsequent iterations such as layout adjustments, material replacement, and geometry optimization. For developers in architectural visualization, game asset creation, and industrial design, this means more time for creative design and less time on repetitive modeling work.

3. Finance and Office Scenarios: Deep Adaptation for Vertical Domains

3.1 Improved Office QA Capability

On an office QA benchmark, the model scored 64.4, leading among models in the same batch. Office QA tasks typically involve compound abilities such as document understanding, table reasoning, and multi-hop QA, placing high demands on context comprehension and information integration.

3.2 Full Support for Financial Workflows

On financial domain benchmarks, the model covered the complete investment workflow from target screening to profitability analysis, scoring 58.6 — a significant improvement over the previous generation. Financial tasks are characterized by:

• Data sensitivity: precise handling of structured information such as numbers, dates, and ratios;
• Logical rigor: reasoning chains must be airtight, with no jumps or contradictions;
• Timeliness: some tasks require dynamic analysis combined with the latest market data.

During post-training, the introduction of high-quality financial instruction data and multi-step reasoning chain samples substantially strengthened the model’s performance on these tasks.

4. Hallucination Suppression: The Foundation of Trustworthy AI

4.1 The Nature of Hallucination

Hallucination refers to a model generating content that contradicts facts or is fabricated outright. In code generation, hallucination may appear as calls to non-existent APIs or incorrect parameter signatures; in 3D modeling, as scripts that cannot execute or operations that violate geometric constraints.

4.2 Stress Test Results

In an internal stress test set, the model’s error rate dropped sharply from 23.4% in the previous generation to 8.3%, leading among models of the same tier. This improvement typically stems from the following techniques:

1. Factuality-enhanced training: strengthening rewards for factual accuracy in fine-tuning data;
2. Retrieval-augmented generation: introducing external knowledge retrieval at inference time to reduce over-reliance on parametric knowledge;
3. Self-consistency checking: filtering low-confidence outputs through multiple sampling and voting;
4. Improved refusal mechanisms: teaching the model to say “I don’t know” when uncertain rather than fabricating an answer.

For enterprise applications, lower hallucination rates directly affect system reliability and user trust. Especially in low-tolerance scenarios such as code generation and financial analysis, hallucination suppression is often the deciding factor for production deployment.

5. Engineering Recommendations for Post-Training

Based on the analysis above, here are several suggestions for teams looking to improve model capabilities during post-training:

5.1 Data Quality Over Quantity

In post-training, data quality matters far more than quantity. A small amount of high-quality, human-verified domain data often delivers more significant capability gains than massive volumes of low-quality data.

5.2 Build Capabilities in Phases

Proceed in stages — “maintain general capabilities → strengthen core domains → cover edge cases” — to avoid degrading general capabilities by injecting too much domain data at once.

5.3 Establish a Multi-Dimensional Evaluation System

Beyond standard benchmarks, build internal evaluation sets that mirror real business scenarios, covering correctness, robustness, cost efficiency, and other dimensions.

5.4 Watch Inference Cost and Latency

Post-training should target not only performance gains but also inference efficiency. Techniques such as trajectory compression and caching strategies can significantly reduce serving costs while maintaining performance.

Conclusion

Post-training is becoming the core driver of differentiation among LLMs. From code generation to 3D modeling, from financial analysis to hallucination suppression, systematic post-training strategies can enable base models to achieve qualitative leaps in specific domains. For developers and enterprises, understanding the technical principles and engineering practices of post-training will help them better evaluate, select, and optimize LLM solutions for their own business scenarios.