Post-Training Reinforcement for LLMs: Going Deeper into Finance and 3D Modeling
This article explains post-training for large language models and why it matters for enterprises adapting general-purpose models to specialized domains such as finance and 3D modeling. It clarifies when post-training is worth doing, walks through a three-layer method from supervised fine-tuning to multi-domain joint reinforcement learning with concrete sampling weights and reward priorities, and shows how to apply it to financial workflows that reduce hallucination and meet compliance. It also covers procedural 3D modeling with tool-calling safeguards, a five-metric framework for evaluating post-trained models beyond public leaderboards, and enterprise best practices for scenario selection, regression testing, retrieval-augmented generation, and model-plus-tools orchestration.
1. What Is Post-Training and When Should You Choose It?
As general-purpose LLM capabilities converge, post-training is becoming a key lever for enterprises that want to go deeper into specialized scenarios such as finance and 3D modeling. Post-training refers to making a model produce more reliable, usable outputs on a specific task distribution after its base pre-training is complete, through supervised fine-tuning, preference alignment, and reinforcement learning. It does not redo pre-training; instead, it invests limited compute in the stages that most influence business results.
When is post-training worth doing? Its value becomes clearer when a task has high-frequency inputs and outputs, relatively clear evaluation criteria, data that can be desensitized, and outputs that feed into a real production workflow. If a scenario has no accumulated data, unclear acceptance criteria, or a strong need for one-off exploration, post-training too hastily tends to mean high cost and low return. Common mistakes include: only fine-tuning the output format without any factual constraints, chasing leaderboard scores while ignoring scenario-specific evaluation, and going live without human review or permission isolation.
2. How to Do Post-Training Reinforcement: From SFT to Joint Reinforcement Learning
Post-training typically has three layers. The first is supervised fine-tuning (SFT), which uses high-quality instruction samples to teach the model the task format. The second is preference alignment, which uses rule-based rewards or a reward model to make outputs more accurate and more reliable. The third is multi-domain joint reinforcement learning, which keeps the model balanced across general, finance, coding, and 3D modeling directions.
The key to multi-domain joint training is not simply mixing data, but designing controllable sampling, reward, and evaluation mechanisms. Example parameters could be: a sampling weight of 40% for general capability, 30% for finance, 20% for coding, and 10% for 3D modeling; each step samples tasks according to these weights, while retaining 10%–20% general replay data to reduce the risk of catastrophic forgetting. It is recommended to trigger an evaluation every 500–1,000 steps, covering fixed test sets for general, finance, 3D modeling, and other domains. Reward design needs clear priorities: compliance and factual accuracy rank above output completeness, and output completeness ranks above language style; when the rules of two domains conflict, the stricter domain constraint takes precedence.
3. Financial Workflows: Reducing Hallucination and Meeting Compliance
Financial analysis is a high-value, high-rigor scenario and a key direction for post-training. The core capabilities include cross-file verification, financial modeling, and traceable research. Cross-file verification requires the model to cross-check figures, dates, and reporting bases across financial statements, announcements, and research reports. Financial modeling requires understanding the reconciliation relationships between statements. Traceable research requires that every conclusion can be traced back to its source.
Under a fixed evaluation set, given material constraints, and explicit citation requirements, post-training can measurably reduce the hallucination rate, but the magnitude of the reduction depends on data quality, how strict the rules are, and the human review process. An executable minimal flow for financial modeling is: retrieve data, run reconciliation checks, input assumptions, produce formula output, cite evidence, and complete human review. The model should not deliver unverifiable conclusions directly; instead, when materials are missing, it should flag “insufficient information” and suggest what supplementary material to add. Financial scenarios must also set up permission isolation and human review, especially when the work involves investment advice, credit decisions, and regulatory reports.
4. 3D Modeling: Spatial Understanding Plus Procedural Generation
The focus of 3D modeling is to extend language understanding to spatial relationships and tool operations. The model needs to judge the proportions, positions, and relative relationships of objects from text or reference images, then generate parametric scripts or call professional software. Compared with dragging points one by one, procedural modeling is more reproducible, easier to modify, and easier to version-control.
The following is a clearer table-modeling example. Coordinate system convention: the X axis runs along the length of the tabletop, the Y axis along its width, and the Z axis points vertically upward for height, in meters; positions represent the center-point coordinates of each box, with the tabletop center located above the origin.
# Scene coordinate system: X is the length direction, Y is the width direction, Z is the height direction; unit is meters.spec = {"object": "table","length": 1.2,"width": 0.6,"height": 0.75,"top_thickness": 0.05,"leg_thickness": 0.05}def build_table(scene, spec):# Read fields explicitly to avoid relying on dict orderinglength = spec["length"]width = spec["width"]height = spec["height"]top_thickness = spec["top_thickness"]leg_thickness = spec["leg_thickness"]# Total table height is 0.75m and tabletop thickness is 0.05m, so the tabletop center Z coordinate is the height minus half the tabletop thicknesstabletop_z = height - top_thickness / 2scene.add_box("tabletop", size=(length, width, top_thickness), position=(0, 0, tabletop_z))# Leg height equals the total height minus the tabletop thickness; the bottom touches the ground, so the center Z coordinate is half the leg heightleg_height = height - top_thicknessleg_z = leg_height / 2# Leg centers are inset by half a leg thickness toward the tabletop interior to avoid the legs extending beyond the tabletop boundaryfor x in (-length / 2 + leg_thickness / 2, length / 2 - leg_thickness / 2):for y in (-width / 2 + leg_thickness / 2, width / 2 - leg_thickness / 2):scene.add_box("leg", size=(leg_thickness, leg_thickness, leg_height), position=(x, y, leg_z))
At the tool-calling stage, exception handling is also needed: if a parameter is out of range, it should be intercepted and a correction prompted; if units are inconsistent, the value should be converted automatically or the execution refused; if the modeling software fails, it can be retried a limited number of times with logs preserved; if the generated result affects an existing design, version rollback should be supported. Only this way can the model move from generating text to generating executable results.
5. How to Evaluate Post-Trained Models? Leaderboards vs. Scenario-Specific Metrics
To judge whether a model is truly usable, you cannot rely only on public leaderboards. Leaderboards reflect general capability, but enterprises need a dedicated evaluation set: real desensitized data, fixed samples, scoring rules, and a regression process.
It is recommended to track at least five metrics. Hallucination rate: the number of conclusions that cannot be supported by the given material, or that contradict it, divided by the number of valid conclusions; for example, if 80 of 1,000 conclusions lack source support, the hallucination rate is 8%. Accuracy: the number of correct conclusions divided by the number of valid questions. Citation coverage: the number of conclusions requiring citation that have a locatable source, divided by the total number of conclusions that require citation. Cost per task: the sum of input token cost, output token cost, tool-call cost, and storage cost. Response latency: measured at P50 and P95, avoiding averages alone.
Cost example: if the cost per task is RMB 0.02 and the monthly call volume is 1 million, the base monthly inference cost is about RMB 20,000. Introducing tool calls and storage adds further cost. Cost-effectiveness should be assessed as “cost per task × call volume,” while also setting thresholds for accuracy and hallucination rate, rather than looking at low price alone.
6. Enterprise Best Practices: Scenario Selection and Implementation Path
When choosing scenarios, enterprises can use a five-dimension scoring sheet: frequency, value, verifiability, data availability, and compliance risk. Each is scored from 1 to 5, and scenarios with high scores and manageable compliance risk are prioritized. For example, financial document verification, report generation, parametric modeling, and operations Q&A can all first enter small-traffic validation.
For implementation, it is recommended to build a dedicated evaluation and regression system, with a fixed evaluation set that runs automatically to prevent quality regressions caused by prompt or model-version iterations. At the same time, use retrieval-augmented generation (RAG) and source citation to improve traceability, and treat “every conclusion must have evidence” as a hard constraint. In terms of engineering architecture, adopt a “model plus tools” orchestration approach, using function calling to connect to professional software and internal systems, and use logging, monitoring, and alerting to record each input, output, latency, and exception.
For infrastructure, enterprises can rely on the compute, storage, message queue, container platform, and logging/monitoring capabilities commonly offered by mainstream cloud platforms to build a stable, scalable application pipeline, rather than rebuilding the underlying infrastructure.
7. Conclusion
As base-model capabilities converge, the focus of competition is shifting from “whose base model is bigger” to “whose post-training is more solid and who understands the business scenario better.” For developers and enterprises, post-training is not a one-time project but an engineering process of continuous evaluation, hallucination control, and cost optimization around real scenarios. Only by combining post-training, data, tools, permissions, and evaluation can LLMs genuinely enter high-value scenarios such as finance and 3D modeling and convert into stable, measurable business returns.
