LLM Post-Training: Engineering Practices in Finance and 3D Modeling
As large language model competition shifts from pre-training scale to post-training refinement, domain-specific reinforcement learning has become the key differentiator. This article explores the engineering logic behind post-training enhancements, detailing technical practices in finance and 3D modeling. It covers multi-domain joint reinforcement learning, verifiable reward signals, hallucination governance in financial workflows, spatial understanding and tool-calling in 3D generation, and strategies for building robust evaluation systems.
Introduction: From “Base Model Competition” to “Post-Training Refinement”
Over the past two years, the competitive focus of large language models (LLMs) has experienced a clear shift. In the early stages, the industry focused on pre-training scale, parameter counts, and base capability leaderboards. As base model capabilities converge, however, an increasing number of teams have realized that with targeted post-training, the same base model can achieve significantly different performance in professional tasks. A common characteristic of recent new model releases is that while the base model remains unchanged, the primary improvements stem from a massive increase in post-training compute and joint reinforcement learning across downstream domains.
This article does not discuss specific products but rather deconstructs the engineering logic behind the “post-training enhancement” trend from a technical methodology perspective. Why has post-training become the critical leverage for professional scenario implementation? What unique requirements do high-value scenarios like finance and 3D modeling impose on models? And how should technical teams design evaluation and optimization closed loops?
1. Why Post-Training is the New Capability Differentiator
1.1 Division of Labor Between Pre-Training and Post-Training
Pre-training solves the “general knowledge” problem: language modeling, world knowledge, and basic reasoning. However, the requirements of professional working scenarios extend far beyond this—they require the model to achieve format compliance, logical traceability, and controllable error rates within specific domains. These capabilities rarely emerge naturally from pre-training corpora alone; they require targeted shaping through Supervised Fine-Tuning (SFT) and Reinforcement Learning from Human or Programmatic Feedback (RLHF/RLVR).
Industry practices indicate that once a base model’s general capabilities reach a certain level, the marginal returns of continuing to stack pre-training compute diminish. Conversely, investing the same compute into post-training—especially joint reinforcement learning covering hundreds of downstream domains—often yields a steeper improvement curve in professional tasks.
1.2 Engineering Essentials of Multi-Domain Joint Reinforcement Learning
The core of “joint reinforcement learning across hundreds of downstream domains” is not simply mixing data, but solving three engineering challenges:
- Verifiability of Reward Signals: Scenarios like coding, math, and financial calculations have programmatically verifiable correct answers, making them suitable for training with Verifiable Rewards (RLVR). Open-ended tasks, however, rely on preference models. These two types of signals must be proportioned correctly to prevent the model from becoming “lopsided.”
- Inter-Domain Interference Suppression: Joint multi-task training is prone to gradient conflicts. Engineering solutions typically involve task-grouped sampling, staged curriculum training, and setting differentiated learning rates and sampling weights for different tasks.
- General Capability Anti-Regression: Professional reinforcement training can damage a model’s general conversational and instruction-following abilities. It is necessary to retain a certain proportion of general data as an “anchor” and continuously perform regression validation on general evaluation sets.
2. Finance Scenario: From “Answering” to “Trustworthy”
Finance is a benchmark scenario for professional model implementation due to its extremely low fault tolerance, strict compliance requirements, and high manual review costs. Recent industry data shows that models undergoing specialized financial post-training can reduce hallucination rates in internal tests from around 23% to approximately 8%. An improvement of this magnitude directly determines whether a model can enter real production workflows.
2.1 Three Specialized Optimization Directions for Financial Workflows
Cross-File Verification: Real financial research often involves cross-comparing multiple documents such as annual reports, announcements, and audit reports. The model must possess the ability to align entities and check numerical consistency across documents. In engineering implementation, long documents are typically segmented into structured chunks with metadata. Combined with Retrieval-Augmented Generation (RAG), the model explicitly annotates the source document and page number when citing. During training, “whether the citation is correct and the value is consistent” serves as the verifiable reward.
Financial Modeling: Financial modeling requires the model not only to read statements but also to generate structured calculation processes. A typical training method involves having the model output intermediate steps with formula annotations (e.g., presenting assumptions, formulas, and results in a table format), then using a program to execute and verify the final result, forming a closed-loop reward signal. Below is a pseudo-code representation of the verification process:
def verify_financial_model_output(response, ground_truth):# Parse the structured calculation steps from the model outputsteps = parse_steps(response)reward = 0.0for step in steps:if not step.has_citation: # Each step must have a source citationreward -= 0.2if not recompute(step): # Programmatic recomputation verificationreward -= 0.5if abs(steps[-1].value - ground_truth) < TOLERANCE:reward += 1.0 # Positive reward for correct final resultreturn reward
Traceable Research: Financial compliance requires every conclusion to be traceable. The post-training phase needs to reinforce an output paradigm of “cite first, conclude later,” meaning the model must provide evidence fragments before generating judgments, penalizing unsupported inferences. This is also one of the primary sources of hallucination reduction—hallucinations often occur when models are allowed to “freely improvise,” and mandatory citation constraints fundamentally compress the space for fabrication.
2.2 Cost Structure Changes Brought by Hallucination Governance
Reducing the hallucination rate from approximately 23% to 8% yields step-change commercial value rather than linear value: once the error rate falls below the acceptable threshold for manual review, the review process can be downgraded from “100% line-by-line inspection” to “sampling quality checks,” typically reducing labor costs by an order of magnitude. This is why the finance scenario is generally expected to commercialize faster than anticipated—the inflection point for return on investment has already arrived.
3. 3D Modeling Scenario: Combining Spatial Understanding, Programming, and Professional Software
3D modeling is a relatively new expansion direction in the LLM capability landscape. Unlike pure text tasks, it requires the model to simultaneously possess three capabilities: multimodal understanding, spatial reasoning, and tool calling.
3.1 Technical Connotation of Spatial Understanding
Qualified 3D generation capability first requires the model to infer proportions and spatial relationships from text descriptions and reference images, such as “the table height is about 1.5 times the seat height of the chair” or “the camera is located 45 degrees to the front right of the object.” Training for these capabilities relies on constructing large amounts of synthetic data with spatial annotations: by programmatically generating scenes and rendering multi-view images, “text-image-spatial parameter” triplets are naturally obtained as the data foundation for supervised and reinforcement learning.
3.2 The “Model as Programmer” Tool-Calling Paradigm
A key insight into the implementation of 3D modeling is that the model does not need to directly output 3D meshes; instead, it generates scripts that drive professional modeling software. Mature modeling software generally provides scripting interfaces. The model can generate parametric modeling code, which the software executes to render results. The render is then passed back to the model for visual self-checking, forming a closed loop of “generate-execute-feedback-revise”:
def modeling_loop(prompt, reference_image, max_iters=5):scene_params = model.infer(prompt, reference_image)for i in range(max_iters):script = model.generate_script(scene_params) # Generate modeling scriptrender = execute_in_3d_engine(script) # Execute in the rendering enginefeedback = model.visual_check(render, prompt) # Visual comparison self-checkif feedback.is_satisfactory:breakscene_params = model.revise(scene_params, feedback)return script
The advantage of this paradigm is that scripts are executable, verifiable, and reproducible, making them naturally suitable as reward carriers for reinforcement learning (the consistency between the render result and the target can be quantitatively evaluated). Furthermore, the output is a standard format file that can directly enter existing design production pipelines without requiring downstream workflow modifications.
4. Evaluation Systems: The Yardstick of Professional Capability
Post-training effectiveness must be demonstrated through evaluation. The current industry practice is to build comprehensive evaluation sets covering dimensions such as coding, agent tasks, financial research, and 3D generation, and compare them horizontally with frontier models. Technical teams building their own evaluation systems should note:
- Task Authenticity: Prioritize desensitized real business samples over textbook-style questions to prevent a disconnect between evaluation and production.
- Process Scoring: Evaluate not only the final answer but also score the citation correctness and calculation accuracy of intermediate steps, echoing the reward design in post-training.
- Cost Alignment: Evaluation should simultaneously record inference costs. The competitiveness of a professional model lies not only in its score but also in its price at equivalent performance—there are already industry cases achieving near-frontier-level scores at approximately one-fifth of the cost. Cost-effectiveness itself is a core competency.
5. Trend Forecast: Post-Training Drives Models into High-Value Workflows
Overall, the implementation path of general LLMs is shifting from “one model answers all questions” to a combined format of “base model + domain post-training + toolchain integration.” Scenarios with extremely high requirements for accuracy, compliance, and traceability—such as finance, 3D modeling, cybersecurity, and law—are becoming the first areas to realize the value of post-training. For enterprise technical teams, the actionable insights are:
- When evaluating base models, include “post-trainability” (whether it supports efficient fine-tuning, whether it provides domain evaluation baselines) as a selection metric.
- In professional scenarios, prioritize building verifiable reward and citation constraint mechanisms; this is the leverage for reducing hallucination rates.
- Connect existing production toolchains through “model generates scripts + professional software executes” to lower integration costs.
- Redesign human-machine collaboration workflows by replacing full reviews with sampling quality checks to realize cost-reduction benefits.
Post-training is no longer a concluding step in the training pipeline; it is the main battlefield that determines the success of professional scenarios. Whoever builds a deeper engineering moat in reward design, data construction, and evaluation closed loops will be the first to transform general intelligence into measurable business value.
