How Post-Training Enables Financial Analysis and 3D Modeling

    As foundation models become increasingly similar in general capabilities, post-training is becoming the main way to improve performance on specialized tasks. This article explains how supervised fine-tuning, preference alignment, and reinforcement learning can support financial analysis and 3D modeling, with a focus on data balance, multi-domain optimization, evaluation, verification, cost control, and human oversight.

    1. Background: As General Capabilities Converge, Differentiation Moves to Post-Training

    For a period of time, pre-training determined a model’s knowledge coverage and basic reasoning ability. As foundation models become more capable and increasingly similar, product differentiation is shifting toward specialized tasks: understanding a complete financial statement, correctly interpreting object proportions in three-dimensional space, and explaining the sources behind a conclusion.

    These capabilities usually cannot be achieved simply by adding more pre-training data. They need to be developed after pre-training through post-training.

    Post-training generally includes supervised fine-tuning, preference alignment, and reinforcement learning. Its core purpose is to use high-quality task samples and quantifiable feedback signals to guide the model toward more accurate and consistent outputs in a target domain. This is particularly valuable in highly specialized fields such as finance and 3D modeling, where errors can be costly and tolerance for inaccuracy is low.

    2. Three Key Elements of Post-Training

    2.1 Compute Investment and Data Allocation

    Post-training performance is directly related to compute investment, but simply increasing compute is not enough. Data allocation determines the model’s final capability profile.

    If the training data is concentrated too heavily on one type of task, the model may improve on that task while degrading on others. This is commonly known as catastrophic forgetting. Teams should define the sample ratio for each domain at every training stage and regularly use held-out datasets to monitor changes in each capability.

    2.2 Multi-Domain Reinforcement Learning

    Multi-domain reinforcement learning places the reward signals of multiple downstream tasks into a single training process, allowing the model to optimize several capabilities at the same time.

    This approach enables capabilities to transfer across domains. For example, code understanding can support structured data processing, while logical reasoning can help verify financial formulas. During implementation, each domain needs a clear reward function so that rewards do not conflict with one another.

    2.3 Evaluation

    Evaluation is the foundation of the post-training feedback loop. Evaluation sets should be derived from real business scenarios, cover both common tasks and edge cases, and be reported separately by domain.

    A single composite score can hide important problems. Teams should also monitor metrics such as accuracy, refusal rate, format compliance, and response latency.

    3. Financial Applications: From Usable to Verifiable

    Financial work involves large volumes of numerical data, information distributed across multiple sources, and conclusions that must be traceable. For models, three capabilities are especially important.

    3.1 Cross-Document Verification

    The same metric may appear in an annual report, quarterly report, public filing, or internal spreadsheet. Cross-document verification requires the model to identify the same metric across sources, compare whether the values are consistent, and highlight any discrepancies.

    A typical workflow is as follows:

    1. Input: Multiple documents and a conclusion to verify
    2. Steps:
    3. 1. Extract the target metric and its definition from each document, including currency, unit, and time range
    4. 2. Normalize the definitions and compare the values
    5. 3. If a discrepancy exists, record the source location and the size of the difference
    6. 4. Output the verification result and the paragraph number supporting each item
    7. Output: Verification result, discrepancy list, and evidence index

    3.2 Financial Modeling

    Financial modeling involves multiple stages, including revenue assumptions, cost structures, discount rates, and sensitivity analysis.

    The model must understand the meaning of each assumption, organize calculations using the correct formulas, and explain how results change when key assumptions are modified. During post-training, teams can add extensive formula-derivation and numerical-recalculation samples. This encourages the model to follow a consistent process:

    1. List the assumptions.
    2. Perform the calculations step by step.
    3. Validate the result.

    3.3 Traceable Research

    Traceable research requires every conclusion to be linked back to the original material. In practice, teams can define an output format in which each conclusion is immediately followed by a citation identifier. The identifier maps to a specific location in the source document.

    This allows reviewers to locate the evidence quickly and determine whether the conclusion is supported.

    3.4 Common Methods for Controlling Hallucinations

    A hallucination is content generated by a model that does not match the facts. In professional applications, hallucinations can be costly, so mitigation requires multiple layers of control:

    • Retrieval augmentation: Retrieve information from trusted data sources before answering, reducing unsupported generation.
    • Citation validation: Require every numerical value and judgment to include a source. A program can then check whether the source exists and whether the cited value is consistent.
    • Numerical recalculation: Recalculate the model’s results with deterministic code and compare the outputs.
    • Confidence-based routing: Send low-confidence outputs for human review instead of delivering them directly.

    The following pseudocode illustrates the basic idea of numerical recalculation:

    1. function check_numeric_claim(claim, source_table):
    2. value_in_claim = parse_number(claim.value)
    3. value_in_source = lookup(source_table, claim.metric, claim.period)
    4. if value_in_source is missing:
    5. return "Cannot verify; route for human review"
    6. if abs(value_in_claim - value_in_source) > tolerance:
    7. return "Value mismatch; record discrepancy"
    8. return "Verification passed"

    4. 3D Modeling: Spatial Understanding and Tool Coordination

    3D modeling requires language understanding, spatial reasoning, and interaction with professional software. Unlike purely text-based tasks, its outputs can be rendered, measured, and compared, making it well suited to closed-loop optimization.

    4.1 Spatial Understanding

    The model must infer an object’s proportions, orientation, and relative position from text descriptions and reference images.

    For example, given a description such as “the tabletop is approximately twice as wide as the chair, and the chair back is level with the tabletop,” the model should infer approximate dimensional relationships and preserve them during modeling.

    4.2 Coordination with Modeling Software

    A mature approach is to have the model generate structured modeling instructions that are executed by the modeling software.

    These instructions typically include object types, dimensional parameters, position coordinates, and material settings. The model should produce parseable structured content rather than free-form text, making it easier for programs to validate and execute.

    4.3 A Generate–Render–Compare Loop

    Quality control for 3D modeling can follow this loop:

    1. Start loop:
    2. 1. Generate modeling instructions based on the requirements
    3. 2. Execute the instructions in a sandbox and render a preview
    4. 3. Compare the preview with the reference image or dimensional constraints
    5. 4. If the deviation exceeds the threshold, send the deviation information back to the model and regenerate
    6. 5. Stop when the iteration limit is reached or the comparison passes

    An iteration limit is important. It controls cost and execution time and prevents the model from repeatedly exploring an incorrect direction.

    5. Key Metrics for Production Deployment

    Whether a technical solution can enter production depends on whether its performance can be measured. The following metrics are recommended:

    • Error rate: Sample tasks for manual annotation and measure the proportion of factual, numerical, and citation errors. The sample size and sampling method should be documented to avoid reporting an isolated number without context.
    • Review cost: Record the average human review time required for each task to determine whether automation actually reduces workload.
    • End-to-end latency: Include the time required for retrieval, generation, validation, and rendering.
    • Per-task cost: Combine compute, storage, and human labor costs and compare them with the existing workflow.

    The definition of error rate must be established at the beginning of the project and kept consistent throughout the evaluation period. Changes in measurement criteria can make before-and-after comparisons meaningless.

    6. Risks and Boundaries

    Model applications in professional settings require clearly defined boundaries.

    First, data compliance must be addressed. Financial data and internal enterprise materials may contain confidential or sensitive information. Both training and inference environments should follow the principle of least privilege, and access records should be auditable.

    Second, human review checkpoints are necessary. Model outputs should be confirmed by an accountable person before they are used for decision-making. Content related to investment judgments, financial disclosures, or design deliverables should not be fully automated without appropriate oversight.

    Third, teams must recognize the model’s capability boundaries. Models are well suited to structured and verifiable subtasks, while problems requiring long-term experience or industry judgment should remain led by human experts.

    7. Practical Recommendations

    1. Start with one high-value, verifiable subtask, such as cross-document numerical verification, and expand gradually.
    2. Build an evaluation set and record baseline metrics before the pilot begins.
    3. Apply deterministic recalculation to all numerical outputs instead of relying on the model’s own judgment.
    4. Require every conclusion to include a source and incorporate source validation into the automated workflow.
    5. For 3D modeling tasks, build the rendering and comparison loop as early as possible to reduce repeated manual revisions.
    6. Regularly monitor capability changes across domains to ensure that optimization in one scenario does not damage performance in others.

    Post-training provides a practical path for applying foundation models to specialized scenarios. The key is to break tasks into verifiable steps, measure performance continuously with quantifiable metrics, and retain necessary human review throughout the process.