Post-Training Reinforcement for Finance and 3D: A Reliable Deployment Path
Post-training has become a focal point as base models converge in general capability. This article examines how post-training and reinforcement learning can be applied to finance workflows and 3D scene generation. It covers reducing hallucinations in financial analysis through cross-document verification, financial modeling, and traceable research; designing reward signals and evaluation sets; and converting natural language into executable 3D modeling scripts within a sandboxed workflow. It also outlines model selection and evaluation practices, including building task-specific benchmarks and preserving human review for high-stakes outputs.
1. Why Post-Training Matters
Over the past two years, large-model iteration has centered mainly on pre-training, with competition focused on parameter scale, data volume, and context length. As the general capabilities of mainstream base models gradually converge, the center of competition is shifting to post-training. Post-training means keeping the base model largely unchanged while using supervised fine-tuning (SFT, teaching the model to answer as required with labeled examples), reinforcement learning (RLHF and RLAIF, adjusting output preferences with feedback signals), and domain joint training to improve performance on specialized professional tasks.
Post-training is attractive for three main reasons: compute investment is concentrated on downstream tasks, so large-scale pre-training does not need to be redone; the collection, training, and evaluation loop for domain data can often be completed within weeks; and results can be quantified with vertical evaluation sets. It is not a universal solution, however. If domain data quality is poor, or evaluation sets are disconnected from real business work, the trained model will only appear better in tests.
2. Finance Workflows: Fewer Errors Matter More Than Polished Writing
Financial analysis has a very low tolerance for error. A wrong revenue figure or date can lead to a wrong investment judgment. The optimization goal in finance is therefore to make the model make fewer errors and let people verify results quickly, not to make its writing more fluent.
2.1 Three Core Capabilities
The table below lists the inputs, outputs, and acceptance criteria for three capabilities. Teams can use it directly for internal evaluation.
| Capability | Typical input | Expected output | Acceptance criteria | Human review point |
|---|---|---|---|---|
| Cross-document verification | Annual reports, notes, broker research | Comparison conclusions for the same metric across multiple documents | Values match the source text item by item; conflicts are explicitly marked | Metrics with inconsistent definitions |
| Financial modeling | Three financial statements and specified assumptions | Valuation table and formulas for DCF or the comparable company method | Three-statement articulation balances; formulas can be recalculated | Key assumptions, such as growth rate and discount rate |
| Traceable research | Research question and document library | Conclusions with file name, page number, and paragraph | Every conclusion can be located in the source text | Missing citations or mislocated items |
Cross-document verification requires the model to align information across multiple documents, not merely understand one document. Financial modeling requires the model to understand the relationships among the income statement, balance sheet, and cash flow statement and to keep formula logic self-consistent. Traceable research requires every conclusion to include a source so analysts can check it.
2.2 Basic Pipeline for Hallucination Governance
Reductions in hallucination rate usually come from a combination of multiple techniques:
Hallucination governance pipeline (illustrative)Domain SFT (financial reports and research corpus)→ RL stage with factual consistency reward→ Inference-time requirement to cite retrieved evidence→ Post-output validation (numbers, dates, definition consistency)
In this pipeline, reinforcement learning reward design is the most critical part. If the model still gives a definite answer when evidence is insufficient, it should receive negative feedback. Answering “insufficient evidence, cannot determine” should receive positive incentive. This goal of “rather not answer than fabricate” is the main difference between professional scenarios and general chat scenarios in training objectives.
2.3 How to Measure Hallucination Rate
Most hallucination-rate numbers circulating online lack test conditions. A credible number must at least specify five items: the source and size of the test set, the criteria for judging an error, the baseline model, the test time, and whether annotation was done by one person or multiple people. Comparison numbers missing this information cannot be used directly for procurement decisions.
Enterprises should build evaluation sets from their own real business questions, test the baseline model first, then test the post-trained model, and compare the difference between the two on the same test set. Only a reduction obtained this way has reference value.
2.4 Deployment Value and Boundaries
If validation shows that the hallucination rate has indeed dropped significantly, the manual review workload will also decrease. At that point, the model is suitable as a junior analysis assistant that requires review, not as a tool that can directly publish conclusions externally. Outputs involving investment decisions and compliance disclosure still require final confirmation by qualified personnel.
3. 3D Modeling: From Natural Language to Executable Scripts
3D scenes place different demands on models and require a combination of three capabilities.
- Spatial understanding: based on text descriptions and reference images, determine object proportions, orientation, and position relationships. For example, for “seat height is 1.2 times the cabin width,” the model needs to understand both semantic constraints and geometric constraints.
- Procedural generation: the model outputs executable modeling scripts or parameters, rather than geometry data directly. Scripts are easier for humans to read and modify, and easier to version-control.
- Workflow embedding: scripts are executed through software interfaces, results are returned to the model for self-check, and designers review at key nodes.
The pseudocode below illustrates the input and output relationships of each step. It cannot run directly; actual interfaces need to be implemented according to the documentation of the software being used.
# Pseudocode: natural-language-driven modeling pipeline# Input: user description string request# Output: scene object scene, or error messageplan = llm.parse(request) # Parse into structured constraints; return clarification questions on failureif not plan.is_valid():return ask_user(plan.missing_fields) # Do not generate a script when information is incompletescript = llm.codegen(plan, target="blender_api")result = sandbox_execute(script) # Execute only in an isolated environment; no external resource accessif result.error:return retry_or_escalate(result.error, max_retries=2)review = llm.check(result.scene, constraints=plan.constraints) # Spatial consistency self-checkif not review.passed:return flag_for_human_review(review.issues)
Before actual deployment, note the following:
- The execution environment must be a sandbox; scripts can only access specified scene files and model libraries.
- Back up scene files before modification and keep rollback versions.
- Apply the principle of least privilege to API permissions, for example allowing only object creation and modification, not deletion or filesystem writes.
- A passed self-check does not mean the modeling is correct; critical dimensions still need human measurement or review.
This type of training requires multimodal data, namely combinations of images, text, and 3D annotations, as well as code execution results as reward signals. It is more difficult than pure text tasks. Game assets, industrial design, and architectural visualization have real demand for these capabilities, but deployment requires complete testing and permission design.
4. Selection and Evaluation: How to Do It and How to Choose
How should post-training reinforcement be done? First define a concrete task, collect real examples for that task, build an evaluation set, and then decide whether to perform SFT or RL training. Starting training without an evaluation set usually makes it impossible to judge whether the effect has really improved.
How should a finance model be selected? Do not rely only on general leaderboards. Test with your own document library and question set, and focus on three metrics: error rate, citation accuracy and traceability, and integration cost with existing systems.
What should be noted when executing 3D modeling scripts? Run only in a sandbox, back up first, limit API permissions, and validate output dimensions and constraints. Any deletion or overwrite operation requires human confirmation.
For developers, domain data construction, reward design, RL training pipelines, and output validation engineering are the most practically valuable skills in this direction. For enterprise users, real test results in professional scenarios should come first; general leaderboard rankings can only serve as a reference. Model outputs still require human review, especially results involving capital, compliance, and physical safety.
5. Conclusion
Post-training reinforcement is moving large models from general chat tools toward collaborators in professional work. Finance and 3D are two typical scenarios, and their common difficulty is accuracy and verifiability. Whether deployment succeeds depends on whether teams can establish their own evaluation standards and keep human gatekeeping at key steps.
