All researchDomain adaptation / Financial reasoning

ALPHEVA RESEARCH / TECHNICAL CASE STUDY

Teaching Kimi K2.6
to Perform Financial Work

How Alpheva turned expert judgment, agent environments, and outcome-based training into specialized model capability.

FIG. 01Aggregate expert score1,200 held-out tasks · Higher is better
100%50%0
67.4%
Base modelKimi K2.6
75.8%
SFTExpert demonstrations
79.6%
+ PreferenceProfessional judgment
81.9%
Full pipelineEnvironments + outcomes

FROM GENERAL CAPABILITY
TO SPECIALIZED PERFORMANCE

+14.5 ptsAggregate expert score
8.73.1%Critical-error rate
See the full evaluation
Financial professionals
42
Expert-reviewed examples
8,400
Preference judgments
6,700
Agent trajectories
52,000

OVERVIEW

General capability does not automatically become professional competence. Financial work requires a model to interpret customer context, calculate accurately, recognize missing information, weigh interacting trade-offs, use tools correctly, and communicate uncertainty without losing the decision at the center of the task.

01 · Research question

Can expert work make a general model reliably better at a defined financial workload?

We worked with a U.S.-based financial vertical AI company whose identity is withheld at its request. The engagement used Kimi K2.6 as the open-weight base model and focused on the layer around the model: the system that converts expert judgment and real workflows into training and evaluation signals.

The goal was deliberately narrow. We were not trying to make the base model universally superior to frontier proprietary systems. We wanted to make it materially better at the financial reasoning and multi-step workflows the application actually performs.

H1

Demonstrations transfer reasoning patterns

Expert examples should improve how the model structures decisions in context.

H2

Preferences transfer professional judgment

Pairwise choices should distinguish merely plausible answers from useful ones.

H3

Environments improve execution

Realistic workflows should improve tool use, state management, and task completion.

H4

Signals are complementary

The combined pipeline should improve more dimensions than supervised tuning alone.

02 · The training system

Expert judgment became a training signal, not just a source of labels.

Tasks were built around realistic customer states: income, expenses, cash, investment holdings, retirement accounts, liabilities, tax circumstances, documents, objectives, and constraints. Experts identified the actual decision, material facts, missing information, calculations, alternatives, trade-offs, and the conclusion supported by the evidence.

FIG. 02A closed improvement loop
  1. 01Observe real work
  2. 02Generate expert tasks
  3. 03Train and roll out
  4. 04Verify outcomes
  5. 05Target failures
Expert data

Demonstrations

8,400 reviewed examples captured contextual reasoning, assumptions, calculations, and decision structure.

Judgment

Preferences

6,700 pairwise choices taught distinctions in relevance, actionability, risk awareness, and uncertainty.

Execution

Agent trajectories

Controlled environments recorded tools, documents, calculations, intermediate state, and final outcomes.

Measurement

Verification

Deterministic checks, structured rubrics, and expert adjudication matched the verifier to the task.

03 · Experimental design

Four configurations isolated the value of each training stage.

A fixed 1,200-task held-out set represented investment, tax, debt, retirement, insurance, cash-flow, and cross-domain decisions. No evaluation example was used directly in training. A separate 300-workflow agent evaluation tested end-to-end execution, while approximately 240 outputs received independent expert review.

Model A

Base

Kimi K2.6 with no financial-specific post-training.

Model B

SFT

Supervised fine-tuning on expert demonstrations.

Model C

SFT + preference

Supervised tuning followed by preference optimization.

Model D

Full pipeline

SFT, preferences, environment trajectories, and outcome-based optimization.

EVALUATIONWeighted expert rubric
  • Financial correctness30%
  • Contextual reasoning25%
  • Decision quality20%
  • Completeness15%
  • Communication10%

04 · Overall performance

The full pipeline improved every scored dimension and cut critical errors.

Supervised fine-tuning produced the first large improvement. Preference optimization added alignment with professional judgment. Environment trajectories and outcome-based training produced the strongest gains in decision quality and multi-step work.

TABLE 01Held-out evaluation across four model configurations
Evaluation dimensionBaseSFT+ PreferenceFull Alpheva
Aggregate expert score67.4%75.8%79.6%81.9%
Financial correctness74.8%80.6%84.3%88.1%
Contextual reasoning61.2%72.1%77.4%82.7%
Completeness64.7%72.9%77.1%80.4%
Decision quality60.9%70.4%75.9%79.8%
Expert preference54.8%68.3%73.9%78.6%
Critical-error rate8.7%5.4%4.0%3.1%

All values are percentages. Higher is better, except critical-error rate.

05 · Agentic workflows

The largest shift was coherence across steps, not isolated tool-call accuracy.

The 300-workflow agent evaluation required models to inspect information, call tools, perform calculations, maintain intermediate state, and verify a defined outcome. The post-trained model was more likely to use intermediate results to guide later actions, recognize insufficient information, and confirm the final state.

FIG. 03Performance on 300 agent workflows
Base modelFull pipelineGain, percentage points
End-to-end completion
52.4%76.9%
+24.5
Correct tool selection
68.1%89.3%
+21.2
Intermediate calculations
73.6%91.5%
+17.9
Handling missing information
49.7%78.4%
+28.7
Outcome verification
57.2%84.6%
+27.4
Score (%)
0255075100
pts

Scores from the full paper’s agent evaluation. Gains are measured in percentage points.

06 · Ablation analysis

Different training signals corrected different classes of failure.

Removing one signal at a time showed that the stages were not interchangeable. The result is a practical map from observed failure to the kind of data or feedback most likely to address it.

Without demonstrations

Context became generic

Broad knowledge remained, but the model lost domain-specific patterns for structuring financial decisions.

Without preferences

Judgment alignment fell

Correctness held up better than the ability to prioritize, explain trade-offs, and calibrate uncertainty.

Without trajectories

Multi-step execution degraded

Static answers remained comparatively stable while tool use and long-horizon completion declined.

Without deterministic checks

Plausibility displaced outcomes

Language graders could reward strong-looking responses that failed numerical or tool-based checks.

07 · Limits and conclusion

Benchmark gains are evidence of capability, not a substitute for production proof.

Expert data is expensive, professional judgment is not always deterministic, graders can encode their own failures, and financial distributions change with regulation, markets, products, and customer behavior. Specialization can also create trade-offs on unrelated tasks. Production monitoring and recurring expert calibration remain part of the system.

Preserve legitimate disagreement

Ambiguous financial decisions should not be forced into a false single-answer reward.

Calibrate the graders

Automated evaluation must be checked against the behaviors domain experts actually value.

Monitor distribution shift

Tax rules, products, regulations, markets, and customer circumstances continue to change.

Measure the deployed workflow

Held-out evaluation supports a claim; it does not prove reliability in every customer interaction.

Conclusion

The competitive asset is the improvement loop itself.

A vertical AI company does not need to build a foundation model to own differentiated model capability. It can own the workflows that expose failures, the experts who diagnose them, the environments that reproduce them, the evaluators that measure them, and the post-training pipeline that converts those signals into better behavior. For financial AI, the objective is not simply a model that knows finance. It is a model that can perform financial work reliably.

THE COMPLETE RESEARCH

Read the full technical paper.

Domain Adaptation and Post-Training of an Open-Weight Model for Financial Reasoning

23 sections · 6 appendices · Original document
Download paper

DOMAIN MODEL RESEARCH

Turn expert work
into model capability.

We build the data, environments, evaluators, and post-training systems that help models perform specialized work reliably.

Build with us