GeorgeZhang
ZH
Home Work Lab CV Focus Writing Contact 中文
Back to list

DPO

DPO for Tool-use Preference

Turns tool-use issue cases into preference data connecting replay evaluation, policy optimization, and model improvement.
  • DPO
  • Tool-use
  • Preference Data

Hypothesis

If tool-use failures are converted into chosen/rejected preference pairs, the Agent platform can route production issue cases into evaluation, policy optimization, and model improvement interfaces.

Dataset

The data design keeps task goal, available tools, business constraints, chosen behavior, and rejected behavior, then stratifies samples by failure type, industry scenario, tool risk level, and human-handoff policy.

Evaluation

The evaluation covers win-rate, tool selection accuracy, parameter validity, refusal quality, and human-handoff triggers, supporting platform release gates and policy iteration.

Result

Validate tool selection, parameter validity, and refusal boundaries through replay evaluation, then route improvements to prompts, policies, or model-level optimization through a reviewable release gate.