CRAFT: Two-Stage Post-Training for Schema-Grounded Coding Agents
A recent study presents CRAFT, a two-phase post-training method designed for enterprise coding agents that converts natural language requests into executable code across proprietary APIs and schemas. This technique tackles the problem of needing to include comprehensive schema documentation in every prompt, which raises inference costs and complicates schema updates. Initially, CRAFT employs schema-stripped supervised fine-tuning (PLAN) to acquire domain-structured plans and executable actions from validated trajectories without the need for extensive schema injection during prompting. Subsequently, execution-shaped reinforcement learning fine-tunes the policy for selecting tools, enhancing code quality, and ensuring consistency between plans and code. The goal is to enhance reliability in multi-turn analyses for production-oriented analytics.
Key facts
- CRAFT is a two-stage post-training recipe for schema-grounded coding agents.
- It addresses inefficiencies of injecting exhaustive schema documentation into each prompt.
- First stage: schema-stripped PLAN supervised fine-tuning.
- Second stage: execution-shaped reinforcement learning.
- The approach targets enterprise analytics over proprietary APIs and schemas.
- It aims to improve reliability in multi-turn analysis.
- The paper is published on arXiv with ID 2607.22642.
- The method preserves consistency required for production-facing analytics.
Entities
Institutions
- arXiv