Source record

Published by arXiv cs.MA (T1 source). The original is at https://arxiv.org/abs/2607.22642.

Pipeline notes

The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.

CRAFT is a two-stage post-training recipe for enterprise coding agents: schema-stripped supervised fine-tuning followed by execution-shaped reinforcement learning, with trajectories filtered by execution validation, data-integrity checks, and an LLM-judge audit. Evaluated on advertising analytics (25 entities, 30 workflows), it improves Agent Score by 9.6 pp and cuts input tokens ~9x versus schema-stuffed prompting.

Concrete benchmark evidence that moving schema knowledge from prompts into post-training reduces token cost and improves multi-turn consistency for enterprise agent deployments.

Cited by

No citations on record.