Source record

Published by arXiv cs.MA (T1 source). The original is at https://arxiv.org/abs/2609.12746.

Pipeline notes

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

Ablation of LAST-CQ, a five-agent training-free Text-to-Cypher framework, run across 2,471 live database queries and six LLM backbones. Key findings: detecting failure and routing to a retry drives recovery, not feedback sophistication or parallel sampling; LLM-synthesized feedback barely beats raw DB error strings. Recovers 91.7% of single-pass failures. Flags execution-BLEU as a flawed evaluation metric and an LLM judge as optimistic by ~9 points.

Actionable ablation that strips away assumed essentials: skip elaborate feedback synthesis in agentic query loops, spend budget on simple failure detection and retry routing.

Cited by

No citations on record.