RL post-training Qwen 27B as a Cypher agent lifts graph-query accuracy 6.2 points for $119

sophiamyang · x · 2026-10-11

Adithya Giridharan RL post-trained Qwen 3.8 27B on Fireworks' serverless training API to answer questions over a 650K-node Neo4j graph by writing Cypher queries.

Key facts:

The task remodeled the BIRD codebasecommunity dataset as a property graph (649,846 nodes, 1,380,394 relationships); the agent can call a runcypher tool up to 8 times before answering.
Training used only machine-generated questions; evaluation was on 186 unseen human-written questions from the BIRD benchmark.
Strict accuracy rose from 0.598 to 0.660 (+6.2 points, 95% interval 3.1–9.4) against an untrained adapter, clearing a pre-registered threshold.
Notably, the gain came entirely from greater reliability on questions it could already sometimes solve — it never solved a previously unsolved question. The strict/lenient scorer gap also shows how answer formatting affects scores.
The whole experiment cost about $119 with no GPUs to provision.

Original post →

RL post-training Qwen 27B as a Cypher agent lifts graph-query accuracy 6.2 points for $119

More from coding & agent