MRKL Systems: Modular neuro-symbolic agents¶
Explicit routing, expert registry and aggregation, not a single prompt returning the correct tools.
Paper information¶
| Field | Details |
|---|---|
| Paper | Paper or original source |
| Institution | AI21 Labs / Hebrew University / Technion / University of British Columbia. |
| First public release | 2022-05-01 |
| Author code | The complete Jurassic-X/MRKL implementation was not released. |
| Local adapter or method key | mrkl |
| Local reproduction code | src/auto_research/agent_research/ |
Paper method and evidence¶
Route inputs to neural or discrete experts such as knowledge bases and calculators. Discrete modules provide deterministic computation; the LM handles language understanding and the interface. The contribution is routing and modular boundaries, not a new LM loss.
flowchart LR
A[Route incoming task] --> B[Select neural or discrete expert]
B[Select neural or discrete expert] --> C[Aggregate result]Original paper figure¶
The original figure shows modular routing. From the original paper; copyright remains with the authors.
Core equation¶
Reported results¶
Jurassic-X cases illustrate controllability on computation and fresh knowledge; no reproducible production A/B or unified offline benchmark table.
Local reproduction¶
ScaleMCP mini, 120 episodes, seed 42. Route calculators, calendars, maps, weather, databases and tables to discrete experts. Baseline/MRKL joint success 1.0000/1.0000, mean cost 64.5000/1.2500, router/symbolic calls 0/0 vs 360/170. Metrics. No Jurassic-X, live search or private KB; cost is a common proxy, not latency.