JOVE: Joint Execution and Verification for Resource-Aware LLM Task Graphs
JOVE treats verification as a budgeted learning signal, not just a correctness check.
The paper proposes an online framework that assigns subtasks in a DAG to heterogeneous LLMs while choosing which intermediate outputs to pay to verify. Those verification results feed future allocation decisions under a long-term budget and per-query latency constraint. The method solves per-query mixed-integer linear programs and updates task-specific estimates of model quality from feedback. Across four reasoning benchmarks, the authors report competitive accuracy with average cost and latency reduced by at least 3.17x. ArXiv · AI/CL/LG's note
The paper proposes an online framework that assigns subtasks in a DAG to heterogeneous LLMs while choosing which intermediate outputs to pay to verify. Those verification results feed future allocation decisions under a long-term budget and per-query latency constraint. The method solves per-query mixed-integer linear programs and updates task-specific estimates of model quality from feedback. Across four reasoning benchmarks, the authors report competitive accuracy with average cost and latency reduced by at least 3.17x. ArXiv · AI/CL/LG's note
score 5