Beyond Corrected Memory: Execution Consistency in Multi-Agent Systems
Correct shared memory is not enough to prove a multi-agent run actually followed the task.
The paper defines “execution consistency” as evidence that agents used state properly, handed off information, and reached a valid final agreement. Its central example is that two executions can retain the same records while one complies and the other violates the same rule. In tests, removing evidence such as receipt, action dependence, or response validity made 82.4% of opposite-label pairs indistinguishable; restoring it separated 97.9%. The authors’ CAVERT framework beat contract-prompted LLM and rule-based baselines across 12 benchmark-executor settings, and also improved recovery in four evaluated environments. ArXiv · AI/CL/LG's note
The paper defines “execution consistency” as evidence that agents used state properly, handed off information, and reached a valid final agreement. Its central example is that two executions can retain the same records while one complies and the other violates the same rule. In tests, removing evidence such as receipt, action dependence, or response validity made 82.4% of opposite-label pairs indistinguishable; restoring it separated 97.9%. The authors’ CAVERT framework beat contract-prompted LLM and rule-based baselines across 12 benchmark-executor settings, and also improved recovery in four evaluated environments. ArXiv · AI/CL/LG's note
score 4