Evidence can exist in the wrong part of a workflow
The SPLICE paper examines multi-agent systems whose communication paths are fixed in advance. One agent may retrieve an important fact that never reaches the agent responsible for using it, leaving an information gap even though the overall workflow found the evidence.
Adding a connection can help, but it can also spread an incorrect upstream conclusion. The authors therefore treat delivering information and deciding whether to adopt the resulting answer as separate problems.
Make a candidate change, then inspect its effects
After the original run, SPLICE uses the query and node outputs to locate a possible gap. It inserts one directed edge, reruns the receiving agent and affected downstream nodes, and produces a candidate answer.
A trained selector then considers the final answers and process features: how far the receiver changed, where its new content came from and whether the change reached the output. It can adopt the candidate or retain the original rather than assuming that a more connected graph is always better.
Include the cost of the repair
Across five benchmarks, the paper reports an eight-percent Web-tool-use score improvement and a 5.7-percent mean improvement on four reasoning and planning benchmarks relative to its static workflow. The reported cost is 1.25 to 1.44 times the baseline, so the gain includes additional execution rather than a free improvement.
For an evaluation, track whether the added connection transfers useful evidence or merely transfers confidence in a wrong answer. Compare the selected output with both candidate runs and include the re-execution cost. The study offers a method for targeted communication repair in the evaluated workflows, not a guarantee that more agent messages improve reliability or that a selector always recognizes a harmful change.
Comments