IBM uses AI agent teams to speed up a quantum error-correction decoder chip design

IBM

IBM Research used a team of AI agents to search for faster hardware designs for a quantum error-correction decoder. A coordinator agent directs parallel worker agents that propose alternative designs, and a verifier checks each one before designs are ranked by speed and hardware use. The work is a research study, not a finished product.

• The open-model configuration reached a 52x speedup over baseline for about $26 of inference cost, across 100 hypotheses.
• A proprietary frontier model reached a 204x speedup for about $299, about 11.5 times the open configuration's cost.
• The open configuration reached 60x at 200 hypotheses; doubling the search budget brought diminishing returns.
• The speedups are high-level synthesis estimates, and the costs cover model inference only, not synthesis or verification.
• The decoder tested is Relay-BP on a [[144,12,12]] bivariate bicycle code; llm-d, an open-source distributed inference framework, served the models.