Understand this story
Google DeepMind published a paper proposing an auction-style routing method: letting each model bid to answer each question, with the highest bidder answering. The paper simulated auctions across 110 models and over 30,000 EmbedLLM prompts, with results showing the method reached 99.2% of the optimal possible value, while routing to the best model by prompt type reached 89.4%. On the prompt clusters where each method performed worst, the auction maintained 96.0%, while the single best model reached only 6.0%. The paper also proves that when multiple agents typically give high valuations to prompts, the loss shrinks. Currently this work remains at the paper stage, with no qualified factual evidence of subsequent deployment or productization progress.
AI-generated from public reports
Latest development · Report
Google DeepMind uses bidding auctions to select models, achieving 99.2% of the optimal resultDevelopments
1 confirmed developments
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Oct 23
Rohan Paul
Google DeepMind uses bidding auctions to select models, achieving 99.2% of the optimal result
A new Google DeepMind paper proposes having each model bid on every question and letting the highest bidder answer; simulating auctions across 110 models and over 30,000 EmbedLLM prompts, it reached 99.2% of the best possible value, whereas routing to the best model by prompt type achieved 89.4%. On the prompt clusters where each method performed worst, the auction maintained 96.0%, while a single best model achieved only 6.0%. The paper demonstrates that losses shrink when multiple agents generally place high valuations on prompts.