AITOPAI热报
PapersSelected

Google DeepMind paper: model auctions hit 99.2% of best routing

rohanpaul_ai@rohanpaul_ai
Post
View the post
Why we picked it

Bidding removes the need to know in advance which model handles each question, giving a self-selecting alternative to router systems that degrade badly on outlier prompts. Marketplace and orchestration platform builders can apply a pay-your-bid mechanism when their model pool contains several strong competitors per prompt type. The gap on worst-case clusters (96.0% versus 6.0%) suggests auctions also act as a safety net where fixed routing fails hardest.

A Google DeepMind paper proposes letting AI models bid money for the right to answer each prompt, with the highest bid answering. The method was simulated in auctions among 110 models on over 30,000 EmbedLLM prompts. With all 110 models bidding, the auction reached 99.2% of the best possible value, versus 89.4% for routing each prompt type to its best model. On each method's worst prompt cluster, the auction held 96.0% while the single best model kept 6.0%. The paper proves the loss shrinks when several agents value a prompt highly, as when distilled models perform close to the best.

Original · rohanpaul_ai

New Google DeepMind paper shows that if AI models offer money to answer each question and the highest offer answers, results are 99.2% as good as always picking the best model.

i.e. say you don't know in advance which model is best for each new question. So who should answer?

The paper’s answer: let each model offer money for the job. The model that knows it is good at this question offers more. Highest offer gets to answer.

They proved the loss shrinks when several agents usually value a prompt highly, as when distilled models perform close to the best. They simulated auctions among 110 models on over 30,000 EmbedLLM prompts.

With all 110 models bidding, the auction reached 99.2% of the best possible value, versus 89.4% for routing each prompt type to its best model. On each method's worst prompt cluster, the auction held 96.0% while the single best model kept 6.0%.

If you run a model marketplace, a simple pay-your-bid auction works well when every bidding pool has several strong models.

#Google DeepMind#EmbedLLM#routing#reasoning models#paper

Keep reading

Recent stories with the same tag or section
PapersBuildersSelected02:49
Google DeepMind paper finds self-improving agents forget old tasks

The fastest-learning method caused the worst forgetting, so picking a self-improvement approach by learning speed alone is misleading. Teams deploying agents that edit their own prompts or memory should isolate task-specific rules and retest prior tasks after every update. The findings apply directly to anyone building persistent agent memory on top of frameworks like GEPA or ReasoningBank.

#Google DeepMind#GEPA#ReasoningBank#agents
Event
rohanpaul_ai@rohanpaul_ai2 sourcesSource
PapersBuildersSelected01:52
Gemini Agent Formally Verifies New Busy Beaver Math Results

The workflow itself is the lesson: an agent chose its own research targets, then gated every output through Lean verification, a pattern that transfers to other open math and code-verification domains. Researchers and agent builders can replicate the propose-then-verify loop without trusting model outputs. Compared with earlier autoformalization work, this one attacks uncomputable problems where no fixed benchmark list exists.

#Gemini#Google DeepMind#Lean#agents
elvis@omarsar0Source