Strategy Auctions for Agent Routing and Self-Improvement

Speaker: Lisa Alazraki (Imperial College)
Series: NLIP Seminar Series

Can smaller language-model agents compete with larger, more expensive ones? On short tasks the gap is small, but it widens sharply on longer-horizon workflows. This talk presents Strategy Auctions for Workload Efficiency (SALE), where the agents’ own plans become both a routing signal and a learning signal, letting small agents improve and take on more of the work over time. On deep search and coding, SALE outperforms the largest agent in the pool at substantially lower cost. We find that plans show early signs of success or failure before any expensive trajectory begins, including failure modes that size alone does not protect against.

Lisa Alazraki is a PhD student in the NLP Group at Imperial College London, advised by Marek Rei. She is also currently a Research Associate at Ai2. Previously, she held research internships at Meta Superintelligence Labs, Cohere and Google Research.