Invited Talk 5: Siva Reddy (Lifelong Agents from Small Language Models)
Abstract
Online live invited talk from Siva Reddy
Frontier API agents are too expensive to deploy per user, per task, per interaction; the deployable unit for lifelong agents is a small language model. Committing to small agents raises three questions. First, how does a small agent acquire frontier-level competence? I will describe Agent-as-Annotators (A3), a structured distillation framework that produces a 9B web agent surpassing GPT-4o and Claude 3.5 Sonnet on WebArena. Second, how does a fixed small agent adapt to each user? I will present AdaptArena, a benchmark for test-time personalization from implicit interaction histories, where even frontier models reach less than half of oracle performance. Third, how does a small agent retrieve and communicate without a separate encoder? I will describe LLM2Vec-Gen, which turns a frozen LLM into its own output-space encoder — inheriting reasoning and safety for retrieval, and opening a path to embedding-space agent-to-agent communication. The thread: lifelong behavior does not require lifelong retraining.