Caltech's knowledge-centric self-improvement cuts AI costs while boosting solve rates

Current self-improving AI systems focus on optimizing the agent—tweaking prompts, workflows, or even code. This agent-centric approach makes improvements expensive to maintain and hard to transfer, as gains are tied to specific designs, task distributions, or adaptation runs. Caltech researchers introduce a complementary paradigm: knowledge-centric self-improvement. Here, agents remain generic and disposable, while the persistent object is a curated knowledge base that agents leverage for future tasks. The protocol has agents attempt one task, contribute evidence-grounded insights via task-level and cross-task forums, then undergo knowledge distillation. This makes improvements inspectable, transferable, and portable.

Across abstract reasoning, coding, and terminal benchmarks, the knowledge-centric approach improved solve rates and reduced dollar cost relative to agent-centric baselines. Notably, the distilled knowledge transferred to held-out tasks and across different LLM families (e.g., GPT-4, Claude), proving the improvements are not model- or run-specific. The researchers argue this supports a new view: progress in agentic systems can be driven primarily by curated persistent knowledge rather than by continually optimizing agents. This could lead to more efficient, scalable AI systems that share knowledge across tasks and environments.