Knowledge-Centric AI Self-Improvement Boosts Performance, Reduces Cost

Traditional AI self-improvement often focuses on optimizing the agent itself, leading to improvements tied to specific designs or task distributions that are hard to maintain or transfer. This paper introduces a complementary "knowledge-centric" self-improvement paradigm. In this model, agents are kept generic and disposable, while the core object of improvement is a shared, curated knowledge base that agents leverage for future tasks. The researchers operationalized this concept through a simple protocol: agents attempt a task, contribute evidence-grounded insights to a shared knowledge base via forums, and then this knowledge is distilled. Controlled case studies across abstract reasoning, coding, and terminal benchmarks demonstrated that this protocol significantly improved solve rates while simultaneously reducing dollar costs compared to agent-centric baselines. Crucially, the distilled knowledge proved transferable to held-out tasks and even across different large language model families, indicating that the improvements are robust and not merely specific to a particular LLM or run. This research suggests that progress in self-improving agentic systems can be primarily driven by enhancing persistent, curated knowledge.