Negative Knowledge as Failure-aware Shared Memory for AutoResearch

The article presents a novel approach called the negative knowledge memory layer, which allows AI-assisted research systems to retain and utilize information from failed experiments as structured knowledge. Evaluated on ScienceAgentBench and nonlinear math-physics PDE problems, this method demonstrated improved performance over traditional AutoResearch baselines, using fewer tokens and enabling agents to solve previously unsolvable tasks. This advancement highlights the importance of maintaining a comprehensive knowledge repository that includes both successes and failures, enhancing the overall efficacy of AI in scientific research.