Reproduce “ADEMA: A Knowledge-State Orchestration Architecture for Long-Horizon Knowledge Synthesis with LLMAgents” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhou Hanlin, Chan Huah Yong
ADEMA, a knowledge-state orchestration architecture, enables Large Language Model (LLM) agents to perform long-horizon knowledge synthesis with governable, inspectable, and recoverable knowledge progression. It explicitly tracks hypotheses, milestones, and artifacts, ensuring continuity and artifact integrity, particularly when facing interruptions.
Key insights
- Checkpoint persistence is a completion-critical mechanism, as its absence was the only condition leading to task failure in interruption-sensitive scenarios, highlighting its role in recoverable continuity.
- Architectural elements like dual-agent consensus governance and dynamic resource allocation introduce "governance overhead" (increased runtime) but contribute to trajectory discipline, explicit artifact progression, and cost-quality optimization.
- The explicit management of epistemic state, milestone progress, and artifact lineage enables a more robust and verifiable approach to long-horizon knowledge synthesis, moving beyond unstructured conversational paradigms.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.