Reproduce “HyEvo: Self-Evolving Hybrid Agentic Workflows for Efficient Reasoning” 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
Beibei Xu, Yutong Ye, Chuyun Shen, Yingbo Zhou, Cheng Chen, Mingsong Chen
The HyEvo framework autonomously constructs heterogeneous agentic workflows by synergistically combining probabilistic Large Language Model nodes with deterministic code nodes. This system achieved an average performance of 84.82% across five diverse benchmarks and demonstrated substantial efficiency gains, reducing inference costs by up to 19 times and execution latency by up to 16 times on specific datasets compared to baselines.
Key insights
- Synergistically combining probabilistic LLM reasoning with deterministic code execution in workflows yields superior task performance and substantial efficiency gains over homogeneous LLM-only approaches.
- An LLM-driven multi-island evolutionary search, enhanced by reflection and a quality-diversity (MAP-Elites) strategy, is crucial for effectively navigating the expanded and hybrid search space of heterogeneous workflows.
- Optimal solutions for complex reasoning tasks emerge from hybrid architectures that intelligently divide labor between LLM and code components, rather than relying solely on one paradigm.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.