Graph-based multi-agents smash long-context benchmarks–89% MMLU-Pro on 8B modelsgithub.com/tjoo5121 pointFeelTheAGI24 months ago