According to arxiv.org, researchers have introduced DataFlow-Harness, a platform that guides LLM agents to construct directed acyclic graphs (DAGs) through typed, incremental mutations rather than free-form scripts. On a 12-task data-engineering benchmark, the system achieved a 93.3% end-to-end pass rate while reducing monetary cost by 72.5% and generation latency by 49.9% compared to Vanilla Claude Code, according to the technical report.
In related research on code review workflows, arxiv.org reports that a controlled experiment on 116 LeetCode tasks found asymmetric effectiveness in cross-model code review. According to the study, Claude review raised Codex drafts from 71.6% to 89.7% (p_BH=.001), while Codex reviewing Claude drafts actually decreased the pass rate from 91.4% to 82.8% (p_BH=.046). The paper, accepted by Agentic SE @ KDD’26, concludes that “the useful pairing is asymmetric: use Claude to review Codex, not the other way around.”
Separately, arxiv.org describes MineValiCoder, a Test-Driven Development framework that achieved Pass@1 scores of 96.34% on HumanEval, 87.40% on MBPP, 64.00% on APPS, and 51.33% on LiveCodeBench across four LLMs.
According to arxiv.org, MedKGent uses over 10 million PubMed abstracts from 1975 to 2023 to construct what researchers describe as “the largest LLM-derived medical KG to date,” containing 156,275 entities and 2,971,384 triples.