New Research Explores LLM Agent Frameworks for Data Pipelines and Code Generation

New Research Explores LLM Agent Frameworks for Data Pipelines and Code Generation

Researchers have published multiple studies examining how large language model (LLM) agents can be used to construct data workflows and improve code generation reliability.

According to arxiv.org, DataFlow-Harness introduces a platform that guides LLM agents to construct platform-native 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 paper.

A separate study on cross-model code review found asymmetric effectiveness in LLM pairings. According to arxiv.org, in experiments with 116 LeetCode tasks, Claude reviewing Codex drafts raised the pass rate from 71.6% to 89.7%, while Codex reviewing Claude drafts actually decreased performance from 91.4% to 82.8%. The paper states this indicates “the useful pairing is asymmetric: use Claude to review Codex, not the other way around.”

MineValiCoder, another framework described on arxiv.org, addresses test-driven development by filtering faulty test cases through self-validation. The system 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 the research.

Additionally, arxiv.org reports that MedKGent constructs medical knowledge graphs using over 10 million PubMed abstracts, resulting in a graph with 156,275 entities and 2,971,384 triples.