AutoResearch: Execution-grounded multi-agent framework automates verified research workflows

Automated research agents today can write code and draft papers, but they often generate broken experiments or cite unsupported claims. To address this, Rajesh Kumar and four co-authors from multiple institutions present AutoResearch, a multi-agent framework designed to ground every step in verifiable execution. The system runs code in a sandboxed Python/PyTorch environment, detects runtime errors, and iteratively repairs them. It also cross-checks every cited source against the claim it supports, using a dedicated claim‑support auditor. A central decision controller filters outputs based on these verification signals before producing structured LaTeX artifacts.

In controlled evaluations, AutoResearch outperformed comparable baselines on coding benchmarks (HumanEval, MBPP, a SciCode subset), citation‑validation tasks, and local claim‑support audits. The framework is open source and intended as a reliability-oriented research assistant, not a fully autonomous scientist. By treating execution errors and citation failures as filtering signals, AutoResearch makes automated research workflows more trustworthy, reducing the risk of hallucinated experiments and fabricated references.