DeltaML-Bench (Modular Agent, 2 x 12h): leaderboard

DeltaML-Bench (2026) gives ML agents 48 tasks drawn from research papers: improve the published baseline inside the original, imperfect open-source repository, with audits that invalidate specification gaming. This board is the per-run success rate in percent with the Modular agent scaffold adapted from RE-Bench and two 12-hour runs per task; higher is better.

Metric: Per-run success rate (%; two independent runs of 12 hours per task; Modular Agent scaffold; 48 tasks from research papers in five ML domains, each requiring the agent to beat the published baseline inside the original imperfect repository; a run succeeds when the audited solution improves on the baseline, and runs flagged as specification gaming are invalidated). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 2 models tracked.

Top models

Interactive version: theaggregate.ai/benchmark?slug=deltaml-bench-modular-agent-2-x-12h · How It Works · Data refreshed daily, snapshot 2026-11-13.