ML-Agent ICML 2026 Reproduction 17897
Collection
Trackio claim logbook, artifact Buckets, and dataset mirror for the ICML 2026 ML-Agent reproduction; source links are in the logbook. • 6 items • Updated
Dataset Viewer
The dataset could not be loaded because the splits use different data file formats, which is not supported. Read more about the splits configuration. Click for more details.
Couldn't infer the same data file format for all splits. Got {NamedSplit('train'): (None, {}), NamedSplit('test'): ('text', {})}
Error code: FileFormatMismatchBetweenSplitsError
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
ML-Agent ICML 2026 reproduction (#17897)
This bundle audits and partially reproduces the five requested claims for
"ML-Agent: Reinforcing LLM Agents for Autonomous Machine Learning Engineering"
(OpenReview kcPPWaoegr, arXiv 2505.23723).
The headline benchmark cannot be independently rerun from the public release: the official repository still withholds the trained checkpoint, evaluation code, RL code, and training trajectories. The bundle therefore separates:
- source verification against arXiv v1 and v2;
- release/readiness auditing at official commit
15932e7; - a disclosed proxy reproduction of the 100-to-10 diversity-selection mechanism;
- exact executable tests of both the v1 and revised v2 reward equations.
Local smoke test:
uv run --with numpy --with matplotlib scripts/audit_claims.py
uv run --with numpy --with matplotlib scripts/reproduce_mechanisms.py \
--backend hash --random-trials 1000 --output-prefix local
The GPU run uses sentence-transformers/all-MiniLM-L6-v2 embeddings in place
of the hash smoke-test backend. All results explicitly label this as a mechanism
reproduction rather than a full ML-Agent replication.
- Downloads last month
- 42