Feature scaling improves logistic regression on breast_cancer.
Survives reshuffling across seeds. Reported.
Autonomous research · evidence-gated
Predict before compute.
Verify before claim.
Lucky Loop is an autonomous ML research agent. It forecasts an experiment's outcome with a language world model, runs the real code, and only claims what the evidence supports — negative results included.
AThe loop
01
PREDICT
language world model, before compute
02
RUN
real code on real hardware
03
VERIFY
effect vs. seed noise
04
CLAIM
only what evidence supports
BClaim ledger
effect/noise = 10.7
Language world model beats trivial heuristics on tabular ML.
No measurable benefit. Reported anyway — honest negative result.
Our pipeline reaches +98.7% improvement.
Did not reproduce. Our own verifier killed it. Verify before claim — applied to ourselves.
CAblation · world model vs baseline
DReferences · 6
- 01Agent Laboratory: Using LLM Agents as Research AssistantsSamuel Schmidgall et al. · arXiv:2501.04227
- 02AI Scientists Fail Without Strong Implementation CapabilityMinjun Zhu et al. · arXiv:2506.01372
- 03Qwen-AgentWorld: Language World Models for General AgentsQwen Team · arXiv:2606.24597
- 04AI Research Agents for Machine Learning: Search, Exploration, and Generalization in MLE-benchEdan Toledo et al. · arXiv:2507.02554
- 05The AI Scientist: Towards Fully Automated Open-Ended Scientific DiscoveryChris Lu et al. · arXiv:2408.06292
- 06MLE-bench: Evaluating Machine Learning Agents on Machine Learning EngineeringJun Shern Chan et al. · arXiv:2410.07095