JAMLE-STAR (Modified MLE-STAR)

How my solution addresses the problem statement.

My solution is a modified MLE-STAR implementation tailored to this specific challenge and its requirements. It is a synchronous multi-agent system where agent of different roles does not run concurrently but rather sequentially in different stages. The agent is fully autonomous in all aspects of things like ablation, code block replacement, data leakage detection and repair, self repair issues. I've implemented multiple fallbacks to ensure that there are zero human interventions. During refinement, it runs on an outer (ablation) and inner loop (experiment hypothesis). It is also web search driven when it does it's researching loop for candidate models to use. This modified MLE-STAR seeds the baseline so that it does not perform worst than the baseline provided in the starter kit in all scenarios. Anything worst than baseline will be dropped off and won't make it to the outer and inner loop.

During the whole run, there were zero human interventions and I only had to manually stop it before the 6 hour wallclock to adhere to the competition rules.

Development tools used

1. VSCode
2. Antigravity CLI with Gemini (Planning)
3. OpenCode with DeepSeek V4 (Implementation of plan)

APIs used

DeepSeek API (V4 Flash and Pro for running agent) Tavily (For enabling web search for research)

Libraries and frameworks used

Pydantic-ai (Agentic Framework) ddgs (web search fallback incase Tavily fails)

Datasets and assets used

KuaiRand-Pure

Built With

• antigravity
• opencode
• pydantic
• python
• vscode