Artifacts from the first complete run of the LossFunk AI scientist pipeline, including prompts, research ideas, and failure analyses from our paper accepted at Agents4Science 2025.

• MARL-idea/ - Multi-agent RL coordination (failed at implementation)
• SALVO-WM-idea/ - Perceptual loss for world models (failed at evaluation)
• SDTS-WM-idea/ - Stochastic tree search in world models (failed at evaluation)
• SemEnt-ALGN-idea/ - Semantic entropy jailbreak detection (successful paper)

Complete workflow prompts used in our system:

• idea_generation/ - Paper pair evaluation and mashing
• hypotheses_generation/ - Converting ideas to testable hypotheses
• experiment_planning/ - Implementation planning for Claude Code
• paper_creation/ - Paper outlining and readiness checks

If you use these artifacts, please cite our report: "Supporting LLMs from Research Idea to Paper" (2025)

You can also refer to the successful submission to Agents4Science 2025:

• OpenReview: https://openreview.net/forum?id=B6ZrLXou3u
• Conference: https://agents4science.stanford.edu/submissions.html