A Skill Suite That Turns Five Years of NeurIPS Into a Research-Idea Machine
The most-upvoted paper on Hugging Face today isn't a model. It's a set of skills that read what actually got published, and rejected, at the top ML conferences, and uses that to generate research ideas that might survive review.
ResearchStudio-Idea (arXiv 2607.04439, out July 5) is three agent skills stacked together. Paper-Search finds the relevant literature. Scoop-Check verifies whether your idea is actually novel or someone already did it. IdeaSpark runs the whole loop end to end. The clever part is the fuel: they mined 1,947 papers from ICLR, ICML and NeurIPS between 2021 and 2025 and distilled 15 reusable ideation patterns, the recurring moves that turn a hunch into a publishable contribution. So instead of a language model free-associating, it's pattern-matching against five years of what the field actually rewarded.
What separates this from the endless AI-scientist demos is the auditing. It checks whether the evidence is actually ready to support the claim, surfaces the unresolved bottleneck the idea has to clear, and pulls up conflicting prior work before you fall in love with your own hypothesis. Outcome-informed, meaning it's grounded in whether similar ideas historically landed or bounced. That's the difference between a brainstorm and a reviewer sitting on your shoulder.
The bigger arc: research ideation is one of the last purely-human steps people assumed agents couldn't touch. Papers like this, plus the week's agent-environment and skill work, keep chipping at that assumption. The claim isn't that AI replaces the researcher. It's that the what-should-I-even-work-on step is becoming a searchable, auditable skill you can install.
Paper: arxiv.org/abs/2607.04439
← Back to all articles
ResearchStudio-Idea (arXiv 2607.04439, out July 5) is three agent skills stacked together. Paper-Search finds the relevant literature. Scoop-Check verifies whether your idea is actually novel or someone already did it. IdeaSpark runs the whole loop end to end. The clever part is the fuel: they mined 1,947 papers from ICLR, ICML and NeurIPS between 2021 and 2025 and distilled 15 reusable ideation patterns, the recurring moves that turn a hunch into a publishable contribution. So instead of a language model free-associating, it's pattern-matching against five years of what the field actually rewarded.
What separates this from the endless AI-scientist demos is the auditing. It checks whether the evidence is actually ready to support the claim, surfaces the unresolved bottleneck the idea has to clear, and pulls up conflicting prior work before you fall in love with your own hypothesis. Outcome-informed, meaning it's grounded in whether similar ideas historically landed or bounced. That's the difference between a brainstorm and a reviewer sitting on your shoulder.
The bigger arc: research ideation is one of the last purely-human steps people assumed agents couldn't touch. Papers like this, plus the week's agent-environment and skill work, keep chipping at that assumption. The claim isn't that AI replaces the researcher. It's that the what-should-I-even-work-on step is becoming a searchable, auditable skill you can install.
Paper: arxiv.org/abs/2607.04439
Comments