LLM-based AutoResearch systems often improve one behavior at another's expense, since scalar feedback hides those trade-offs. The paper finds that adding competing-behavior feedback when task gains slow down boosts…

#AutoResearch #MachineLearning #AIresearch
https://arxiv.org/abs/2609.39933

TraceML pairs human and agent ML development under a shared version-level schema, showing experts make larger, more deliberate edits than auto-research agents across paired Kaggle trajectories. The trace-based…

#TraceML #AIResearch #MachineLearning #AutoResearch
https://arxiv.org/abs/2608.26086

A new arXiv paper applies Karpathy's AutoResearch paradigm to production embedding systems, running 220+ multi-GPU experiments over twelve weeks and documenting five recurring failure modes at scale. Multi-agent frameworks continue to…

#AutoResearch #LLM #MLInfra
https://arxiv.org/abs/2609.30541