New arXiv work finds that real-world LLM agent harnesses are undertested, with LLM-dependent harness code covering less than half of its lines and branches, highlighting reliability risks in agentic software infrastructure.

#AI #LLMAgents #SoftwareTesting #MLInfra
https://arxiv.org/abs/2610.04921

A new arXiv paper examines why using frontier models as meta-agents to generate terminal tasks and verifiers for RL training often stalls, identifying benchmark invalidity, harness brittleness, and reward misalignment as core failure modes.…

#AI #RLA #LLM #MLInfra
https://arxiv.org/abs/2610.02405

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