Preprint, not yet peer-reviewed. Researchers at Meta and Carnegie Mellon University study reinforcement learning for tool selection by a small language model and transfer to different retrievers and tool collections. The authors report that their post-trained Qwen3-8B outperforms Llama 4 Maverick on ToolBench; this is the paper’s reported result, not an independent benchmark verification.
Evaluates open-weight and proprietary AI models on exploit generation, vulnerability repair and subsequent attacks in five nonpublic software environments, using researcher-developed deterministic graders rather than LLM judges. Repair scores exceed attack scores in two environments and fall below them in three. Preprint, not yet peer-reviewed.
A February 2026 working paper (Csaszar, Peterson, Wilde) that had AI models and 346 experienced managers rank 30 live Kickstarter tech ventures before their fundraising ended. The best model, Gemini 2.5 Pro, ranked outcomes far more accurately than the humans, and human-AI teams did not beat it.
Why I recommend it: A preprint that has not been peer reviewed, and it covers just 30 crowdfunding campaigns. It is a striking result, but it is not proof that AI can pick winning businesses.
The open-access preprint server for physics, mathematics, computer science, and related fields — a primary source for cutting-edge AI and machine-learning research papers.
Why I recommend it: The best place to read AI research before it hits journals or the press; search by tag or author to follow a specific line of work.
