ML researchers debate code-sharing safety

The traditional "reproducibility first" mandate is colliding with a cutthroat research culture where "simple but novel" ideas are increasingly vulnerable to theft.

• –NeurIPS 2026 guidelines strongly encourage code but do not strictly mandate it for the main track, providing a tactical choice for cautious authors.
• –The use of AI agents for rapid code refactoring has lowered the barrier for reviewers to "cleanly" replicate and steal ideas without leaving a clear trail.
• –arXiv remains the primary "proof of stake" for researchers, yet it simultaneously signals competitors to accelerate their own parallel work.
• –Reviewers frequently demand code for evaluation but rarely execute it, turning a reproducibility requirement into a performative risk for authors.
• –The Datasets & Benchmarks track's stricter code requirements highlight an internal inconsistency in how conferences handle academic trust.