Across 278,790 review conversations in 300 open-source projects, AI agent suggestions were adopted far less often than human ones, and when adopted they made the code bigger
AspireAI community · arXiv
Code review is a critical software engineering practice, and AI agents that understand code context, plan review actions and interact with development environments are increasingly integrated into it, but there is limited empirical evidence comparing AI agents and human reviewers in collaborative workflows. The authors conduct a large-scale analysis of 278,790 code review conversations across 300 open-source GitHub projects, comparing the feedback humans and AI agents provide and how interaction shapes outcomes. Human reviewers provide additional feedback that AI agents do not, including understanding, testing and knowledge transfer, and exchange 11.8% more rounds when reviewing AI-generated code than human-written code. Code suggestions made by AI agents are adopted into the codebase at a significantly lower rate than human suggestions, and over half of the unadopted AI suggestions are either incorrect or addressed through alternative fixes by developers. When adopted, AI agent suggestions produce significantly larger increases in code complexity and code size than human ones. The authors conclude AI agents can scale defect screening but human oversight remains critical for suggestion quality and contextual feedback. The limitation is that this is observational analysis of existing repositories rather than a controlled trial, so adoption rate reflects reviewer behaviour and suggestion wording as much as agent capability.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.