Reproduce “Controlling Foundation Models via Internal Interventions: A Survey of Inference-Time Methods” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Pradeep Bolleddu, Karn Tiwari, Varnith Chordia, Abhishek Pandir

The authors survey inference-time control in large generative models, organizing methods by whether they modify representations, attention routing, or decoding. The taxonomy spans LLMs, VLMs, and diffusion models, while deployed base weights remain fixed; preparation and runtime costs may still apply. In one COCO POPE evaluation, CG-VLM raises LLaVA F1 from 79.5 to 86.2, not establishing cross-method rankings.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.