Reproduce “Latent-Lens: Visual Perception in Small Language Models Through Latent Communication” 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
Mo Shahdloo, Faegheh Sardari, Mathew Salvaris
Latent-Lens sends visual information from a frozen VLM to a separately pretrained text-only SLM as hidden states rather than generated text. It maps signals from early, middle, and late layers into the SLM’s input space and adds LoRA adapters. On CLEVR, it scores 81.0 ± 7.9% versus 58.3 ± 5.7% for text communication with the same LoRA budget, though Latent-Lens also trains a projection network; the best band count varies by task.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.