Reproduce “Towards Human-Agent Communication via the Information Bottleneck Principle” 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
Mycal Tucker, Julie Shah, Roger Levy, Noga Zaslavsky
The paper introduces VQ-VIB, which trains communicating agents to balance task utility, informativeness, and signal complexity. A speaker samples a continuous representation, quantizes it to a discrete code, and helps a listener reconstruct the input. Reconstruction improved convergence across tested settings, while complexity penalties reduced complexity and utility; color-distance correlation reached 0.97 ± 0.01, without testing human understanding.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.