Reproduce “Meaning Representations from Trajectories in Autoregressive Models” 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
Tian Yu Liu, Matthew Trager, Alessandro Achille, Pramuditha Perera, Luca Zancato, Stefano Soatto
The method represents a sentence through an autoregressive model’s scores for possible continuations, rather than a single vector. It samples continuations from two sentences and compares their scores on shared samples, enabling similarity and directional containment tests. On STS-B, Falcon-7B scores 67.7, versus 46.7; on SNLI-Entailment, it reaches 65.70% accuracy. These results are task-bounded and model-relative, not evidence of understanding.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.