Reproduce “Language Models Share a Readout Bottleneck in Working Memory” 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
Dongyu Gong
The authors test why pretrained models make lure errors in a synthetic N-back task, comparing failures to locate the target with failures to use it amid competing letters. Across 15 models, suppressing half the letter-identity component of non-target items raises d′ more than sharpening attention: median gains are +0.60 versus +0.05. This supports interference in this setup, not a fix.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.