Reproduce “Repurposing Pre-trained LLMs as High Fidelity Continuous Text Autoencoders” 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
Arkanath Pathak, Unnat Jain, Alexander C. Berg
LLMAE encodes text as fixed-length continuous latents by using learned tokens to gather input at an intermediate layer, then reconstructing with later layers that cannot access source text. On 500 held-out C4-News-Stratified documents up to 1,024 tokens, Qwen2.5-0.5B achieves 97.4% exact match, compared with 13.4% for ICAE. This is bounded evidence: exact match is 82.4% on CreationMMBench.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.