Reproduce “Latent Action Reparameterization for Efficient Agent Inference” 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

Wenhao Huang, Qingwen Zeng, Qiyue Chen, Zijie Guo, Yu Sun, Cheng Yang, Siru Ouyang, Jiri Gesi and 6 more

LAR compresses recurring action fragments in language-model agents into learned symbols while keeping task-specific values, such as search queries, explicit. It identifies predictable fragments from agent trajectories and trains a student model to use their symbols. On Qwen3-8B TriviaQA, LAR scores 80.09 versus ReAct’s 77.84 and uses 27.1% fewer action tokens than Vanilla; results vary, including a Llama TriviaQA decline.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.