Reproduce “Selective Critique for Cost-Aware LLM Agents in Long-Horizon Decision Making” 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
Heewon Park, Somin Im, Minhae Kwon
SAG uses the actor’s uncertainty about valid next actions to decide when an auxiliary critic should review an action, rather than checking every step. After successful episodes, it fine-tunes the actor on actions taken, including critic-suggested actions. On ALFWorld, Qwen2.5-7B SAG achieved 78.3% success versus ReAct’s 24.6%, with 56k versus 55k average tokens per episode in sequential evaluation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.