ReFlect: Training-Free Error Recovery for Long-Horizon LLM Reasoning
Long-horizon reasoning is where production LLM agents tend to quietly break. A model can produce a plausible-looking chain of thought, accept a wrong intermediate answer, and continue building on that error for every step that follows. By the time th...
