1 paper
Ankush Kadu, Aswanth Krishnan
We present ReflexGrad, a dual-process architecture for within-episode failure recovery in LLM agents without demonstrations. When agents commit to a wrong approach early and exhaus…