DeReAct introduces a modular agent architecture that breaks the coupling in traditional ReAct where a single LLM policy simultaneously proposes actions, interacts with the environment, and decides task completion. Two gating policies are externalized: a Critic that validates actions before execution, and a Context Manager that reconstructs an environment‑supported \textsc{State} and certifies whether the task is finished. This separation allows independent control of action authorization and completion, preventing error propagation and unsupported early termination.
On the GAIA and SWE-bench Verified benchmarks, DeReAct yields the largest Pass@1 gains for weaker models—about 6.5–7.0 points for Qwen3‑Coder‑480B and 4.2–5.2 points for Claude Sonnet 4.5—while the improvement diminishes as the Brain model becomes stronger. Trajectory and ablation analyses show that external gating is effective when targeted failures are sufficiently frequent and the gating policy itself is reliable. With Claude Opus 4.5, Pass@1 remains comparable to ReAct, but DeReAct produces trajectories that are more evidence‑complete and constraint‑satisfying, indicating a trade‑off between earlier termination and stronger grounding.
Overall, DeReAct boosts weaker agents while preserving grounding benefits as model capabilities increase.
Review