ReflectiChain: Mitigating Semantic-Execution Drift in Long-Horizon LLM Agents via Retrospective Reflection and Double-Loop Policy Adaptation

Large Language Model (LLM) agents in long-horizon planning often exhibit Semantic-Execution Drift (SED), where executed actions progressively deviate from original language constraints. To formalize this phenomenon, we model SED as a stochastic drift process, expressed by the recurrence D(t+1) = alpha D(t) + epsilon(t) + beta P(t), and show that policies with an alpha below one induce semantic contraction. We propose ReflectiChain, a framework integrating Retrospective Reflection, a Latent World Model, and Double-Loop Policy Adaptation to preserve semantic consistency during long trajectories. To evaluate SED, we introduce Sema-Sim, a multi-agent supply chain benchmark containing 10 policy constraints, six adversarial perturbations, and 30-step planning horizons. We further propose the Semantic Fidelity Index (SFI) for measuring instruction adherence. Experiments on DeepSeek-V3.2 across seven reasoning strategies show that ReflectiChain achieves the highest SFI (88.7) and stable semantic contraction (alpha = 0.823, below one). The results are consistently validated on Qwen3.5-122B and Qwen2.5-72B. Ablation studies demonstrate that Retrospective Reflection contributes most to performance gains. Additional analyses on scalability, failure modes, and cost efficiency further verify the robustness and practicality of the proposed framework. All code and evaluation resources are publicly released.