Failure-Aware Adaptive Inference Improves Reliability in Long-Horizon Language Model Agents

<title>Abstract</title> <p>Large language model agents increasingly operate in sequential environments where early decision errors can propagate across multiple interaction steps, leading to complete trajectory failure. Existing approaches rely on successful demonstrations, similarity-based memory retrieval, or post-failure reflection, but provide limited mechanisms for explicitly learning from previous execution failures before a new action is taken. We introduce Failure-Aware Adaptive Inference (FAAI), a framework that represents failure experiences as reusable decision knowledge and integrates adaptive memory retrieval, failure-risk estimation, and proactive self-correction during inference. Across sequential decision-making tasks with interaction horizons up to 40 steps, FAAI consistently improves task success while reducing failure propagation. Compared with direct prompting, chain-of-thought reasoning, ReAct, retrieval-augmented agents, and reflection-based methods, FAAI achieves the largest gains in long-horizon settings where error accumulation is most severe: 95.83% success at H = 40, versus below 7% for all baselines. Ablation studies confirm that adaptive memory, risk prediction, and pre-execution correction each independently contribute to reliability, with risk prediction being the most critical component. These findings indicate that explicitly modelling prior failures provides a practical mechanism for improving the robustness of language-model-based agents without requiring additional model training.</p>