LLMs may use outdated choices even after recognizing an update

A post describing an LLM paper says models can recognize a changed preference or deadline yet still draw on older mentions in a long conversation. It says nudging attention toward the newest value fixed most of these mistakes in five open models without retraining. In another test it cites, GPT-5.6 Sol got nine of 40 questions right on long agent logs, compared with 40 of 40 when given the current state.