未来可预测吗？面向可靠时间推理的未来可预测性建模
原标题：Is the Future Predictable? Modeling Future Predictability for Reliable Temporal Reasoning
AI 摘要
研究提出 ChronoRegime 任务，要求模型判断历史时间证据是否足以支持对未来事实做出预测，并据此提出 Confidence-Aware Temporal Reasoning（CTR）框架，从时间动态中估计未来可预测性并调节回答承诺。在 LLaMA-3.1、Qwen3、GPT-4o 等模型上，CTR 相比普通问答基线平均提升严格总体准确率 6.7 个百分点，降低幻觉率 18.4 个百分点。
正文节选
Is the Future Predictable? Modeling Future Predictability for Reliable Temporal Reasoning Abstract Large language models (LLMs) remain unreliable when reasoning about the future. Existing models often overlook the difference among future states, committing to specific answers even when the queried facts are inherently volatile. To capture this challenge, we introduce ChronoRegime, a future-oriented temporal reasoning task that requires models to determine whether historical temporal evidence jus