答对得 1 分,回答“不知道”得 0 分。答错扣 c 分时,只有把握超过 c ÷ (1 + c) 才值得作答。思路见 Kalai 等(2025)。
六、信息太多时,关键内容被忽略
给代理更多的上下文,并不总是更好。刘等人(2024)的实验发现,当关键信息位于一长串材料的开头或结尾时,模型能较好地利用它;位于中间时,表现明显下降,呈现出 U 形。上下文越长,这种现象越明显。对金融任务而言,一份年报、几十条新闻与一年的交易记录全部放进上下文,关键的一条可能恰好被淹没。更可靠的做法,是先按任务筛选出相关的信息,再交给模型处理,并要求它标明每个结论来自哪一处材料。
Kwa, T., West, B., Becker, J., et al. (2025). Measuring AI ability to complete long tasks. arXiv:2503.14499(本书引用 v1,2025;v3 起题为 Measuring AI ability to complete long software tasks).
METR (2026). Task-completion time horizons of frontier AI models (METR-Horizon-v1.1). https://metr.org/time-horizons/
Dell'Acqua, F., McFowland, E., Mollick, E. R., et al. (2023). Navigating the jagged technological frontier: Field experimental evidence of the effects of AI on knowledge worker productivity and quality. Harvard Business School Working Paper 24-013.
Kalai, A. T., Nachum, O., Vempala, S. S., & Zhang, E. (2025). Why language models hallucinate. arXiv:2509.04664.
Liu, N. F., Lin, K., Hewitt, J., Paranjape, A., Bevilacqua, M., Petroni, F., & Liang, P. (2024). Lost in the middle: How language models use long contexts. Transactions of the Association for Computational Linguistics, 12, 157–173.
Schick, T., Dwivedi-Yu, J., Dessì, R., et al. (2023). Toolformer: Language models can teach themselves to use tools. Advances in Neural Information Processing Systems, 36.
Gao, L., Madaan, A., Zhou, S., et al. (2023). PAL: Program-aided language models. Proceedings of the 40th International Conference on Machine Learning, PMLR 202, 10764–10799.