研究人员提出的递归语言模型（Recursive Language Models, RLMs）通过引入Python交互环境，让模型学会写代码去分块读取、检索并递归调用自身处理信息，成功在千万级Token的超长文本任务中实现了性能与成本的双重突破。

你有没有发现，你让AI读一篇长文章，结果它读着读着就忘了前面的内容？ 你让它处理一份超长的文档，结果它给出来的答案，牛头不对马嘴？ 这个现象，学术界有个专门的名词，叫做上下文腐化。 这也是目前AI的通病：大模型的记忆力太差了，文章越长，模型越傻！

传统基础模型在S-NIAH单针大海捞针等简单检索任务中尚能维持表现，但在信息密度更高的复杂任务中，其推理性能随输入长度增加而下降。相比之下，RLM在输入长度超过特定阈值区间后，依然保持得分稳定性。

2025年的最后一天， MIT CSAIL提交了一份具有分量的工作。当整个业界都在疯狂卷模型上下文窗口（Context ...

AI data trainers who ensure the accuracy and viability of training data going into AI models are well-compensated, in-demand professionals. Two new studies projected potential annual incomes ranging ...

Run oprn source Chatterbox on CPU or GPU with Python 3.11 with watermarking support, giving creators fast, traceable voice ...

Chatbots can be overly agreeable. To get less agreeable responses, ask for opposing viewpoints, multiple perspectives, and a ...

High school sophomore Abigail Merchant has made it her mission to use technology to reduce flood-related deaths. The ...

Geoffrey Hinton warns AI’s rapid advance could eliminate millions of jobs by 2026, putting white-collar work, regulation, and ...

The SAVE Act if signed into law would require states to collect and document proof of citizenship from each voter, in the ...

Which cyber defenses actually work against current threats? An analysis of cyber-insurance claims data shows six technologies ...

With a new year upon us, software-security experts disagree on SBOM utility — in theory, SBOMs are great, but in practice, ...