arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2610.10118cs.LGcs.CL

YANchor-4B:O(N) 时间与 O(1) 内存下的高效长时程推理

YANchor-4B: Effective Long-Horizon Reasoning in O(N) Time with O(1) Memory

Huishan Ji, Hua Xu, Weiming Zhang, Qirui Ye

首次发表
浏览论文内容

中文总结 AI 辅助

YANchor-4B 是一种通用循环模型,通过锚点记忆机制在 O(N) 时间和 O(1) 内存下实现高效长时程推理,在数学基准上显著优于同类模型,并提升长序列生成吞吐量。

中文摘要 AI 辅助

长时程推理要求在可控的生成成本下访问早期信息。全历史注意力机制会导致存储和计算量不断增长,而循环压缩则可能丢失精确细节。为此,我们提出了 YANchor-4B,一个通用的循环模型,它将关键记忆保留为锚点(ANchors),供后续推理时检索。除了 O(N) 时间的生成和 O(1) 的内存占用,YANchor 还通过其多维记忆机制实现了有效的长时程推理。例如,在具有挑战性的数学问题上,它在 AIME 2024--2026 上实现了 82.93% 的平均 pass@1,在 HMMT 上实现了 63.64%,显著优于包括更大模型在内的线性时间、恒定状态模型。此外,在 H100 上,它的批量长序列生成吞吐量比 Transformer 和混合基线高出数倍。最后,在数十个基准上的评估证明了 YANchor 在通用能力上的优越性。

英文摘要

Long-horizon reasoning demands access to earlier information at a manageable generation cost. Full-history attention incurs growing storage and computation, while recurrent compression can lose precise details. Therefore, we present YANchor-4B, a general-purpose recurrent model that preserves crucial memory as ANchors for retrieval during subsequent reasoning. Beyond $O(N)$-time generation and $O(1)$ memory, YANchor enables effective long-horizon reasoning through its multidimensional memory mechanism. For example, on challenging math problems, it achieves 82.93% mean pass@1 on AIME 2024--2026 and 63.64% on HMMT, substantially outperforming linear-time, constant-state counterparts, including larger models. It also delivers several-fold higher batched long-generation throughput than Transformer and hybrid baselines on H100. Furthermore, evaluations across dozens of benchmarks demonstrate YANchor's superiority in general-purpose capabilities.

发表机构

  • Rocore Matrix

机构由 AI 辅助整理,请以论文原文为准。

补充信息

↑