arXiv cs.CL· Huishan Ji, Hua Xu, Weiming Zhang, Qirui Ye·· 10 天前AI 评分46
YANchor-4B：以 O(N) 时间和 O(1) 内存实现高效长程推理
YANchor-4B: Effective Long-Horizon Reasoning in O(N) Time with O(1) Memory
AI 导读
YANchor-4B 是一个通用循环模型，通过将关键记忆保留为 ANchor 供后续推理检索，实现 O(N) 时间生成与 O(1) 内存。在 AIME 2024–2026 上取得 82.93% 平均 pass@1，HMMT 上达 63.64%，大幅超越线性时间、恒定状态的同类模型（包括更大模型）。
正文 · 原文
Abstract: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.
| Comments: | 24 pages, 8 figures. Code: this https URL ; Model: this https URL |
| Subjects: | Machine Learning (cs.LG); Computation and Language (cs.CL) |
| Cite as: | arXiv:2610.10118 [cs.LG] |
| (or arXiv:2610.10118v1 [cs.LG] for this version) | |
| https://doi.org/10.48550/arXiv.2610.10118 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Huishan Ji [view email]
[v1]
Wed, 7 Oct 2026 14:03:50 UTC (1,480 KB)
来源：arXiv cs.CL · arxiv.org