Show HN: 在单张 46GB 显卡上实现 600 万 Token 的滑动窗口
Show HN: A 6M-token movable window on a single 46GB GPU

原始链接: https://arxiv.org/abs/2607.23806

论文《A Frozen 12B Beats Frontier Models on Verified Work》引入了一种人工智能效率的范式转移:研究人员不再重新训练模型,而是利用持久的验证性记忆系统来绕过重复计算。 通过存储针对特定问题系列的已验证解决方案,该系统使模型能够以零生成标记(zero generation tokens)检索答案,从而确保 100% 的准确性和确定性的位级精确结果。在 180 个测试案例中,这种方法使一个规模适中的 12B 模型获得了满分,有效地将高阶能力与参数规模解耦。 该系统采用了优于传统基于相似度方法的精确寻址记忆检索流程,而传统方法往往会失效。此外,其记忆存储支持高达 600 万个标记的“可移动窗口”,与 vLLM 等标准引擎相比容量巨大。 尽管前沿模型在原始的从头推理方面仍具优势,但该架构为重复性任务提供了一种变革性的替代方案。通过消除冗余的生成过程,它不仅显著节省了成本和能源,还将冻结模型转化为一种用于复杂推理和解决问题的高性能、可验证引擎。

抱歉。
相关文章

原文

[Submitted on 26 Jul 2026]

View a PDF of the paper titled A Frozen 12B Beats Frontier Models on Verified Work: 100% Accuracy, 0 Tokens, Bit-Exact, Forever, by Sietse Schelpe (Corbenic AI)

View PDF
Abstract:Improving a language model today means retraining it: enormous compute, a new opaque model each cycle, non-deterministic output. We take the opposite path: the model stays frozen, and a persistent memory of verified solutions grows beside it. Once a problem family is solved and has passed an independent verification step that never consults the answer key, every new instance of that family is answered at zero generation tokens, bit-exact, deterministically. Across 180 fresh instances spanning nine problem families, four architectures from four vendors - dense and mixture-of-experts - each score 180/180 at zero generation tokens per answer: execution-bound capability decoupled from parameter scaling. A negative control attributes the capability fully to the memory: emptied, it solves nothing. The same verify-before-store contract holds for open-ended reasoning: 88/88 consistency-gated acceptances across all four models, machine-checked formal proof, and reasoning-method transfer at 77/80. Memory selection takes 1.4 microseconds; a full reuse completes in 6-23 ms at 36 mWh. Approximate similarity retrieval selects the wrong item 94.3% of the time on a 4,500-item verified store where exact addressing makes zero errors. The store also serves as working context at a scale no shipped engine matches: a 6,000,000-token movable window on a single 46 GB GPU at flat memory, where vLLM stops at 30,399 tokens and SGLang silently truncates past 32,000. On published benchmarks, frontier models remain far ahead of any 12B at raw from-scratch reasoning; on everything this system has solved and verified, the comparison inverts: a frontier API call pays a fresh generation pass on every query, forever, while verified reuse costs zero tokens and returns the identical bits every time. A public testbench with free, rate-limited access accompanies this report: this https URL
From: Sietse Schelpe [view email]
[v1] Sun, 26 Jul 2026 19:08:33 UTC (1,421 KB)
联系我们 contact @ memedata.com