[论文] How far does a hash-only sender go? A pre-registered map of sender mem...
研究领域: NLP 作者: Marcelo Samoilenko 发布时间: 2026-11-06 arXiv: 2611.09848
论文概要
研究领域: NLP 作者: Marcelo Samoilenko 发布时间: 2026-11-06 arXiv: 2611.09848
中文摘要
仅计算哈希的发送器(接收端只有语言模型)能否比在同领域上训练的经典发送器更便宜地传输短文本?作者在 Taskmaster-1 的1000条英语用户话语上做了预注册单次实验。主要终点是投递100%消息的单条消息成本(哈希位 + 首次失败时的2字节否定确认和原始文本),与域内 PPM 一阶发送器对比。哈希发送器在首次尝试中成功投递857/1000条消息,平均成本192.8位/消息,比 PPM 一阶多18.4位(95% bootstrap 区间10.9到26.3);按预注册规则,它更昂贵且不等价。次要结果绘制了实测发送器内存与比特的关系。哈希发送器(主机构建8,190字节,含键控BLAKE2b标签)在8,190字节起直至PPM一阶(24,844字节)前的比特成本最低,比 smaz2 和 Unishox2 低45-46比特;从 PPM 一阶起,域内发送器比特更少(除剪枝至16KiB的二阶外)。在空中时间(airtime)上,它在五种定价链路上都输给 PPM 一阶,在四条上输给 Unishox2,在两条上输给带标签的原始文本——因为每次失败都要付出确认帧和回退帧的代价。更宽的接收器搜索能恢复一半失败,使成本比 PPM 一阶低7.7位(每次失败约3分钟接收器时间),但在五条链路中仍有四条在空中时间上落败。作者以边界划定的形式报告测量结果及运行中的所有偏差。
原文摘要
Can a sender that only computes hashes, with a language model only at the receiver, transmit short human text more cheaply than a classical sender trained on the same domain? We answer with one pre-registered run on 1,000 English user utterances from Taskmaster-1. The primary endpoint is the cost per message to deliver 100% of the messages (hash bits, plus a 2-byte negative acknowledgement and the raw text when the first attempt fails), paired against an in-domain PPM order-1 sender. The hash-only sender delivered 857 of 1,000 messages on the first attempt and cost 192.8 bits per message, 18.4 bits more than PPM order 1 (95% bootstrap interval 10.9 to 26.3); by the pre-registered rule it spends more and is not equivalent within one byte. As secondary results we map measured sender memory a...
*自动采集于 2026-11-10*
#论文 #arXiv #NLP #小凯