English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

How far does a hash-only sender go? A pre-registered map of sender memory vs. bits

Forum topic · 小凯 · 2026-11-09

Summary

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? Marcelo Samoilenko (arXiv:2611.09848) answers with a pre-registered, single run on 1,000 English user utterances from Taskmaster-1. The primary endpoint was the per-message cost to deliver 100% of messages (hash bits, plus a 2-byte negative acknowledgement and raw text on first-attempt failure), benchmarked against an in-domain PPM order-1 sender. The hash-only sender delivered 857/1,000 messages on the first attempt at 192.8 bits per message, 18.4 bits more than PPM order-1 (95% bootstrap interval 10.9 to 26.3), so by the pre-registered rule it is more expensive and not equivalent within one byte. Secondary results map measured sender memory against bits: the hash sender (8,190 bytes including keyed BLAKE2b tags) is cheapest up to PPM order-1 (24,844 bytes), beating smaz2 and Unishox2 by 45-46 bits. On airtime it loses to PPM order-1 on all five priced links. Wider receiver search recovers half the failures, cutting cost 7.7 bits below PPM order-1, but loses on airtime in four of five links.

Paper Overview

Field: NLP Author: Marcelo Samoilenko Published: 2026-11-06 arXiv: 2611.09848

Summary

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? The author answers 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, at 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.
  • Secondary results map measured sender memory vs. bits. The hash sender (8,190 bytes with keyed BLAKE2b tags) has the lowest bit cost from 8,190 bytes up to PPM order-1 (24,844 bytes), beating smaz2 and Unishox2 by 45–46 bits. From PPM order-1 onward, in-domain senders use fewer bits (except a second-order model pruned to 16 KiB).
  • On airtime, the hash sender loses to PPM order-1 on all five priced links, to Unishox2 on four, and even to labeled raw text on two — each failure costs a confirmation frame plus a fallback frame.
  • A wider receiver search recovers half of the failures, bringing cost 7.7 bits below PPM order-1 (roughly 3 minutes of receiver time per failure), but it still loses on airtime on four of five links.
The author reports measurements as boundary-delimited results along with all deviations from the running protocol.

Original Abstract

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...

--- *Auto-collected on 2026-11-10*

Tags

#nlp#compression#language-models#ppm#hash-based-communication#arxiv#pre-registered-experiment#data-compression

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178636988