File size: 2,840 Bytes
86415de 7420c83 86415de 3d17eb3 86415de 7420c83 81bb1aa 86415de 7420c83 86415de 7420c83 86415de 7420c83 86415de 3d17eb3 86415de 81bb1aa 86415de | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 | cff-version: 1.2.0
message: "If you use this work, please cite it as below."
title: "Semantic Strands: A Pre-registered Negative Result on Learned Soft-Slot Context Compression — Trained Compact Text Wins"
type: software
authors:
- family-names: "Sanchez"
given-names: "Erik"
affiliation: "Independent researcher"
orcid: "https://orcid.org/0009-0001-4908-3084"
email: "sanchezerik836@gmail.com"
version: "2026-10-16-rev5"
doi: "10.5281/zenodo.23448095"
identifiers:
- type: doi
value: "10.5281/zenodo.23448095"
description: "Zenodo DOI of the technical report"
- type: url
value: "https://huggingface.co/Eriksanchezolvera/cuerdas"
description: "Code, logs, sealed tests and checkpoints"
license:
- MIT
- CC-BY-4.0
repository-artifact: "https://huggingface.co/Eriksanchezolvera/cuerdas"
url: "https://huggingface.co/Eriksanchezolvera/cuerdas"
abstract: >-
Technical report (revision 5) presenting a pre-registered negative result.
A 3.84M-parameter encoder maps four-field structured records to K=2 soft
input slots for a frozen SmolLM2-135M-Instruct (17 vs 35 positions per
event against a verbal template). On unseen combinations of known words it
matches the template on two hash-sealed tests evaluated once (99.0 vs
99.25%; 98.7 vs 98.5%). It fails on unseen values (29% vs 98% with four
unseen names, phrases or multi-token values), a larger training pool does
not help, and compact text with a matched trained soft prompt (3.84M
parameters) scores 100% at 22 (+8 shared) positions. Probes read the slots
better than the frozen LLM. Recommendation: compare soft-slot encoders
against compact text with a matched trainable budget before claiming
compression.
keywords:
- soft prompts
- context compression
- prompt compression
- frozen language model
- knowledge distillation
- SmolLM2
repository-code: "https://huggingface.co/Eriksanchezolvera/cuerdas"
preferred-citation:
type: report
title: "Semantic Strands: A Pre-registered Negative Result on Learned Soft-Slot Context Compression — Trained Compact Text Wins"
authors:
- family-names: "Sanchez"
given-names: "Erik"
affiliation: "Independent researcher"
orcid: "https://orcid.org/0009-0001-4908-3084"
email: "sanchezerik836@gmail.com"
year: 2026
month: 10
status: preprint
doi: "10.5281/zenodo.23448095"
publisher:
name: "Zenodo"
notes: "Technical Report, revision 5 (updated 18 October 2026). Not peer reviewed."
url: "https://huggingface.co/Eriksanchezolvera/cuerdas"
license: CC-BY-4.0
references:
- type: software
title: "SmolLM2-135M-Instruct"
authors:
- name: "Hugging Face"
url: "https://huggingface.co/HuggingFaceTB/SmolLM2-135M-Instruct"
commit: "12fd25f77366fa6b3b4b768ec3050bf629380bac"
license: Apache-2.0
|