license: cc-by-4.0
language:
- en
pretty_name: SYNC per-row experiment records
size_categories:
- 1M<n<10M
tags:
- language-models
- multi-agent-communication
- interpretability
SYNC: per-row experiment records
Per-row records behind the paper Share the Symbol, Not the Vector: Training-Free Fact Communication Across Language Models (Xin Li, Mengbing Liu, Hao Jiang, Xin Gao, Annan Wang, Xingwei Qu, Yichi Zhang, Chau Yuen).
- Paper: https://arxiv.org/abs/2610.19696
- Project page: https://lixin.ai/SYNC/
- Code: https://github.com/LiXin97/SYNC
Contents
1.6 million rows in 322 JSON-lines files (2.3 GB). One row per scene, world, question and condition: the receiver, the message source and arm, the write configuration, the parsed answer, correctness, coverage, the generated text and the message content. To fit the size limit, token ids, slot tables, per-layer patch reports and most cache/logit hashes were removed from the full rows; every point estimate in the paper can be recounted from these rows, and the engineering checks that need the removed fields are recorded in the analysis outputs of the code repository. Rows of the Countries task omit message contents.
The folder layout mirrors the code repository, so the files land where its scripts expect them:
| Folder | Study |
|---|---|
artifacts/latest0924/expA/ |
new senders and receivers (Table 1) |
artifacts/latest0924/predict/ |
depth prediction (Table 2) |
artifacts/latest0924/layout/, artifacts/latest0924/realtask/ |
new layouts and the Countries task (Table 3) |
artifacts/latest0924/graded/, artifacts/latest0924/graded_textfmt/ |
graded weights and other text formats (Figure 6) |
artifacts/latest0924/diag/ |
late-writing diagnostics |
artifacts/receiver_recovery/ |
development receivers (Figures 4-5) |
artifacts/word_interface_followup/ |
four-object questions and simpler representations |
Usage
git clone https://github.com/LiXin97/SYNC.git && cd SYNC
hf download XINLI1997/SYNC-data --repo-type dataset --local-dir .
python scripts/latest0924/layout/recount_layout.py
python scripts/latest0924/predict/score.py
How the data were made
Every scene is generated mechanically from fixed word lists and a seeded random generator; no third-party text and no personal data are included. The Countries task follows the published protocol of Ramesh & Li (2025) and is rebuilt from hand-written lists of landmarks and countries. The model outputs were produced by open-weight models (Qwen, gpt-oss, Gemma, Llama, GLM, Mistral, Granite, OLMo and Seed-OSS checkpoints); text generated by a model remains subject to that model's license.
License
CC BY 4.0.
Citation
@misc{li2026sync,
title = {Share the Symbol, Not the Vector: Training-Free Fact Communication Across Language Models},
author = {Xin Li and Mengbing Liu and Hao Jiang and Xin Gao and Annan Wang and Xingwei Qu and Yichi Zhang and Chau Yuen},
year = {2026},
eprint = {2610.19696},
archivePrefix = {arXiv},
primaryClass = {cs.CL},
url = {https://arxiv.org/abs/2610.19696}
}