SYNC-directions: fitted per-receiver word directions

This archive holds the word directions used in the paper "Share the Symbol, Not the Vector: Training-Free Fact Communication Across Language Models".

Each receiver model gets its own set of directions, fitted only from that receiver's own activations; no other model is involved in the fit. A direction for symbol w at layer l is the mean of 12 contrasts. Each contrast is the receiver's residual-stream state at the final token of w in a short context, minus the state at the same position in a token-length-matched neutral context. The 12 contexts cover list widths 2 and 4 and every object/rank position. The archive contains one direction per symbol per write layer. No communication outcome was used to fit any file (receipts carry "no_communication_outcomes": true).

Layout (paths are relative to this folder)

Folder	Experiment	Receivers	Symbols	Write layers per receiver
latest0924/fits/	New receivers (Table 1)	gpt-oss-20b, granite-4.1-8b, mistral-small-3.2-24b, olmo-3-7b-instruct, qwen3.5-2b, qwen3.5-4b, qwen3.5-9b, qwen3.5-27b, qwen3.5-122b-a10b, qwen3.6-27b, qwen3.6-35b-a3b, seed-oss-36b-instruct (12)	221 objects	4
latest0924/predict/fits/	Depth-prediction study	gpt-oss-20b, granite-4.1-8b, mistral-small-3.2-24b, olmo-3-7b-instruct, qwen3.5-2b, qwen3.5-4b, qwen3.5-9b, qwen3.5-27b, qwen3.6-27b (9)	221 objects	5 (a depth sweep)
latest0924/realtask/fits/	Landmark (Countries) task	mistral-small-3.2-24b, qwen3-32b, qwen3.5-27b, qwen3.6-27b (4)	275 landmarks (multi-token names)	4
receiver_recovery/fits/	Development receivers	gemma-3-27b-it, glm-4-9b, qwen3-14b-base, qwen3-32b (4)	221 objects	7
word_interface_followup/directions/	Four-object / representation follow-up	gemma-3-27b-it, glm-4-9b, qwen3-14b-base, qwen3-32b (4)	221 objects	1

Receiver names are the short identifiers used throughout the paper. Each maps to the public checkpoint of the same name, for example glm-4-9b = GLM-4-9B-0414 and qwen3-14b-base = Qwen3-14B-Base.

The first four folders share one file layout:

• directions_<receiver>.pt: all symbols for one receiver, aggregated into one file. Use this one.
• <receiver>/wNNN.pt: per-symbol fit files (NNN = symbol address), which include the 12 individual contrasts.
• scale_<receiver>.json: per-layer local scale c[l] (median over symbol and template of ‖delta‖/‖d‖, ≥ 1 by construction), the median direction norm per layer, and the number of ratio samples.
• fit_<receiver>_s<k>.done.json: fit receipts for shard k, one row per symbol (address, path, sha256), plus layers, elapsed time and status.

Loading

import torch

d = torch.load("latest0924/fits/directions_qwen3.5-9b.pt", map_location="cpu", weights_only=False)
d["directions"]  # float32 tensor [n_layers, n_words, hidden]
d["words"]       # list[str], length n_words; index i corresponds to directions[:, i]
d["layers"]      # list[int], write-layer indices; index j corresponds to directions[j]
d["receiver"], d["n_words"], d["construction"], d["source"]  # metadata; source = repo-relative folder of the per-word fits

v = d["directions"][d["layers"].index(13), d["words"].index("accordion")]  # one direction

Per-symbol file <receiver>/wNNN.pt (dict):

• direction: float32 [n_layers, hidden]. Equals the matching slice of directions_<receiver>.pt.
• contrasts: float32 [12, n_layers, hidden]. The 12 per-context contrasts that direction averages.
• word, address, receiver, layers, construction, manifest_sha256: metadata.
• contexts: list of 12 dicts. Each holds source and neutral (receiver token-id lists for the word context and the matched neutral context), position (the token position read out), word_positions, width, object and index.

Follow-up file word_interface_followup/directions/p0a_directions_<receiver>.pt (dict):

• constructions: dict of float32 [221, hidden] at a single layer:
  • c12: the 12-context mean.
  • single_t0, single_t6, single_t11: single-context directions from context 0, 6 or 11.
  • single_t*_nm: the same single-context directions, norm-matched to c12 by the scalars in nm_scalars.
• words, layer, receiver, single_context_positions, fitting_contexts, source_fits: metadata.

p0a_directions_report.json summarizes norms, cosines to c12, and fitting positions for each receiver.

Sanitization and integrity

The internal absolute paths in metadata strings were rewritten. Paths in this archive are relative to the repository (artifacts/<folder>/...), and the archive root corresponds to artifacts/. Host names were replaced with gpu-node. No numerical content was changed.

• Per-symbol wNNN.pt files contain no path strings. They are byte-identical copies of the originals, so the sha256 values in the fit_*.done.json receipts still verify against them.
• The 33 files directions_<receiver>.pt and p0a_directions_<receiver>.pt were re-saved with their source / source_fits string rewritten. Every tensor was checked against the original with torch.equal (same dtype and shape) and matched.
• The sha256 fields inside scale_*.json and p0a_directions_report.json are historical records. They pin the original (pre-sanitization) files. SANITIZED_DIGESTS.json maps every file's original sha256 to its sanitized sha256, so those pins can still be followed.
• A few receipts carry a verified_against field. It names a file from an earlier fit that is not included in this archive.

Licenses

These files are derived from the base models (fitted from their activations). Each file is governed by the license of the base model it was fitted from:

Base model (receiver)	License	File
Qwen3 / Qwen3.5 / Qwen3.6 (all qwen* receivers)	Apache-2.0	licenses/LICENSE-Apache-2.0.txt
gpt-oss-20b	Apache-2.0	licenses/LICENSE-Apache-2.0.txt
Mistral-Small-3.2-24B	Apache-2.0	licenses/LICENSE-Apache-2.0.txt
granite-4.1-8b	Apache-2.0	licenses/LICENSE-Apache-2.0.txt
OLMo-3-7B-Instruct	Apache-2.0	licenses/LICENSE-Apache-2.0.txt
Seed-OSS-36B-Instruct	Apache-2.0	licenses/LICENSE-Apache-2.0.txt
GLM-4-9B-0414 (glm-4-9b)	MIT	licenses/LICENSE-MIT-GLM.txt
gemma-3-27b-it	Gemma Terms of Use	licenses/GEMMA_NOTICE.txt

The files derived from gemma-3-27b-it are Model Derivatives under the Gemma Terms of Use. Gemma is provided under and subject to the Gemma Terms of Use found at ai.google.dev/gemma/terms. The Prohibited Use Policy (ai.google.dev/gemma/prohibited_use_policy) applies to these files and to any redistribution of them.

No file in this archive derives from a Llama model, so the Llama license does not apply.

Paper & links

Share the Symbol, Not the Vector: Training-Free Fact Communication Across Language Models

• Paper: https://arxiv.org/abs/2610.19696
• Project: https://lixin.ai/SYNC/
• Code: https://github.com/LiXin97/SYNC
• Model (this repository): https://huggingface.co/XINLI1997/SYNC-directions

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}
}