SRT Verbalizer v1 — a 0.6B reads a large model's record aloud

When a large vision-language model looks at a photograph, it forms an internal reading of it: a few thousand numbers, partway through the network. That record is not hidden and not mysterious. The SRT program's whole finding is that it is orderly enough to read with small, auditable instruments. What it is not, is in words.

This is the translator. A frozen Qwen3-0.6B, plus a ~44M-parameter prefix network, takes one such record and writes a sentence describing the photograph. The small model is text-only. Its entire input is the vector.

Measured on 5,000 held-out photographs, the sentence it writes retrieves the correct image out of 123,287 at median rank 20, against 39 for a human-written reference caption.

The record was always legible. This puts it in English.

What is in this repository

† Not comparable to the others. See Which checkpoint to use. ‡ A documented negative. See Negative results.

The result, and the controls that make it one

A plausible-sounding caption proves nothing: a model that learned COCO's caption prior would produce fluent sentences about nothing. So the arms below matter more than the headline. All are val2017 photographs, held out of both head training and verbalizer training, scored against the full 123,287-image gallery. Chance median is ~61,644.

Qwen3.8-27B layer 52 (qwen38_27b_L52.pt)

gemma-4-31B layer 47 (gemma4_31b_L47.pt)

Both controls sit at chance, and they are legible in the text itself:

• another image's record: the vectors are rolled by one position, and the captions come out rolled by one position. The words follow the vector.
• the mean record: one sentence for every input, every time. "A man is standing on a bench next to a basket." Fluent, plausible, about nothing. That is the floor a real reading has to clear.

The cross-model control

The gemma-4-31B checkpoint is scored by a gallery an unrelated Qwen3.8-27B tower built. No shared representation is available to carry that result, and it lands within five ranks of the matched pair. If a shared representation were doing the work rather than the sentence being descriptive, the matched arm should dominate. It does not.

Caveats that travel with the number

It is not better captioning than a human. It scores above a human reference for two measurable reasons, neither of which is superior description.

1. Register. The model enumerates whole-scene inventory, which is what this retrieval head recovers well (detection AUC 0.883 over 80 COCO categories). Human references foreground arrangement and oddity: "a woman stands in the dining area", "a stop sign mounted upside-down". That is exactly the information the head is documented not to carry. The humans encode more; the instrument reads less of it.
2. Length. The metric rewards naming more true things about a scene, and a human reference names one scene once. Given more tokens this model continues the reading and names more, and enough of that is correct to help. Forcing length artificially by repetition makes it worse (see below), so the verbosity that helps is genuine continued reading, not filler.

Gold is one reference caption, not the best of an image's five.

Scored in one instrument. These are retrieval numbers through one head and one gallery. A caption-similarity metric such as CIDEr would very likely favour the human references, and that would not contradict anything here.

Which checkpoint to use

• qwen38_27b_L52.pt if your records come from the tower that built the shipped gallery. Best number, single coherent system.
• gemma4_31b_L47.pt if you want the scientifically cleanest claim. Its scoring gallery shares no weights with the model being read, so it is the arm that rules out a shared-representation shortcut. This is the checkpoint the Sunstone Lab serves.
• browser_gallery_1024.pt for deployment where only the projected gallery vector is available, such as a browser holding an index. Every image in the gallery becomes describable with no extra download. Its median of 18 is not comparable to the others: the gallery row it reads is also the retrieval target, so the input carries the answer more directly than a hidden state does. The controls still discriminate and training was cross-entropy on captions rather than on retrieval, so the reading is real, but do not put that 18 in a table beside the 20 and 25.

Negative results

Kept because they were expensive to learn and they constrain what to try next.

Teaching it to stop costs retrieval. No training target originally contained an end-of-sequence token, because Qwen's tokenizer does not append one, so the model never learned that a caption ends and greedy decoding loops the last clause until the budget runs out. Appending EOS fixes the text completely: it stops on its own in 11–15 tokens, no repetition, no trimming needed. It also drops to median 46, below the human caption it used to beat. Saying its piece in eleven tokens names fewer objects. That checkpoint is included as gemma4_31b_L47_eos.pt because its prose is strictly better and some uses will prefer it.

Length by repetition is worse than brevity. Joining three of an image's five reference captions into one target gives length, and the model fills it by emitting the most likely caption three times: given only the vector it cannot know which three references are wanted, so the mode repeated is the lowest-loss output. Median 68. Abandoned at a third of training.

How it works

image ──► large model ──► hidden state (5376-d or 5120-d)
                              │
                              ▼
                    prefix MLP  d_in → 2048 → 16 × 1024      (the only trained part)
                              │
                              ▼
                16 soft tokens ──► frozen Qwen3-0.6B ──► a sentence

Training: cross-entropy on COCO captions for 118,287 train2017 images and 591,753 captions, 3 epochs, batch 32, lr 3e-4, bf16, ~30 min on one RTX PRO 6000. The backbone is frozen throughout; only the prefix learns.

Raw states are strongly anisotropic, so inputs are centred by the train-split mean and scaled by its mean radius. Those statistics ride along in the checkpoint (mu, sd) and must be reapplied at inference: feeding an uncentred vector silently degrades the output rather than failing.

Usage

import numpy as np, torch
from transformers import AutoModelForCausalLM, AutoTokenizer

ck  = torch.load("gemma4_31b_L47.pt", map_location="cpu", weights_only=False)
tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-0.6B")
lm  = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-0.6B",
                                           dtype=torch.float32).eval()

class Prefix(torch.nn.Module):
    def __init__(self, d_in, d_model, n_tok, hidden=2048):
        super().__init__()
        self.n_tok, self.d_model = n_tok, d_model
        self.net = torch.nn.Sequential(
            torch.nn.Linear(d_in, hidden), torch.nn.GELU(),
            torch.nn.Linear(hidden, n_tok * d_model))
    def forward(self, v):
        return self.net(v).view(v.size(0), self.n_tok, self.d_model)

pre = Prefix(ck["d_in"], ck["d_model"], ck["n_tok"], ck.get("hidden", 2048))
pre.load_state_dict(ck["prefix"]); pre.eval()

v = np.load("one_image_state.npy")              # (5376,) gemma-4-31B layer 47
x = torch.tensor((v - ck["mu"]) / ck["sd"],     # the frame it was fit in
                 dtype=torch.float32).unsqueeze(0)

with torch.no_grad():
    soft = pre(x)
    out = lm.generate(inputs_embeds=soft,
                      attention_mask=torch.ones(soft.shape[:2], dtype=torch.long),
                      max_new_tokens=32, do_sample=False,
                      pad_token_id=tok.eos_token_id)
print(tok.decode(out[0], skip_special_tokens=True).strip())

The non-EOS checkpoints will not stop on their own; trim to the last complete sentence for display, or use gemma4_31b_L47_eos.pt.

Try it

The Sunstone Lab serves this live at lab.sunstonenorth.com under 02 · Read an image: upload a photograph, gemma-4-31B encodes it, and this model says what is in the record beside the same reader given no record at all. Warm latency is about 2 s to encode and 2 s to read.

Note that the prefix was fit on CUDA/bf16 states and reads Apple-MLX/Q4 states from the Lab correctly with no adaptation, which is the program's substrate-invariance result appearing in a serving path.

Provenance

• Paper: paper_nla.md §11.8
• Code: scripts/train_shared_space_verbalizer.py, eval_shared_space_verbalizer.py, build_fullstate_pairs.py
• Every arm above is a banked JSON under artifacts/nla/verbalizer/
• Input states, already published, no re-encode required: srt-nla-gemma4-artifacts (procrustes/train_pairs/) and srt-qwen38-coco-states (raw118k/)
• Scoring head and gallery: srt-browser-head-118k

Citation

@software{lancaster_srt_verbalizer_2026,
  author = {Lancaster, Burton},
  title  = {SRT Verbalizer v1: reading a large model's record aloud with a 0.6B},
  year   = {2026},
  url    = {https://huggingface.co/RiverRider/srt-verbalizer-v1}
}