The multilingual pivot pilot
Does SONAR translate French into an “English embedding” before decoding Chinese?
Exploratory experiment run on 4l, 9 September 2026. Direct French→Chinese works without an explicit English-text intermediate. The experiment does not identify an internal English pivot, and it does not establish that adding one improves translation. It does show that an English decode/re-encode is a real transformation of the vector, rather than an identity operation.
This is the completed first step of T05 in the consolidated things-to-do document. The proposed learned embedding-to-embedding translator and independently annotated multilingual benchmark have not been run.
What ran
The retained text_sonar_basic_encoder and text_sonar_basic_decoder, in float32, on one NVIDIA RTX A4000. Twenty-four assistant-authored French/English/Chinese triples cover ten ordinary contrast pairs—role, negation, number, time, spatial relation, ownership, comparison, quantification, relation and modality—plus two tu/vous pairs. No training, paid API or new model installation was needed.
Every source language was decoded into every target language: 9 routes × 24 examples. French inputs also went through an extra French, English or Chinese text round trip before final Chinese decoding: another 72 generations. Each run repeated its first direct French→Chinese item as a reproducibility check. Greedy decoding and beam size 5 each produced 289 outputs; both repeat checks passed. Recorded run times were 29.70 and 44.87 seconds including model loading, excluding process startup and file transfers.
The “Chinese decoder” and “English decoder” boxes are the same multilingual decoder with different target-language settings. There is no fitted vector translator in these routes. French and Chinese intermediate round trips control for the extra decode/re-encode step.
Observed results
| Comparison | Greedy | Beam size 5 |
|---|---|---|
| Direct Chinese exactly matches Chinese via English | 13/24 | 14/24 |
| Direct Chinese exactly matches Chinese via French | 21/24 | 23/24 |
| Direct Chinese exactly matches Chinese via Chinese | 15/24 | 18/24 |
| Mean cosine: original French vector vs English-pivot vector | 0.920 | 0.918 |
| Mean cosine: original French vector vs French-round-trip vector | 0.981 | 0.992 |
| Mean cosine: original French vector vs Chinese-pivot vector | 0.766 | 0.769 |
Exact agreement is a route-consistency diagnostic, not translation accuracy. Several differing strings are valid paraphrases or name transliterations. A French round trip also has a structural advantage on this measure because many French reconstructions reproduce the input exactly. These counts do not establish that direct decoding is more faithful than an English pivot.
For the supplied reference translations, mean paired embedding cosine was 0.918 French–English, 0.655 French–Chinese and 0.654 English–Chinese. In each of those three directed comparisons, all 20 ordinary contrast items ranked their intended equivalent above its paired wrong-meaning alternative. These are just ten hand-authored contrast families, not 60 independent benchmark successes. The four tu/vous items were excluded from this ranking summary because the supplied English strings deliberately collapse each pair.
The lower Chinese cosine could reflect language, tokenization, phrasing, translation-reference choices and model geometry. It is not evidence that French is encoded “in English.” The references were not independently checked, and raw language strings were passed to the standard SONAR pipeline without external Chinese segmentation.
Concrete examples
| Input / contrast | Output observed | What it supports |
|---|---|---|
Le chat poursuit le chien. / Le chien poursuit le chat. |
Direct Chinese: 猫追着狗. / 狗追着猫. in both runs; English-pivot outputs agree |
These two opposing roles survive direct translation through one vector. This is not the held-out global-role probe from H1. |
| Three versus five books on a table | Direct and English-pivot Chinese retain 三本书 versus 五本书 in both runs |
The number distinction survives these routes on this pair. |
Tu peux partir maintenant. |
Greedy direct: 你可以现在走.; via English: 你可以去. |
The greedy English-pivot output drops “now.” Beam size 5 restores “now” in both routes, so this is decoding-sensitive, not a robust route disadvantage. |
Pouvez-vous fermer la porte ? |
Greedy English: Can you close the door?; beam-5 English: Can you close the door, please? |
English can express politeness through wording. The premise that an English pivot must always erase formal address is too strong. |
La sœur de Paul connaît Marie. |
Both Chinese routes use 保罗的妹妹认识玛丽. |
Both choose “younger sister,” although French does not specify age. A shared decoder interpretation can add detail absent from the source. |
| Sentences about children eating apples | Chinese generations contain ⁇ 果 in both runs |
Output corruption remains even with beam search. Investigate decoder/tokenizer behavior before using Chinese outputs as a clean semantic oracle. |
| Red versus blue bicycles | Direct Chinese often uses generic 车; English-pivot output sometimes restores 自行车 |
An English pivot can improve an explicit detail in some examples; the observed effects are mixed. |
These interpretations are the assistant's inspection of saved outputs, not blinded bilingual human scores. The sister and ownership examples illustrate that languages differ in what they specify: “son livre” does not uniquely identify the owner's gender, and Chinese sibling terms can force choices absent in French/English. The isolated French vous examples also permit plural or formal-singular readings. Do not score one hand-authored Chinese reference as the only correct answer.
What this says about the hypothesis
- The functional route exists: French text can go through SONAR's sentence vector straight to Chinese. This matches the official multilingual encoder/decoder interface.
- An English text pivot changes the representation: its mean cosine to the original is below 1, and its Chinese decodes often differ. In the greedy tu/vous examples, identical generated English strings necessarily produce identical English-pivot vectors, illustrating a text bottleneck's potential to merge distinctions.
- An internal English-centered computation remains untested: translation success, pairwise similarity and route agreement do not distinguish a shared semantic code from English-associated internal features or other mixed representations. There is no unique “language of a vector” identifiable from these measurements alone.
- A learned vector translator is a separate hypothesis: it might align language-associated distributions or improve decoder compatibility, but it might also delete useful distinctions. It needs enough training pairs, held-out meanings and an identity-map baseline.
The useful next experiment
Build a disjoint, independently checked parallel corpus with contextualized information-loss contrasts. Keep English, French and Chinese pivots, but add another non-English pivot before calling English special. Compare identity, mean offset, orthogonal Procrustes and regularized linear maps from French vectors to English-reference vectors, and compare a Chinese-reference target map too. Select settings only on validation. Score target-language meaning and unsupported details independently of SONAR cosine; also measure geometric alignment and reconstruction.
For an internal-computation claim, add layerwise language/meaning readouts under held-out content and calibrated interventions, with random-direction and semantic-damage controls. A language-identification probe alone only establishes readable language information; it does not establish translation through that language.
Artifacts and reproducibility
- Greedy results, all routes and embeddings.
- Beam-5 results, all routes and embeddings.
- Readable direct/pivot generation tables.
- Current runnable script; exact archived greedy-run script. Their SHA-256 hashes match their respective result metadata.
Remote working directory: /workspace/HOME/guest/tae_status_2026_09_09/. Environment: /workspace/HOME/guest/night8/.venv/bin/python, Torch 2.9.1, sonar-space 0.5.0, fairseq2 0.8.1. Physical GPU 1 was selected using CUDA_VISIBLE_DEVICES=1; the program sees it as cuda:0. Batch size 8, maximum generation length 64, seed 0. Encoder/decoder asset names and package versions are recorded; this pilot did not checksum the large pretrained weight files or establish compatibility across other package versions.
CUDA_VISIBLE_DEVICES=1 HF_HOME=/workspace/hfcache OMP_NUM_THREADS=4 \
/workspace/HOME/guest/night8/.venv/bin/python \
/workspace/HOME/guest/tae_status_2026_09_09/multilingual_pilot.py \
--beam-size 5 --out /workspace/HOME/guest/tae_status_2026_09_09/multilingual_beam5_repeat
QwenAR and LatentSeal were shortlisted from public releases, not executed in this pilot. Their immediate value is as alternative TAE substrates; multilingual suitability remains to be tested. See T09 in the consolidated plan.