agmt-en-vi-20m

AGMT v3.2 --- Elastic Recurrent English ↔ Vietnamese Translator

AGMT (Adaptive/Elastic Recurrent Translation Architecture) is an experimental end-to-end English ↔ Vietnamese translation model built around shared recurrent refinement and translation-value routing.

The central idea is simple:

A translation model should not always use the same amount of computation.
It should refine only while additional latent computation is expected to improve the translation.

AGMT combines established components --- shared subword tokenization, embeddings, Transformer-style encoding/decoding, residual paths and autoregressive decoding --- with an experimental architecture built around recurrent parameter sharing, adaptive HALT/REFINE routing, lexical/base shortcuts, and a stable memory bridge.

This is therefore not a claim that every component is new. AGMT is partly a new architecture and partly a composition/reworking of existing ideas into one translation system.

Author

Trần Tuấn Phi --- Vietnam

• Email: [email protected] --- I do not check email frequently.
• Facebook: https://www.facebook.com/share/1HerWkmghN

---

What makes AGMT different?

A conventional fixed-depth encoder executes a predetermined stack for every input. AGMT instead has a non-recurrent base encoder followed by a shared recurrent refiner:

H_base
   │
   ▼
Shared Refiner ──► H1
   ▲               │
   └───────────────┘
        repeat

The same refiner parameters are reused at multiple refinement steps.

Therefore:

stored parameter count ≠ effective computation depth

Increasing r does not add a new set of model weights. It runs the shared refinement transformation again on the current hidden state.

Importantly:

larger r ≠ automatically better translation

A later refinement can improve, preserve, or occasionally degrade a translation. The router exists to decide whether another refinement step is expected to have enough translation value to justify its compute cost.

---

Architecture

SOURCE
  │
  ▼
Shared Byte-Safe SentencePiece Tokenizer
  │
  ▼
Shared Embedding
  │
  ├────────────────────────► Lexical Highway
  │
  ▼
Prelude Encoder
  │
  ▼
H_base
  │
  ▼
Shared Recurrent Refiner
  │
  ├────► Translation-Value Router
  │          ├─ HALT
  │          └─ REFINE
  │
  └──── recurrent refinement
  │
  ▼
H_refined
  │
  ├──────── H_base
  ├──────── E_lex
  ▼
Stable Memory Bridge
  │
  ▼
Coda Adapter
  │
  ▼
Shallow Autoregressive Decoder
  │
  ▼
TRANSLATION

Shared tokenizer and embedding

AGMT uses one tokenizer and a shared representation space for both translation directions. Direction tokens select the target language.

The tokenizer is embedded directly inside translate.py as a Base64-encoded serialized SentencePiece model. Do not delete _SPM_B64 unless you also change the loader to obtain the exact same tokenizer from an external file.

The embedded tokenizer keeps the release self-contained: translate.py reconstructs the SentencePiece processor at startup before loading the ONNX graphs.

Lexical Highway

The lexical path preserves source-surface evidence such as names, numbers, technical identifiers and other information that should not have to survive every recurrent transformation.

Prelude Encoder

The Prelude creates H_base, the contextual foundation before recurrent refinement. H_base is also retained as a stable shortcut for the decoder-side memory interface.

Shared Recurrent Refiner

The refiner reuses the same learned transformation at every recurrent step. A step ID tells the shared block which refinement stage it is executing.

Conceptually:

H_(r+1) = Refiner(H_r, step=r)

r is therefore the number of recurrent refinement steps actually executed, not a confidence score and not a separate reasoning model.

Translation-Value Router

For fast and balanced inference, the router runs before the next refiner step. If the predicted remaining value is not above the active compute price, AGMT halts without performing speculative extra refinement.

The runtime router features include information derived from the current state, state change, input length, recurrence step and translation direction.

Stable Memory Bridge

The decoder does not rely only on the final recurrent endpoint. It receives three information streams:

H_refined  — current refined contextual state
H_base     — pre-recurrence contextual state
E_lex      — lexical shortcut

This is intended to reduce information loss across repeated refinement and to keep a stable decoder interface across different recurrent depths.

Decoder

AGMT uses an autoregressive decoder. The provided runtime currently performs greedy decoding.

---

Repository files

Keep these files in the same directory:

.
├── base.onnx
├── refine.onnx
├── router.onnx
├── decode.onnx
└── translate.py

base.onnx

Runs the base/source path and returns the lexical representation, base contextual representation and source padding mask.

refine.onnx

Runs one shared recurrent refinement step. The same ONNX graph is called repeatedly with the current hidden state and a refinement step ID.

router.onnx

Predicts the value used by the adaptive HALT/REFINE policy.

decode.onnx

Runs the target decoder from target token IDs plus H_refined, H_base, E_lex and the source mask.

translate.py

translate.py is the complete Python inference runtime. It:

• embeds and reconstructs the SentencePiece tokenizer;
• loads all four ONNX graphs from its own directory;
• normalizes and tokenizes input;
• inserts the translation-direction token;
• rejects over-length input instead of silently truncating it;
• runs base encoding;
• builds router features;
• performs adaptive or forced recurrent refinement;
• performs greedy autoregressive decoding;
• exposes normal, forced-depth and full-depth trace inference;
• provides an interactive Vietnamese CLI.

The runtime explicitly uses ONNX Runtime's CPUExecutionProvider, with ONNX graph optimization enabled.

---

Requirements

Only three third-party Python packages are required:

pip install onnxruntime numpy sentencepiece

They are used as follows:

---

Package Purpose

---

onnxruntime Execute base.onnx, refine.onnx, router.onnx, and decode.onnx

numpy Tensor preparation, router feature construction and runtime numerical operations

sentencepiece Load the embedded tokenizer and encode/decode text

The remaining imports used by translate.py come from the Python standard library.

---

Quick start

Put all four .onnx files beside translate.py, install the three dependencies, then run:

python translate.py

The interactive CLI starts with:

AGMT v3.2 — ONNX Translator
Gõ /help để xem toàn bộ lệnh tiếng Việt.

Choose a target language:

/vi

then type English text to translate into Vietnamese, or:

/en

then type Vietnamese text to translate into English.

You can also translate immediately:

/vi Hello world
/en Xin chào

Command-line one-shot usage:

python translate.py /vi "Hello world" balanced
python translate.py /en "Xin chào" balanced
python translate.py /vi "The man was sleeping." quality

---

Interactive commands

---

Command Meaning

---

/en Set target language to English

/en <text> Translate the supplied text directly to English

/vi Set target language to Vietnamese

/vi <text> Translate the supplied text directly to Vietnamese

/mode fast Use a higher effective compute price, making earlier halting easier

/mode balanced Default adaptive routing policy

/mode quality Force all trained recurrent refinement steps

/depth auto Let the active mode/router choose the depth

/depth 0 Decode directly from the base representation

/depth 1..R_MAX Force exactly that many refinement steps

/trace on Decode and print every depth from r=0 through R_MAX, including router values

/trace off Return to normal inference

/max <tokens> Change the maximum number of generated output tokens

/price <value> Change the balanced-mode compute price

/fastmult <value> Change the multiplier used by fast mode

/status Show the current inference configuration

/help Show the built-in Vietnamese help

/q Exit (/quit and /exit also work)

What does [r=...] mean?

Example:

[r=3 | balanced]

means the model actually executed three recurrent refinement steps before decoding.

It does not mean "confidence 3", "reasoning level 3", or "three extra parameter layers".

---

Inference modes

Fast

/mode fast

Uses a higher effective compute price, so the router is more willing to halt early.

Balanced

/mode balanced

Uses the default adaptive HALT/REFINE policy.

Quality

/mode quality

Forces the maximum trained recurrent depth. In the current runtime, quality ignores router halting.

Important: maximum compute does not mathematically guarantee the best output for every sentence. AGMT can have a non-monotonic depth-quality curve, so quality should be understood as force maximum trained compute, not as a guarantee that r=R_MAX is always the optimal translation.

---

Trace mode

Trace mode is useful for studying what recurrence is doing:

/trace on

For each input, the runtime decodes every recurrent depth and reports the router value/decision.

Conceptually:

r=0  → translation before recurrent refinement
r=1  → after one refinement
r=2  → after two refinements
...

This makes it possible to observe transitions such as:

wrong → wrong → correct → correct

or:

correct → correct → degraded

Trace mode is diagnostic only; it does not modify the model weights.

---

Manual qualitative tests

The following are small hand-written diagnostic examples, not a formal benchmark. They are included because they show the behavior that motivated AGMT's elastic recurrent design.

1. Simple translation

EN: The dog chased the cat.
VI: Con chó đuổi con mèo.

A subject/object reversal test also preserved the changed roles:

EN: The cat chased the dog.
VI: Con mèo đã đuổi theo con chó.

2. Recurrent refinement correcting a structural translation

Input:

The man the boy saw was sleeping.

Observed trace:

r=0  incorrect
r=1  incorrect
r=2  "Người đàn ông mà thằng bé thấy đang ngủ."   ✓
r=3  same / router would halt
r=4  still correct
r=5  degraded: information about "the boy" was lost
r=6  degraded

This is an important qualitative example: additional recurrent compute first repaired the relation, but excessive refinement later degraded it. It illustrates why AGMT does not assume that the deepest recurrence is always the best recurrence.

3. Recursive semantic binding

Input:

I know that you know that I know that he knows her.

Observed behavior:

r=0  incorrect binding
r=1  incorrect binding
r=2  corrected:
     "Tôi biết anh biết là tôi biết anh ta biết cô ấy."
r=3..r=6  remained stable in this test

This provides a small qualitative example where repeated application of the shared refiner corrected a semantic-role/binding error.

4. Cases that remain difficult

AGMT v3.2 is experimental and still has clear weaknesses. Manual testing found failures on phenomena such as:

idioms
counterfactual constructions
long-distance relative-clause attachment
some English tense realization
technical/code-switched text
some VI → EN grammatical constructions

For example, an idiom such as:

It's not really my cup of tea, but I can see why people like it.

was translated too literally in manual testing, and forcing deeper recurrence did not solve the underlying knowledge/translation limitation.

These examples are intentionally reported rather than hidden: adaptive routing cannot create translation capability that the translator itself has not learned.

---

How AGMT is intended to be trained

The architectural training concept separates translator learning from final router learning:

Static Data Sanitation
        ↓
Train Translator
        ↓
Learn useful multi-depth recurrence
        ↓
Freeze Translator
        ↓
Profile translations across valid depths
        ↓
Build noise-aware / ε-tolerant recurrence oracle
        ↓
Train source-only translation-value router
        ↓
Calibrate HALT/REFINE compute policy
        ↓
Evaluate translation + compute + semantic challenge sets

The router is intended to predict remaining useful translation value, rather than simply learning that a high-loss or long sentence must receive more compute.

At inference time the router uses source-derived state only; reference translations are not available to it.

---

Design principles

AGMT follows several core principles:

Translation quality is the final authority.

More recurrence is not automatically better.

Sentence length is not a depth rule.

A hard or corrupted sample should not automatically mean "refine more".

The router should allocate compute according to expected translation benefit.

Lexical and base shortcuts should protect useful information from recurrent degradation.

The deepest recurrence is a fallback/diagnostic path, not an assumption of optimality.

In one sentence:

AGMT v3.2 is a shared recurrent translator with translation-value routing, lexical/base shortcuts and a stable memory bridge, designed to spend additional latent computation only when that computation is expected to improve translation.

---

Current status

AGMT v3.2 should currently be treated as an experimental research model.

The examples above are qualitative tests, not standardized BLEU/chrF/COMET results. Formal evaluation should compare fixed recurrent depths, adaptive routing, compute usage, latency, and semantic challenge sets before making broader quality claims.

---

Contact

Trần Tuấn Phi --- Vietnam

Email: [email protected]
I do not check email frequently.

Facebook: https://www.facebook.com/share/1HerWkmghN