| --- |
| library_name: transformers |
| license: apache-2.0 |
| tags: |
| - decision-model |
| - latent-reasoning |
| - recurrent-depth |
| - text-classification |
| - adaptive-compute |
| - multilingual |
| base_model: jhu-clsp/mmBERT-small |
| --- |
| |
| # Dazo |
|
|
| **Dazo** is an experimental recurrent latent decision model. It is designed to spend variable test-time compute on structured decisions without generating a natural-language chain of thought. |
|
|
| Dazo is not a pretrained general-purpose checkpoint yet. The initial release is a research architecture and training harness. |
|
|
| ## Architecture |
|
|
| 1. A bidirectional encoder reads the evidence/context. |
| 2. Candidate actions/labels are encoded separately rather than packed into the context sequence. |
| 3. A Perceiver-style cross-attention compressor maps the evidence into a fixed latent workspace. |
| 4. A weight-tied recurrent Transformer repeatedly refines that workspace. |
| 5. Candidate options query the workspace in parallel to produce a calibrated probability distribution. |
| 6. A separate critic predicts correctness/abstention, while recurrent convergence and a learned halt probability provide stopping signals. |
|
|
| The v0 workspace uses evidence, hypothesis, critic, and control slots. The option decoder is permutation-equivariant for categorical decisions; ordinal options carry explicit semantic rank IDs. |
|
|
| ## Intended research questions |
|
|
| - Does accuracy increase when Dazo receives more recurrent inference steps on problems requiring deeper composition? |
| - Can Dazo generalize from shallow training chains to deeper unseen chains by increasing recurrence at test time? |
| - Can adaptive stopping avoid the known recurrent-depth failure mode of overthinking? |
| - Does separate option encoding eliminate the high-cardinality and option-order bottlenecks seen in prompt-packed decision models? |
|
|
| ## Status |
|
|
| Research prototype. Do not treat probabilities as production-calibrated until the model is trained and calibrated on the deployment domain. |
|
|
