Update README.md
Browse files
README.md
CHANGED
|
@@ -7,4 +7,44 @@ sdk: static
|
|
| 7 |
pinned: false
|
| 8 |
---
|
| 9 |
|
| 10 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 7 |
pinned: false
|
| 8 |
---
|
| 9 |
|
| 10 |
+
# BEMSH-BVV Research
|
| 11 |
+
|
| 12 |
+
Independent research in representations, memory, and modular computation.
|
| 13 |
+
|
| 14 |
+
BEMSH-BVV Research is an independent scientific initiative exploring how learning systems represent information, organize memory, and coordinate computation and action. Our current work focuses on language-model architectures, with a longer-term interest in robotics and embodied interaction.
|
| 15 |
+
|
| 16 |
+
## Research directions
|
| 17 |
+
|
| 18 |
+
- **Representations and input interfaces:** distinguishing token identity from learned representations and studying fixed alternatives to trainable input tables.
|
| 19 |
+
- **Modular architectures:** investigating interfaces between contextual computation, persistent memory, and executable skills.
|
| 20 |
+
- **Experimental methodology:** evaluating capabilities, testing interfaces, and separating observed behavior from proposed explanations.
|
| 21 |
+
|
| 22 |
+
## Current projects
|
| 23 |
+
|
| 24 |
+
### Fixed token representations
|
| 25 |
+
|
| 26 |
+
Experiments on language modeling without independently trainable token-specific input vectors.
|
| 27 |
+
|
| 28 |
+
[Paper](https://huggingface.co/papers/2610.04002) 路 [Models and artifacts](https://huggingface.co/collections/Bochkov/do-language-models-need-a-trainable-input-embedding-table)
|
| 29 |
+
|
| 30 |
+
### Modular language-model systems
|
| 31 |
+
|
| 32 |
+
Experiments with reconstructive memories, executable skills, and explicit interfaces for coordinating independently constructed components.
|
| 33 |
+
|
| 34 |
+
[Paper](https://huggingface.co/papers/2610.04012) 路 [Models and artifacts](https://huggingface.co/collections/Bochkov/beyond-the-parameter-monolith-modular-language-modeling)
|
| 35 |
+
|
| 36 |
+
## Open research
|
| 37 |
+
|
| 38 |
+
We publish models, code, experimental methods, and findings鈥攊ncluding limitations and negative results鈥攖o support independent evaluation, further research, and teaching.
|
| 39 |
+
|
| 40 |
+
Each release documents its own scope, requirements, and limitations. Research prototypes should not be treated as validated deployment systems.
|
| 41 |
+
|
| 42 |
+
Existing project artifacts are currently hosted under [Bochkov](https://huggingface.co/Bochkov); this organization provides a shared home for the developing research initiative.
|
| 43 |
+
|
| 44 |
+
## Collaboration
|
| 45 |
+
|
| 46 |
+
The initiative is currently maintained by [Andrey Bochkov](https://huggingface.co/Bochkov).
|
| 47 |
+
|
| 48 |
+
Contributions are welcome in independent replication, evaluation, architecture design, representation analysis, and educational materials. To discuss a specific project, please open a discussion in the relevant model repository.
|
| 49 |
+
|
| 50 |
+
When using an artifact, please consult its model card for licensing, usage conditions, and citation information.
|