Growing Transformers: Modular Composition and Layer-wise Expansion on a Frozen Substrate
Paper
•
2507.07129
•
Published
•
2
Andrey
Bochkov
AI & ML interests
None yet
Recent Activity
reacted
to
sergiopaniego's
post
with 🔥
2 days ago
New REPL environment in OpenEnv available! ✨
Used in the Recursive Language Models (RLM) paper by Alex Zhang.
Ready for inference & post-training using trajectories. Handles long contexts:
> Run Python code in a sandbox
> Make recursive calls to LMs
> Explore data programmatically
> Return final result
Docs: https://meta-pytorch.org/OpenEnv/environments/repl/
Inference script: https://github.com/meta-pytorch/OpenEnv/blob/main/examples/repl_oolong_simple.py
posted
an
update
2 days ago
Curious reproducible fact: I trained a GPT-like decoder-only Transformer where the entire input embedding table is frozen and reduced to a 16‑D binary token-ID code (0/1) — this is NOT 16-bit quantization.
Key details:
- vocab_size = 65536, n_embed = 16 (2^16 = 65536 unique IDs)
- deterministic expansion 16 → d_model=1024 via repeat_interleave (scale=64)
- full embedding table is published (embeddings.txt) for auditability
Repro note + verification script:
https://huggingface.co/blog/Bochkov/emergent-semantics-beyond-token-embeddings
Model repo:
https://huggingface.co/Bochkov/emergent-semantics-model-16-bit-269m
License: Apache-2.0
Organizations
None yet