VQ-VAE on Coconut latent thoughts (Qwen3-0.6B)
VQ-VAE bottleneck trained on the frozen Stage-3 Coconut latent thoughts of Qwen3-0.6B,
from the VQ-CoT: Discretising Latent Chain-of-Thought project (team RateLimit Achieved, EPFL CS-552).
The language model is frozen; only this bottleneck is trained on the dumped latents
(385,620 x 6 latent thought vectors, GSM8K).
- Checkpoint:
vq.pt
- K=4096 codes, code dim 64, EMA 0.999, k-means init + AE warmup
- Result: inserting this bottleneck into the frozen latent loop costs ~0 GSM8K test accuracy.
