SAE on Coconut latent thoughts (Qwen3-4B)
SAE bottleneck trained on the frozen Stage-3 Coconut latent thoughts of Qwen3-4B,
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:
sae.pt
- TopK SAE, dictionary 8x hidden, k=32 active features
- Result: inserting this bottleneck into the frozen latent loop costs ~0 GSM8K test accuracy.
