datasysdev
/

clsd

@@ -24,6 +24,10 @@ base_model:
 ---
 ## Thesis
 Autoregressive language models are bottlenecked by sequential generation. CLSD deploys a hybrid causal Diffusion Transformer (DiT) -- a strided 12-layer slice of Qwen3.5-9B -- operating in the continuous embedding space of the same frozen Qwen3.5-9B verifier. Both models share the exact same 4096-dimensional manifold, the same tokenizer, and the same attention geometry. No projection bridges, no dimensional translation loss.

 ---
+## What is cool
+The base verifier has fixed weights, but its inference process is not exhausted by ordinary left-to-right decoding. A learned continuous proposer can search for hidden-state trajectories and token paths that the verifier can recognize as correct, even if the verifier would rarely or never reach them under standard autoregressive rollout.
 ## Thesis
 Autoregressive language models are bottlenecked by sequential generation. CLSD deploys a hybrid causal Diffusion Transformer (DiT) -- a strided 12-layer slice of Qwen3.5-9B -- operating in the continuous embedding space of the same frozen Qwen3.5-9B verifier. Both models share the exact same 4096-dimensional manifold, the same tokenizer, and the same attention geometry. No projection bridges, no dimensional translation loss.