DoubleCat/SemiCat drops self-conditioning, lifts CDLM one-hot flow from 4% to 50% on TinyGSM
tarantulae · x · 2026-10-19
New research on continuous diffusion language models (CDLMs) systematically explores their design space and argues that self-conditioning has been more of a crutch than a principled technique.
The proposed DoubleCat/SemiCat approach removes the need for self-conditioning entirely, boosting the one-hot flow baseline from 4% to 50% on TinyGSM — a major recipe improvement for a paradigm that went from niche to contender this year.
More from Models
- Same agent build task: DeepSeek costs $0.21 vs Opus 5.5's $2.20 — teortaxesTex · 2026-10-19
- Gemini 4 Argon's real wins: 6M lines stripped from Chromium, C++ to Rust at 2.7x speed — ITower__Education · 2026-10-19
- Qwen3.5-0.8B 4-bit spotted running on Rabbit R1 via wllama — ngxson · 2026-10-19
- Microsoft MAI reveals in-house 3T-parameter base models with native encoder-free VLM pretraining — burny_tech · 2026-10-19
- Two Years On, the General LLM Bet Has Paid Off With No Wall in Sight — adivinemessenger · 2026-10-19
- Tencent says rogue agent web-browsing came from third-party model evals, not Hunyuan — CFGeek · 2026-10-19