New survey maps the fast-rising field of continuous diffusion and flow language models
_akhaliq · x · 2026-11-10
A new survey on Continuous Diffusion and Flow Language Models is out, and Niels Rogge added a "Continuous Language Model" method page on Papers with Code to index all related papers.
These models generate text by transforming noise into continuous token representations (embeddings, simplex/one-hot vectors, bits, or autoencoder latents) rather than predicting tokens one at a time, trained with Gaussian diffusion or flow matching; deterministic trajectories enable few-step distillation and gradient-based guidance. The page already lists 151 papers — 43 in the last 90 days — spanning pre-training, few-step generation, guidance, and scaling. Author @sophiatang notes the field flipped from hard-to-train in early 2025 to exploding in 2026.
Related event: New Survey on Continuous Diffusion and Flow Language Models Released(5 posts)→
More from Research
- Scholars: The Pygmalion Myth Is Now a Central AI Question — begusgasper · 2026-11-10
- Reverse-engineering Google's image labels: 100 tags per photo, product attributes reuse Shopping filters — gaganghotra_ · 2026-11-10
- GPT-6 Astra edits Qwen3.5-0.8B weights directly: 0.016% of params, MATH-500 from 7.9% to 48.1% — sebkrier · 2026-11-10
- Fable 5.2 benchmarks leak: 49.1% FrontierCode, robotics jumps to 64.9%, 89.3% computer use — ChrisGPT · 2026-11-10
- ICLR reviewer spots agent paper whose drawer-task trajectory shows potatoes — yet TASK_SUCCESS — ziv_ravid · 2026-11-10
- Reviewer recomputes all 7,000 benchmark answers — then finds the $38,320 sausage — ziv_ravid · 2026-11-10