Asynchronous Flow Map Language Models Generate All Tokens in a Single Forward Pass
kalyan_kpl · x · 2026-11-02
A new paper introduces asynchronous Flow Map Language Models (aFMLMs), a way to bring token prediction ordering into one-step parallel language modeling.
- Problem: Autoregressive and masked diffusion models iteratively fill in tokens and exploit orderings across generation steps, but leveraging an ordering for one-step parallel generation was unclear.
- Method: aFMLMs learn flow maps of asynchronous flows, where a tokenwise time schedule controls the relative denoising progress of each token, specifying a prediction ordering.
- Result: The model internalizes the ordering into its structure and can generate all tokens in a single forward pass.
More from Research
- Berkeley's Evolving Agent Teams beats production FlashInfer by 1.2x on CUDA kernels from scratch — berkeley_ai · 2026-11-02
- Swarm-based auto-research engine runs 500+ experiments on one agent thread — vincent_koc · 2026-11-02
- ICML paper: finite context windows impose impossibility results for adversarial regret — burny_tech · 2026-11-02
- Labs now train on mini-swe-agent to game evals, says Xeophon — xeophon · 2026-11-02
- WASABI: Blei lab trains diffusion models to sample Boltzmann distributions with 3.3x fewer energy evaluations — skoularidou · 2026-11-02
- A complete math roadmap for machine learning: linear algebra, calculus, probability — TivadarDanka · 2026-11-02