Paper 2603.17051
Astrolabe: Steering Forward-Process Reinforcement Learning for Distilled Autoregressive Video Models
- Published
- Mar 2026
- Research lab
- Independent
- Citations
- 10
- GitHub
- 151 stars
01 In brief
Summary
Astrolabe is an online reinforcement learning (RL) framework designed to align distilled autoregressive (AR) video models with human preferences without re-distillation.
It addresses limitations of existing RL methods, which require expensive reverse-process optimization or lack active exploration.
Astrolabe uses a forward-process RL formulation based on negative-aware fine-tuning, contrasting positive and negative samples at inference endpoints to establish an implicit policy improvement direction.
For long videos, it employs a streaming training scheme with a rolling KV-cache, applying RL updates to local clip windows while conditioning on prior context.
To mitigate reward hacking, it integrates a multi-reward objective (visual quality, motion quality, text alignment) stabilized by uncertainty-aware selective regularization and dynamic reference updates.
Experiments on models like Self-Forcing, LongLive, and Causal-Forcing show consistent improvements in generation quality across short, long, and multi-prompt settings, with maintained inference speed.
Ablations validate the streaming training scheme, multi-reward formulation, and selective KL penalty, which reduce memory usage and prevent reward hacking.
02 From the paper
Abstract
Distilled autoregressive (AR) video models enable efficient streaming generation but frequently misalign with human visual preferences. Existing reinforcement learning (RL) frameworks are not naturally suited to these architectures, typically requiring either expensive re-distillation or solver-coupled reverse-process optimization that introduces considerable memory and computational overhead. We present Astrolabe, an efficient online RL framework tailored for distilled AR models. To overcome existing bottlenecks, we introduce a forward-process RL formulation based on negative-aware fine-tuning. By contrasting positive and negative samples directly at inference endpoints, this approach establishes an implicit policy improvement direction without requiring reverse-process unrolling. To scale this alignment to long videos, we propose a streaming training scheme that generates sequences progressively via a rolling KV-cache, applying RL updates exclusively to local clip windows while conditioning on prior context to ensure long-range coherence. Finally, to mitigate reward hacking, we integrate a multi-reward objective stabilized by uncertainty-aware selective regularization and dynamic reference updates. Extensive experiments demonstrate that our method consistently enhances generation quality across multiple distilled AR video models, serving as a robust and scalable alignment solution.