Venue: NeurIPS 2025 · Author: ByteDance Seed · arXiv: 2506.09990 Category: VLA Architecture / CoT

flowchart RL
  Goal[Goal keyframe] --> KT[Token at t+H]
  KT --> KTm1[Token at t+H-1]
  KTm1 --> Dots[...]
  Dots --> Kt1[Token at t+1]
  Kt1 --> Kt[Token at t]
  Kt --> Exec[Execute]
  note[Generation is BACKWARD:<br/>goal → current, not current → goal]

Standard AR VLAs generate actions forward in time (t → t+H). But manipulation tasks are often defined by their end state (goal keyframe). Forward AR has the model reason "what do I do first?" — a harder question than "what's the last step before the goal?"

Generate the trajectory backward — from a predicted goal keyframe to the current state — in a single autoregressive model. This is action-level CoT: each step is "what action would produce the next keyframe on the path from goal to current?"

At inference, the model:

1. Predicts the goal keyframe — the first token is a stable keyframe action encoding the task goal (predicted by the model, not externally provided).
2. Autoregressively generates trajectory tokens backward from the goal, enforcing a global-to-local structure where each local action is constrained by the final goal.
3. Executes the earliest (current-time) action first.

Actions are emitted as continuous action tokens (regression via linear projection, not discrete tokenization or a separate diffusion/flow head) within a single unified autoregressive transformer. The backbone is lightweight: a ResNet18 (ImageNet-pretrained) vision encoder over multi-view 128×128 RGB, feeding a ~512-d, 8-head encoder-decoder transformer (4 encoder / 7 decoder layers, including a multi-token-prediction layer). The keyframe is defined as a step where the gripper state changes or joint velocities approach zero.

• SOTA on 60 RLBench tasks + 8 real tasks (described as the most comprehensive RLBench evaluation to date).
• Outperforms ACT by 16% and Diffusion Policy by 23% averaged across the 60 RLBench tasks; surpasses ACT by 15% in real-world manipulation.
• On the 60-task suite the absolute averages are roughly CoA 55% vs ACT 39% vs Diffusion Policy 33%; real-world 8-task kitchen suite ~61% (CoA) vs ~46% (ACT).
• Ablations show backward (reverse) AR is the core driver (≈0.76 vs 0.67 forward), and each of the four supporting designs — continuous action tokens, dynamic stopping (variable-length trajectories), reverse temporal ensemble, and multi-token prediction — contributes positively.
• Decisive ablation: regularizing the latent action space via a latent consistency loss is critical (≈0.76 vs 0.21 without it) — far more important than direct action-space supervision.

A conceptually simple change (flip the AR direction) with large empirical impact. Feeds into the ICLR 2026 action-CoT cluster alongside ThinkAct and Robot-R1, informing Embodied-R1's pointing primitives and Actions as Language's framing.

• arXiv: https://arxiv.org/abs/2506.09990
• NeurIPS virtual page: https://neurips.cc/virtual/2025/loc/san-diego/poster/116620
• GitHub: https://github.com/ByteDance-Seed/Chain-of-Action

• ECoT-Lite (CoRL 2025 ancestor)
• ThinkAct (RL-for-reasoning sibling)
• Embodied-R1 (ICLR 2026 descendant)
• Review: VLA Architectures — §5.G

← Back to NeurIPS-2025