Released g05-libero / g05-robotwin20 disable AR+CoT (ar.ce_weight=0); how does this reconcile with the paper's AR framing?

[Zheng-Liming]

Hello, thanks a lot for open-sourcing G0.5, the checkpoints, and the eval pipeline.

While preparing to evaluate the released checkpoints, I noticed that the sim-benchmark checkpoints and the real-robot checkpoints are trained in fundamentally different modes (discrete_action=true/false observed from .hydra/config.yaml in the checkpoints), which seems inconsistent with how the paper and README describe G0.5, and I want to make sure I'm reading the setup correctly. I would like to ask about the rationale. The two sim benchmarks are evaluated with a pure FM model, while the paper and checkpoints/README.md position G0.5 as an autoregressive VLA that "generates reasoning tokens and action tokens in one unified autoregressive stream". The released sim checkpoints do not exercise that AR / CoT pathway at all.

I would like to understand

1. Why was the AR + CoT branch disabled specifically for the LIBERO and RoboTwin 2.0 checkpoints? Was it because AR + CoT supervision hurt performance in these sim benchmarks in your experiments, because the datasets lack usable reasoning / language annotations, because of training stability or compute considerations, or some other reason?
2. Are the LIBERO and RoboTwin 2.0 numbers in the technical report produced by exactly these released FM-only checkpoints, or by an internal AR + FM variant that was not released? If the reported numbers come from an AR variant, could you release the corresponding checkpoint or training config so the paper's autoregressive claims can be reproduced on these benchmarks?
3. If a user wants to fine-tune an AR + CoT capable model on LIBERO or RoboTwin 2.0 starting from g05-base, is there a recommended configuration (e.g. ar.ce_weight, predict_cot, source of CoT annotations)? A minimal reference config or a short note in the README would be very helpful.
4. For downstream users fine-tuning on new tasks / embodiments, do you recommend AR+FM+CoT as the default (matching droid/so101), or FM-only when CoT annotations are unavailable (matching libero/robotwin20)? Is there a rule of thumb from your internal ablations?

Thanks in advance for the clarification.