acceptodds
Under review as a conference paper at ICLR 2027

Continuous Reasoning for Vision-Language-Action Models

Abstract

A robot's intermediate thoughts can be trained through the actions they enable, and a useful thought should serve more than the network that produced it. We introduce Continuous Reasoning (CR), a vision-language-action (VLA) framework that learns a sequence of continuous thought vectors directly from action supervision, without annotated reasoning traces. Our key idea is self-verification: a thought should support action prediction by both the policy that generates it and an evolving exponential-moving-average (EMA) consumer. The consumer's action-prediction error backpropagates to the code generator, training thoughts through their use for action prediction. Built on π0.5, CR generates thoughts sequentially through a Gaussian-prior bottleneck and uses the resulting prefix to condition block-causal action prediction. The EMA consumer is used only during training, so inference requires no separate verification pass. On LIBERO-PRO, CR improves mean success from 58.0% to 64.0% over its backbone and exceeds DeepThinkVLA-RL in all four suite means, with a 14.0-percentage-point overall advantage. Gains concentrate on position and task perturbations, where policies must adapt to changed layouts and goals. On TX-G2 and HSR, mean subtask success improves by 19.8 and 15.6 percentage points, respectively, with gains also extending to full-task completion. Every component ablation loses on those same two families, and freezing the consumer removes the benefit of self-verification entirely. These results establish action-supervised continuous reasoning as an effective interface for robust VLA control.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Robot Learning, Reasoning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_cbd7OpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Continuous Reasoning, which aims to learn continuous intermediate representations of VLA models through action supervision alone, without requiring explicit reasoning annotations. The core method involves having a student model generate latent thoughts and requiring an EMA consumer to also utilize these representations for the same action prediction, thereby reducing generator-decoder co-adaptation. The approach further incorporates a Gaussian latent bottleneck with block-causal action decoding. Experiments demonstrate that CR achieves consistent improvements on LIBERO-PRO under position/task perturbations, ablation studies, and two real-world robotic tasks.

Critical strengths
  1. The problem is clearly defined, and the method design is concise. The paper directly uses action supervision to learn intermediate representations, avoiding reliance on additional reasoning annotations; the EMA consumer also provides a simple and targeted mechanism to alleviate the co-adaptation issue between the latent and its own decoder.
  2. The experimental results show good consistency. The improvement in CR is primarily observed under position and task perturbations, and the corresponding ablation studies mainly focus on these two settings. Moreover, similar trends are observed in real robot experiments, making the core conclusions more convincing than mere overall performance gains.
Critical weaknesses
  1. It has not yet been sufficiently demonstrated that the EMA consumer itself is essential, rather than a general auxiliary regularization effect. Current comparisons mainly involve removing verification or freezing the EMA consumer, but lack simpler controls such as an additional action head, or standard EMA consistency. Therefore, it remains difficult to determine whether the performance gain stems from the proposed moving-reader mechanism or simply from additional training constraints.
  2. Some experiments fail to cleanly isolate the contributions of individual CR components. For example, removing continuous thoughts also eliminates latent structuring and verification; removing Gaussian latent space simultaneously removes the bottleneck and WAE regularization. As a result, current ablation studies better demonstrate that the full design is effective, but make it harder to confirm whether each core component is individually necessary.
Questions for authors
  1. What is the inference computational overhead of CR compared to the original π0.5?
    Although the EMA consumer is used only during training, CR still requires sequential generation of continuous thoughts at inference time. It is recommended to report latency, FLOPs, or actual control frequency to evaluate the deployment cost associated with performance gains.
  2. Can CR be validated on another VLA backbone?
    The current main results are based on π0.5. If similar improvements can be reproduced on different VLA backbones, it would more strongly support that CR is a general reasoning/representation learning mechanism rather than one dependent on specific backbone design.
Reviewer_WbFROpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Continuous Reasoning (CR), a VLA framework that learns a sequence of continuous thought vectors before action prediction. Unlike approaches that supervise intermediate reasoning with textual traces, visual subgoals, or annotated plans, CR trains the intermediate interface solely through action supervision.Its central mechanism is self-verification: the student-generated latent codes are consumed both by the student policy and by an EMA-updated consumer policy, whose action-prediction loss backpropagates into the code generator. CR additionally uses a Gaussian-prior bottleneck/WAE regularizer and block-causal action decoding. On LIBERO-PRO, CR improves the π0.5 backbone from 58.0% to 64.0% average success, with most gains on position and task perturbations. Ablations support the role of the latent interface, continuous thoughts, block-causal masking, and self-verification. The paper also reports improvements on TX-G2 and HSR real-robot tasks and provides qualitative PCA visualizations suggesting that the learned thought states track instructions and target relocation.

Critical strengths
  • The paper avoids relying on annotated reasoning traces or explicit textual reasoning and instead ties intermediate representations directly to action utility, which is a meaningful direction for VLA reasoning.
  • CR is more than adding latent tokens: it combines sequential thought generation, a Gaussian bottleneck, block-causal action decoding, and EMA-consumer self-verification in a way that aligns with the stated goal of making thoughts usable by a second reader.
  • The LIBERO-PRO gains concentrate on position and task perturbations, while object and semantic perturbations change much less. This pattern supports the paper's claim that CR mainly helps with re-anchoring behavior under changed layouts and goals.
  • Removing the Gaussian latent space, continuous thoughts, self-verification, or the block-causal mask all hurts performance, especially on position/task perturbations. The frozen-consumer control is particularly informative because it supports the need for an evolving EMA reader.
  • Results on TX-G2 and HSR cover both bimanual manipulation and mobile manipulation. Although limited in scale, these experiments make the work more convincing than a purely simulated study.
  • Since the EMA consumer is training-only, the method remains practical at deployment.
Critical weaknesses
  • CR introduces continuous thoughts, a WAE bottleneck, an EMA consumer, block-causal masking, and additional training losses at the same time. The ablations are useful, but more tightly matched controls for parameter count, training compute, and regularization would better isolate how much of the gain comes from the reasoning interface itself versus extra capacity or auxiliary supervision.
  • DeepThinkVLA-RL uses a different backbone and training recipe, and the paper notes that 8.1 points of the 14.0-point gap come from the π0.5 backbone alone. The claim that CR outperforms explicit text reasoning should therefore be stated more cautiously unless supported by same-backbone or closer-budget baselines.
  • TX-G2 uses 10 episodes per task and HSR uses 6 episodes per task. The Wilson intervals are wide, so the real-robot results are encouraging but not yet definitive.
  • The EMA consumer and WAE/verification objectives increase training compute and engineering complexity. The paper emphasizes no additional inference pass, but provides less analysis of training overhead, memory cost, and hyperparameter sensitivity.
  • The main evidence comes from LIBERO-PRO and two in-house real-robot settings. It is still unclear whether CR transfers robustly to more robot embodiments, broader task families, longer horizons, stronger perturbations, or open-vocabulary instructions.
Questions for authors

None

Reviewer_sNbgOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces Continuous Reasoning (CR), which adds continuous "thought" vectors to a π0.5 VLA policy and trains them only through action supervision, using a Gaussian-prior bottleneck and block-causal action decoding. Its key idea is self-verification: an evolving EMA copy of the policy must also act on the thoughts, so they can't become readable only by their own decoder. Experiments on LIBERO-PRO, ablations, and two real robots show gains over the backbone and a text-reasoning baseline.

Critical strengths
  1. The paper is well motivated and removes the need for reasoning annotations. CR learns its intermediate thoughts purely from action supervision, using an evolving EMA consumer to keep the latent from co-adapting with its own decoder. The ablations also support claims in the paper.
  2. The performance is promising in both simulation and the real world. On LIBERO-PRO, CR improves robustness over its backbone and a text-reasoning baseline, especially under position and task perturbations. The improvements are also shown on two different physical platforms (bimanual and mobile).
Critical weaknesses
  1. The claim that the gains come from "reasoning" is not completely supported. With only two thought slots trained purely through the action loss, it is unclear whether the improvement comes from intermediate reasoning or simply from extra capacity and computation added on top of π0.5.
  2. The inference cost of the thought generation is not analyzed. The thoughts are generated sequentially, each requiring its own pass through the transformer plus WAE encoding and decoding before actions can be predicted, which adds latency to every policy query. The paper emphasizes that no verification pass is needed at inference but never reports runtime or control frequency compared with π0.5.
Questions for authors
  1. Can you provide experimental results to further clarify the problem mentioned in weakness 1, for example, the same number of learned query tokens without sequential generation, the WAE bottleneck, or verification? This would show whether the gains come from the structure of the thoughts rather than from extra capacity.
  2. What is the inference latency or achievable control frequency of CR compared with π0.5, on both the simulation setup and the real robots?
  3. How does performance change with the number of thought slots? Do more slots increase performance more, as the reasoning framing implies?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.