Continuous Reasoning in Vision-Language-Action

Updated 4 July 2026

Continuous reasoning for VLA is an approach that embeds an intermediate computation—using latent thought tokens, textual CoT, or continuous action functions—between multimodal perception and fine-grained action control.
It aligns reasoning with temporal and geometric structures by replacing discrete, high-level language subgoals with continuously updated internal representations that directly influence downstream control.
Architectural designs emphasize shareability and verifiability, utilizing techniques such as flow-matching, Gaussian latent interfaces, and hybrid attention to improve efficiency and robustness in tasks like manipulation, navigation, and driving.

Continuous reasoning for vision-language-action (VLA) denotes a family of methods that insert an explicit or implicit reasoning process between multimodal perception and continuous control, with the goal of aligning intermediate computation to the temporal and geometric structure of embodied action rather than to language alone. In recent work, this idea appears as textual chain-of-thought (CoT), visual foresight, tool-invoked perception, latent “thought” variables, shared Gaussian reasoning interfaces, multi-path latent deliberation, and continuous action functions. The common motivation is that natural language and discrete subgoals often operate at task-level granularity, whereas VLA policies must issue action chunks or motor commands at much finer temporal scales; this creates a representation mismatch, inference overhead, and a weak coupling between “reasoning” and the action needed immediately (Wu et al., 29 May 2026, Bai et al., 1 Feb 2026).

1. Conceptual foundations

A central claim in the recent literature is that reasoning in VLA should not be identified solely with the generation of natural-language explanations. “Continuous Reasoning for Vision-Language-Action” argues that a useful reasoning medium for VLA must be shareable across model instances, verifiable through downstream action improvement, and aligned with temporally extended control structure; the paper treats the question of reasoning as a question about the interface by which action is improved, rather than about whether extra text is produced (Wu et al., 29 May 2026). In a related formulation, LaRA-VLA states that existing CoT-based VLA approaches incur high inference overhead and rely on discrete reasoning representations that mismatch continuous perception and control, motivating continuous latent reasoning states in place of explicit textual CoT at inference (Bai et al., 1 Feb 2026).

This reframing changes the status of intermediate computation. In VLA-Thinker, reasoning is modeled as an iterative multimodal loop in a partially observable setting, where “perception” itself is a reasoning action and the trajectory includes textual thoughts, tool invocations, returned visual evidence, and the final control action (Wang et al., 15 Mar 2026). ThinkingVLA makes a similar decomposition explicit by factorizing action generation into forward reasoning about the immediate subgoal, predicted next observation, inverse reasoning grounded in the predicted image, and only then final action prediction (Lu et al., 16 Jun 2026). These formulations treat reasoning as a temporally extended control primitive rather than a post hoc explanation.

A recurring misconception in this area is that any auxiliary latent or any improved policy accuracy constitutes “reasoning.” The strongest rejection of that view appears in Continuous Reasoning for Vision-Language-Action, which states that better action prediction alone does not certify good reasoning if the same internal medium cannot be shared across model instances and independently verified through improved downstream control (Wu et al., 29 May 2026). This has made verification, reuse, and causal influence on action central design criteria rather than optional interpretability add-ons.

2. Reasoning media and representational choices

Recent VLA systems differ primarily in the medium used for intermediate reasoning.

Reasoning medium	Representative instantiations	Characteristic role
Textual CoT	AutoVLA, VLA-R1, DeepThinkVLA, MobileVLA-R1	Explicit stepwise reasoning before action
Visual foresight or image reasoning	HALO, ThinkingVLA, VLA-Thinker	Predict subgoals, revisit evidence, or ground inverse reasoning
Continuous latent thoughts	LaRA-VLA, Continuous Reasoning, MPCoT	Replace or augment explicit CoT with compact latent interfaces
Continuous action functions	NIAF	Move continuity from waypoint prediction to the action representation itself

Textual CoT remains the most direct approach. AutoVLA uses two thinking modes within a single autoregressive model for autonomous driving: a fast mode that directly emits action tokens and a slow mode that inserts four ordered CoT sub-steps—scene description, critical object identification, surrounding-agent intention prediction, and driving decision and meta-action—before trajectory tokens (Zhou et al., 16 Jun 2025). VLA-R1 formalizes explicit reasoning with a “<think>” head and an “<output>” head, where the output can be either bounding-box parameters or waypoint sequences later mapped to continuous 7D robot commands (Ye et al., 2 Oct 2025). DeepThinkVLA and MobileVLA-R1 likewise preserve explicit textual reasoning during training and inference, but couple it more tightly to downstream action generation (Yin et al., 31 Oct 2025, Huang et al., 22 Nov 2025).

A second family uses visual prediction as part of the reasoning process. HALO inserts textual task reasoning and visual subgoal prediction before action chunk prediction, factorizing the embodied policy into textual reasoning generation, visual subgoal prediction, and action prediction conditioned on both (Shou et al., 24 Feb 2026). ThinkingVLA extends this idea by interleaving forward CoT, future-image generation via discrete image tokens, and inverse CoT grounded in the predicted image, so that the target state becomes part of the action reasoning context (Lu et al., 16 Jun 2026). VLA-Thinker makes the visual loop interactive rather than purely predictive: the model can emit a tool command such as ZOOM-IN, receive a cropped sub-image, and continue reasoning with the returned evidence (Wang et al., 15 Mar 2026).

A third family replaces explicit text with compact latent reasoning variables. LaRA-VLA introduces a small set of continuous “thinking” tokens $z_t$ updated by transformer layers according to

$z_{t+1}=f_\theta(z_t,v_t,x),$

where the latent thinker absorbs reasoning structure that was initially supervised by explicit textual and visual CoT (Bai et al., 1 Feb 2026). Continuous Reasoning for Vision-Language-Action instead uses raw thought vectors $\tau$ that are encoded into a Gaussian latent interface $z\sim q_\phi(z\mid\tau)$ and then decoded back into shared thoughts; the design is explicitly intended to make the reasoning medium shared and verifiable across model instances (Wu et al., 29 May 2026). MPCoT moves latent reasoning to test time, creating $M$ parallel hypotheses refined for $K$ recurrent steps and softly aggregating them before action decoding, while preserving the original 8-step action interface and generating zero reasoning tokens (Zhang et al., 4 Jun 2026).

Finally, some work pushes continuity directly into the action representation. NIAF replaces discrete waypoint prediction with a continuous vector-valued action function

$\Phi(\tau;\theta)=f_\theta(t),$

implemented as a SIREN modulated by an MLLM, so that velocity, acceleration, and jerk can be supervised analytically (Liu et al., 2 Mar 2026). This is not a CoT mechanism in the narrow textual sense, but it extends the continuous-reasoning agenda by treating action generation itself as continuous-time function regression rather than discrete token or waypoint prediction.

3. Architectural patterns for coupling reasoning and action

Architecturally, continuous reasoning methods differ in how they couple intermediate representations to action generation. LaRA-VLA builds on a Qwen3-VL backbone and adds three modules: a Latent Thinker that maintains the latent tokens $z_t$ , a Reasoning Dynamics Module implemented as transformer layers, and an Action Decoder based on a Flow-matching Diffusion Transformer that conditions on the multimodal latent context $h_t=[z_t;v_t;\text{text}]$ to produce continuous control trajectories (Bai et al., 1 Feb 2026). The design removes textual CoT generation at inference and replaces it with a single compact latent update per step.

Continuous Reasoning for Vision-Language-Action organizes the pipeline into thought prediction, latent structuring, and chunk-structured action generation. Raw thought vectors are encoded into a Gaussian latent using a Wasserstein autoencoder, then decoded into shared thoughts that condition a flow-matching policy over action chunks. The action horizon $H$ is partitioned into $z_{t+1}=f_\theta(z_t,v_t,x),$ 0 chunks of size $z_{t+1}=f_\theta(z_t,v_t,x),$ 1, with block-causal attention allowing later chunks to attend to earlier ones while preventing access to future information (Wu et al., 29 May 2026). This chunk-level structure is the architectural mechanism by which the reasoning timescale is aligned to action-chunk granularity.

Several papers address the longstanding mismatch between sequential reasoning and parallelizable control. DeepThinkVLA introduces a hybrid-attention decoder that uses causal attention for CoT tokens and then switches to bidirectional attention for the action block, allowing the action vector to be decoded in parallel once reasoning is complete (Yin et al., 31 Oct 2025). HALO and ThinkingVLA solve a related problem through Mixture-of-Transformers architectures: both separate reasoning- and action-related feed-forward processing into specialized experts while preserving shared attention for cross-modal communication (Shou et al., 24 Feb 2026, Lu et al., 16 Jun 2026).

Other architectures pursue efficiency through alternative sequence models or action heads. RoboMamba integrates a frozen CLIP-ViT vision encoder with a Mamba backbone and exploits the selective state space model’s linear inference complexity, $z_{t+1}=f_\theta(z_t,v_t,x),$ 2 per layer, before attaching a minimal SE(3) pose-prediction policy head that constitutes about 0.1% of the model parameters (Liu et al., 2024). Reasoning-VLA, in autonomous driving, dispenses with autoregressive action tokens and instead refines Gaussian-initialized learnable action queries through self-attention and cross-attention to fused vision-language features, finally regressing continuous trajectory coordinates in parallel (Zhang et al., 25 Nov 2025). CoT4AD adopts a different route again: it weaves 3D perception, VQA-style semantic reasoning, future BEV prediction through latent diffusion, and diffusion-based waypoint planning into a perception-question-prediction-action pipeline whose explicit CoT is compressed into internal activations at inference (Wang et al., 27 Nov 2025).

Taken together, these systems indicate that “continuous reasoning” is not a single architecture class. It includes latent recurrence, expert-routed autoregression, state space backbones, diffusion-based planners, continuous function regressors, and query-based parallel decoders, provided that the intermediate computation is explicitly linked to downstream control.

4. Training, supervision, and verification

Most continuous-reasoning VLA systems rely on staged training. LaRA-VLA provides one of the clearest curricula. Stage I performs explicit CoT fine-tuning with teacher-forced textual CoT and future-image latent prediction; Stage II gradually replaces subsets of discrete CoT tokens with latent tokens over 2–6k steps using a mixing coefficient $z_{t+1}=f_\theta(z_t,v_t,x),$ 3 annealed from $z_{t+1}=f_\theta(z_t,v_t,x),$ 4 to $z_{t+1}=f_\theta(z_t,v_t,x),$ 5; Stage III removes explicit CoT and discrete action tokens, training only the continuous flow-matching action objective (Bai et al., 1 Feb 2026). This design is intended to transfer structured reasoning from explicit multimodal supervision into a latent form without discarding the action-conditioning signal.

Continuous Reasoning for Vision-Language-Action makes verification itself part of the objective. Its total loss combines a flow-matching action loss, a Wasserstein autoencoder loss for the Gaussian latent interface, and a self-verification loss in which an exponential-moving-average teacher must consume the same latent $z_{t+1}=f_\theta(z_t,v_t,x),$ 6 and also predict the target vector field successfully:

$z_{t+1}=f_\theta(z_t,v_t,x),$ 7

The point of this construction is not merely regularization. The paper argues that because another network must re-consume the same reasoning medium to recover good action prediction, the latent becomes both shareable and verifiable rather than a model-private shortcut (Wu et al., 29 May 2026).

Reinforcement learning is the other dominant training pattern, especially when reasoning length, action consistency, or long-horizon recovery must be optimized beyond imitation. AutoVLA applies Group Relative Policy Optimization (GRPO) after supervised fine-tuning and adds a CoT length penalty to the reward so that the model reduces unnecessary reasoning in easy driving scenarios and uses fast thinking when it suffices (Zhou et al., 16 Jun 2025). VLA-R1 combines Reinforcement Learning from Verifiable Rewards (RLVR) with GRPO, using region alignment, trajectory consistency, and output-format rewards to strengthen both reasoning robustness and execution accuracy (Ye et al., 2 Oct 2025). VLA-Thinker optimizes complete reasoning-perception-action trajectories with sparse task-success and format rewards, while MobileVLA-R1 uses GRPO with movement, action, and format rewards after supervised CoT alignment to stabilize continuous control for quadruped robots (Wang et al., 15 Mar 2026, Huang et al., 22 Nov 2025).

Dataset construction is correspondingly specialized. VLA-R1 introduces VLA-CoT-13K, aligned with affordance and trajectory annotations (Ye et al., 2 Oct 2025). HALO describes an automated embodied multimodal CoT synthesis pipeline that converts robotic trajectories into tuples of instruction, observation history, reasoning text, visual subgoal, and action (Shou et al., 24 Feb 2026). AutoVLA distills CoT from a larger Qwen2.5-VL-72B engine, yielding 45.6k annotations in nuPlan and 7.2k in Waymo (Zhou et al., 16 Jun 2025). MobileVLA-R1 constructs MobileVLA-CoT with episode-, navigation-, and step-level reasoning annotations from R2R, RxR, and QUARD (Huang et al., 22 Nov 2025). The field therefore depends not only on architectural choices, but also on whether supervision is structured enough to align intermediate reasoning with action semantics and temporal context.

Empirical results show that continuous reasoning is being evaluated in three major regimes: robotic manipulation, autonomous driving, and navigation/mobile robotics.

Domain	Method	Reported outcome
Manipulation	LaRA-VLA	97.9% average success on LIBERO, 96.6% on Long, 68.8% on SimplerEnv, and 135 ms per step on A100 versus about 1 s for explicit CoT methods (Bai et al., 1 Feb 2026)
Manipulation robustness	Continuous Reasoning	LIBERO-PRO average suite mean improves from 58.0% to 64.0%; TX-G2 reaches 68.8% mean subtask success, 40.4% over $z_{t+1}=f_\theta(z_t,v_t,x),$ 8; HSR reaches 75.0%, 26.3% over $z_{t+1}=f_\theta(z_t,v_t,x),$ 9 (Wu et al., 29 May 2026)
Manipulation with interactive perception	VLA-Thinker	97.5% success on LIBERO; RoboTwin 2.0 improves from 21.3% to 62.3% on short horizon, 47.1% to 70.7% on medium horizon, and 46.5% to 64.6% on long/extra-long horizon (Wang et al., 15 Mar 2026)
Test-time latent deliberation	MPCoT	LIBERO average success rises from 96.8% to 98.9% and Long from 95.3% to 98.9%; CALVIN 5-step success rises from 72.9% to 89.4% (Zhang et al., 4 Jun 2026)
Continuous action fields	NIAF	97.9% average success on LIBERO and AvgLen 4.66 on CALVIN ABCD→D; real-world item placement reaches 90% and cup stacking 80% (Liu et al., 2 Mar 2026)
Autonomous driving	AutoVLA / CoT4AD / Reasoning-VLA	AutoVLA reports +10.6% PDMS and 66.8% runtime reduction on nuPlan (Zhou et al., 16 Jun 2025); CoT4AD reaches DS 80.24 and SR 55.22%, with explicit CoT at DS 81.22 and SR 55.78% (Wang et al., 27 Nov 2025); Reasoning-VLA reports average L2 0.23 m, collision 0.08%, and 0.089 s to predict 10 trajectories on H200 (Zhang et al., 25 Nov 2025)

In robotic manipulation, the main empirical pattern is that reasoning benefits increase with horizon length, perturbation, or subtask coupling. LaRA-VLA reports especially strong Long-horizon results on LIBERO and attributes the speedup to eliminating autoregressive textual CoT at inference (Bai et al., 1 Feb 2026). Continuous Reasoning emphasizes robustness under object, position, semantic, and task perturbations on LIBERO-PRO, and its ablation study reports that removing Gaussian latents, chunk masking, thought slots, or self-verification degrades position and task success by 6–9 points (Wu et al., 29 May 2026). VLA-Thinker attributes gains to active zoom-in perception, which allows the model to resolve fine-grained ambiguities and correct drift during long-horizon tasks (Wang et al., 15 Mar 2026).

For autonomous driving, the principal empirical issue is the trade-off between reasoning depth and inference speed. AutoVLA keeps reasoning and action within a single autoregressive generation model and then uses GRPO to reduce unnecessary CoT in straightforward scenarios, reporting both open-loop and closed-loop gains together with substantial runtime reduction (Zhou et al., 16 Jun 2025). CoT4AD reports that removing the perception, VQA, or future-prediction stages of its CoT pre-training pipeline sharply reduces Bench2Drive Driving Score and Success Rate, indicating that its explicit multi-stage reasoning supervision materially affects downstream planning quality (Wang et al., 27 Nov 2025). Reasoning-VLA pushes further toward parallel continuous action prediction, reporting strong generalization across eight datasets and a marked throughput advantage over autoregressive decoding (Zhang et al., 25 Nov 2025).

Navigation and mobile robotics show that continuous reasoning is not limited to arm manipulation or driving. LaViRA decomposes action hierarchically into Language Action, Vision Action, and Robot Action for zero-shot VLN-CE, and reports SR 38.3% and SPL 28.3% with Gemini-2.5-Pro on 100 unseen validation episodes (Ding et al., 22 Oct 2025). MobileVLA-R1 reports about a 5% improvement on VLN tasks, improves QUARD average success from 60% to 73%, and shows higher real-world success under complex quadruped instructions than GPT-4o and NaVILA baselines (Huang et al., 22 Nov 2025). These results suggest that the broader topic is not tied to one embodiment, but to the shared problem of aligning high-level multimodal reasoning with continuous low-level execution.

6. Limitations, debates, and open directions

The main unresolved issue is interpretability. Continuous Reasoning for Vision-Language-Action explicitly notes that continuous thoughts are not human-readable and that bridging them to explainable subgoals remains open (Wu et al., 29 May 2026). This creates a persistent tension in the field: explicit textual CoT is easier to inspect, format-check, and annotate, but latent reasoning is often better aligned with continuous control and substantially cheaper at inference (Bai et al., 1 Feb 2026, Wu et al., 29 May 2026). A plausible implication is that future systems may need dual interfaces: a latent interface optimized for control and a partial symbolic projection optimized for diagnosis or supervision.

A second issue concerns timescale and structure. Continuous Reasoning identifies fixed slot count and chunk size as limitations and suggests dynamically adaptive reasoning horizons as a future direction (Wu et al., 29 May 2026). NIAF, from the action side, shows that discrete waypoint schemes impose rigid sampling rates and quantization artifacts, whereas continuous-time action manifolds admit analytic higher-order supervision and resolution-independent querying (Liu et al., 2 Mar 2026). This suggests that “continuous reasoning” may increasingly encompass both intermediate thought representations and the mathematical form of the action signal itself.

A third issue is supervision quality and reward specification. DeepThinkVLA states that its CoT annotations come from a semi-automated pipeline, that reward is sparse and terminal, and that sim-to-real transfer still requires handling domain gaps and safety constraints (Yin et al., 31 Oct 2025). VLA-R1 addresses reward design through verifiable affordance, trajectory, and format rewards, but its own framing implies that post-training quality depends strongly on which aspects of reasoning are actually rewardable (Ye et al., 2 Oct 2025). In driving and navigation, this difficulty is amplified by the need to balance numerical feasibility, causal reasoning, and real-time throughput.

Finally, several papers challenge the assumption that more visible CoT is necessarily better. Continuous Reasoning states that reasoning in VLA is less about extra tokens than about a shareable, verifiable internal language for action (Wu et al., 29 May 2026). MPCoT obtains long-horizon gains while generating zero reasoning tokens at inference (Zhang et al., 4 Jun 2026). LaRA-VLA reports that latent reasoning can reduce latency by up to 90% relative to explicit CoT while still outperforming prior VLA methods (Bai et al., 1 Feb 2026). These results indicate an ongoing shift from text-first reasoning to action-coupled reasoning interfaces whose success is judged by reuse, verification, robustness, and control-time alignment rather than by the presence of human-readable intermediate text alone.