SITCOM: Scalable Inference for VLAs

• The paper introduces SITCOM, an inference-time planning framework that uses model-based rollouts to enhance VLA performance.
• It integrates a pretrained VLA with a learned dynamics model and a reward model to simulate and score multi-step action trajectories.
• Empirical results indicate a boost in task success from 48% to 72%, nearing the performance of an oracle planner in robotic control.

Scaling Inference-Time COMpute for VLAs (SITCOM) is an inference-time planning framework for robotic Vision-Language-Action (VLA) models that augments any pretrained VLA with model-based rollouts and reward-based trajectory selection, inspired by Model Predictive Control algorithm (Saxena et al., 5 Oct 2025). Rather than using a VLA as a one-shot policy that maps the current observation and language instruction directly to a single control command, SITCOM leverages a learned dynamics model to simulate multi-step action rollouts, scores candidate rollouts using rewards from simulator, and selects the best candidate plan for real-world execution (Saxena et al., 5 Oct 2025). In the reported SIMPLER evaluations, SITCOM combined with a good reward function improves average task completion rate from 48% to 72% using a trained dynamics model, with an oracle planner using the environment simulator reaching 76% (Saxena et al., 5 Oct 2025).

1. Research problem and conceptual position

SITCOM addresses three difficulties identified for robotic control: the high cost of collecting labeled data, limited generalization to unseen environments, and difficulties in planning over long horizons (Saxena et al., 5 Oct 2025). In this framing, VLA models are promising because they ground natural language instructions into single-step control commands, but they often lack mechanisms for lookahead and struggle with compounding errors in dynamic tasks (Saxena et al., 5 Oct 2025).

The framework therefore reallocates computation from training into deployment-time planning. Its central claim is not that the underlying VLA becomes intrinsically more capable, but that extra inference-time computation can be used to evaluate alternative futures before execution. SITCOM is explicitly presented as a way to transform one-shot VLAs into robust long-horizon planners through model-based rollouts and reward-based trajectory selection (Saxena et al., 5 Oct 2025).

This placement is consistent with the broader inference-time scaling literature, where repeated sampling and verification improve performance most directly in settings with reliable verification signals (Brown et al., 2024). SITCOM instantiates that general principle in robotics by replacing answer verification with trajectory scoring.

2. Core architecture

SITCOM is modular. The paper describes three main components: a VLA policy $\pi_{\mathrm{VLA}}$, a learned dynamics model $f_{\mathrm{dyn}}$, and a reward model $r$ (Saxena et al., 5 Oct 2025).

Component	Function	Paper description
VLA Policy $\pi_{\mathrm{VLA}}$	Maps image observation and task instruction to actions	Takes current image observation $\mathcal{I}$ and natural language task instruction $\mathcal{T}$ as input
Dynamics Model $f_{\mathrm{dyn}}$	Predicts future observations	Given an image $\mathcal{I}$ and action $a$, predicts the next image observation $\mathcal{I}'$
Reward Model $f_{\mathrm{dyn}}$0	Scores candidate rollouts	Scores candidate trajectories based on the initial and final observations and the task

The VLA policy supplies candidate actions. The learned dynamics model provides the forward simulation needed to evaluate those actions beyond the current timestep. The reward model provides the criterion for selecting among simulated futures. In the reported experiments, the rewards are simulator-based; the paper also notes task-specific metrics such as object-gripper gap and task completion (Saxena et al., 5 Oct 2025).

The modularity is important because SITCOM is presented as augmenting any pre-trained VLA rather than replacing the policy class. A common misconception is to interpret SITCOM as a new VLA architecture. The paper instead presents it as an inference-time wrapper around an existing policy, with the learned dynamics model and reward-based selection supplying the missing lookahead (Saxena et al., 5 Oct 2025).

3. Inference-time planning procedure

At each replanning step, SITCOM performs a planning loop over candidate actions and simulated futures (Saxena et al., 5 Oct 2025). The reported procedure is:

1. Sample $f_{\mathrm{dyn}}$1 candidate actions $f_{\mathrm{dyn}}$2 from the VLA for the current observation $f_{\mathrm{dyn}}$3 and task $f_{\mathrm{dyn}}$4, using high temperature for exploration.
2. For each candidate action, initialize a simulated trajectory from $f_{\mathrm{dyn}}$5.
3. Roll out the trajectory for $f_{\mathrm{dyn}}$6 steps by alternating the VLA and the dynamics model:
   • $f_{\mathrm{dyn}}$7
   • $f_{\mathrm{dyn}}$8
4. Score the final state of each trajectory with the reward model:

$f_{\mathrm{dyn}}$9

1. Select the highest-scoring trajectory:

$r$0

2. Execute actions from $r$1 in the real world, usually just the first action or a chunk before replanning (Saxena et al., 5 Oct 2025).

The paper also gives the rollout structure in compact form: $r$2 followed by repeated application of $r$3 and $r$4 until the reward is computed on the final simulated state (Saxena et al., 5 Oct 2025).

This planning loop is the operational meaning of “scaling inference-time compute” in SITCOM. Compute is spent on breadth through multiple candidate actions and on horizon through multi-step rollouts. The paper reports that increasing the number of candidate rollouts and rollout length improves performance up to a point, at the cost of increased compute, and notes that this workload can be parallelized (Saxena et al., 5 Oct 2025).

4. Learned dynamics model and Real2Sim adaptation

The learned dynamics model is a transformer-based encoder-decoder that encodes image observations as patches, concatenates them with action embeddings, and decodes to predict the next image or frame in pixel space (Saxena et al., 5 Oct 2025). It is trained autoregressively so that its own predictions can be fed back as inputs during multi-step rollout, matching the inference regime required by SITCOM (Saxena et al., 5 Oct 2025).

Training uses a combined L1 pixel-wise loss and Learned Perceptual Image Patch Similarity (LPIPS) loss (Saxena et al., 5 Oct 2025). The reported data pipeline has two stages. First, the model is pretrained on BridgeV2, described as a large-scale dataset with 60k human and scripted manipulation trajectories in varied scenes and object/task combinations, collected on a real WidowX-250 arm (Saxena et al., 5 Oct 2025). Second, it is fine-tuned on SIMPLER environments to reduce the Real2Sim gap (Saxena et al., 5 Oct 2025).

To reduce prediction drift over multiple rollout steps, the paper uses a Data Aggregation (DAgger)-inspired algorithm in which some rollouts use the model’s own predictions as input, aligning train and test distributions (Saxena et al., 5 Oct 2025). This is a consequential design choice because SITCOM depends on multi-step imagined futures rather than one-step next-frame prediction alone.

The paper evaluates the adapted world model with Frechet Inception Distance (FID) and Optical Flow Loss (OFL). Fine-tuning reduces FID from 17.0 to 11.2 and OFL from 1.67 to 0.99 (Saxena et al., 5 Oct 2025). In the paper’s interpretation, these numbers indicate better realism and temporal coherence after in-domain adaptation.

A recurring theme is that the dynamics model is not merely auxiliary. SITCOM’s performance depends on the quality of its simulated futures. The paper states that initial models struggled in simulation due to distributional mismatch in robot appearance, object physics, and backgrounds, and that both policy and dynamics model require target-domain adaptation for robust transfer (Saxena et al., 5 Oct 2025).

5. Empirical results and performance conditions

The reported evaluations are conducted across multiple tasks and settings in the SIMPLER environment (Saxena et al., 5 Oct 2025). The headline results compare zero-shot OpenVLA, fine-tuned OpenVLA-SFT, SITCOM with an oracle planner using the environment simulator, and SITCOM with the trained world model.

Method	Average success	Setting
OpenVLA	0.01	Zero-shot in SIMPLER
OpenVLA-SFT	0.48	Fine-tuned policy baseline
SITCOM (EnvSim)	0.76	Oracle planner using environment simulator
SITCOM (World Model)	0.72	Trained dynamics model

Task-level results reported in the paper include 0.66 on “Put Carrot on Plate,” 0.83 on “Put Spoon on Cloth,” 0.62 on “Stack Blocks,” and 0.79 on “Put Eggplant in Basket” for SITCOM with the world model (Saxena et al., 5 Oct 2025). The corresponding OpenVLA-SFT scores are 0.50, 0.63, 0.17, and 0.63 (Saxena et al., 5 Oct 2025).

Two conditions are emphasized repeatedly. First, reward quality is decisive. The paper states that SITCOM when combined with a good reward function can significantly improve task completion rate, and further notes that strong performance arises when the reward model is highly informative; with weaker or noisy reward signals, performance may become limited (Saxena et al., 5 Oct 2025). Second, Real2Sim adaptation is necessary. Pretraining on broad real datasets helps, but target-domain fine-tuning remains crucial for the policy and for the world model (Saxena et al., 5 Oct 2025).

These conditions also clarify the framework’s limits. The paper notes that current deterministic dynamics can struggle in stochastic or partially observed domains, and identifies probabilistic diffusion-based models and more general reward designs as future directions (Saxena et al., 5 Oct 2025). SITCOM therefore should not be interpreted as a universal substitute for robust environment modeling or reward specification.

6. Relation to adjacent VLA inference-time compute methods

SITCOM occupies one part of a wider design space for inference-time compute in robotic and multimodal systems. Other papers in that space allocate compute differently.

“RoboMonkey: Scaling Test-Time Sampling and Verification for Vision-Language-Action Models” studies test-time scaling through sampling and verification. It reports that action error follows an exponentiated power law with the number of generated samples, and that pairing existing VLAs with RoboMonkey yields a 25% absolute improvement on out-of-distribution tasks and 8% on in-distribution tasks (Kwok et al., 21 Jun 2025). Its deployment pipeline samples actions, applies Gaussian perturbation and majority voting, and then uses a VLM-based verifier to select the optimal action (Kwok et al., 21 Jun 2025).

“Realtime-VLA FLASH: Speculative Inference Framework for Diffusion-based VLAs” addresses latency rather than long-horizon planning. FLASH introduces a lightweight draft model, parallel verification via the main model’s Action Expert, and a phase-aware fallback mechanism. On LIBERO, it lowers task-level average inference latency to 19.1 ms, a 3.04x speedup, by replacing many 58.0 ms full-inference rounds with speculative rounds as fast as 7.8 ms (Niu et al., 13 May 2026).

“VLASH: Real-Time VLAs via Future-State-Aware Asynchronous Inference” overlaps execution and inference. It estimates the future execution-time state by rolling the robot state forward with the previously generated action chunk, achieving up to 2.03x speedup and reducing reaction latency by up to 17.4x compared to synchronous inference while fully preserving the original accuracy (Tang et al., 30 Nov 2025).

“Running VLAs at Real-time Speed” focuses on systems optimization. It reports pi0-level multi-view VLA at 30Hz frame rate and at most 480Hz trajectory frequency using a single consumer GPU, and a 100% success rate in grasping a falling pen task (Ma et al., 30 Oct 2025).

Outside robotics-specific settings, “Latency and Token-Aware Test-Time Compute” formulates inference-time scaling as dynamic compute allocation and method selection, explicitly balancing expected accuracy, token cost, and wall-clock latency (Huang et al., 11 Sep 2025). A plausible implication is that SITCOM represents the model-based rollout and reward-based trajectory selection end of the inference-time compute spectrum, whereas FLASH, VLASH, RoboMonkey, and real-time serving frameworks emphasize speculative execution, asynchronous control, sampling-and-verification, or systems-level elimination of inference overheads.

Taken together, these works show that “scaling inference-time compute” is not a single technique but a family of deployment-time strategies. SITCOM’s distinctive contribution within that family is to spend extra computation on multi-step imagined futures and reward-based plan selection, thereby supplying lookahead to VLA policies that otherwise operate as single-step controllers (Saxena et al., 5 Oct 2025).