Test-Time Scaling for World Action Models via Zero-Shot Geometric Evaluation

Abstract: Test-time scaling improves foundation-model inference by spending additional computation, but robot control requires deciding whether extra compute is useful before executing an action. World Action Models (WAMs) make this decision natural: each rollout exposes both an action chunk and predicted future observations. We propose \methodgated, a training-free selective test-time scaling framework for WAMs. We first instantiate \method, a fixed-budget Best-of-$N$ selector that ranks sampled rollouts by cross-view depth reprojection consistency of their predicted futures, computed with a frozen geometry foundation model. \methodgated\ adds a lightweight action--future consistency gate that invokes \method\ only when the initial rollout appears internally inconsistent. Across five benchmark--backbone settings on RoboCasa, LIBERO Long, and RoboTwin~2.0, fixed-budget \method\ improves $N{=}8$ task success in every setting, e.g., raising the RoboCasa group average from $66.3\%$ to $68.4\%$ with Cosmos Policy and from $80.8\%$ to $82.5\%$ with X-WAM. With gating enabled, \methodgated\ recovers on average $74.8\%$ of the always-on success gain while triggering additional sampling on only $26.2\%$ of decision points. Offline diagnostics show that cross-view reprojection is a strong task-label-free selector, and we identify false low-score selections as a failure mode that helps explain why performance can saturate or degrade as $N$ increases.

• The paper introduces Gated GeoBoN, a training-free method that uses action–future consistency to trigger extra sampling and cross-view depth reprojection to select coherent robot rollouts.
• Across five benchmark–backbone settings, fixed-budget GeoBoN improves success by 1.1–2.1 percentage points, while the gated version recovers 63.6–85.7% of always-on gains at only 14.2–34.5% of decision points.
• The analysis shows larger Best-of-N pools can select spurious geometric outliers, with false selections rising to 32% at N=16, supporting moderate budgets and selective test-time computation.

Gated GeoBoN is a training-free, selective test-time scaling framework for World Action Models (WAMs) that exploits the dual outputs of WAM inference—action chunks and predicted future observations—to decide both when to spend additional compute and which candidate rollout to execute. The framework consists of two components: GeoBoN, a fixed-budget Best-of-$N$ selector that ranks sampled rollouts by cross-view depth reprojection consistency using a frozen geometry foundation model (VGGT-$\Omega$), and an action–future consistency gate that invokes this selection only when the initial rollout appears internally inconsistent (2607.17454).

Motivation and positioning

WAMs such as Cosmos Policy, DreamZero, LingBotVA, X-WAM, and Motus jointly predict future visual observations and action sequences from multi-view observations and language instructions. This exposes an imagined rollout before execution, creating an opportunity for test-time selection that action-only policies lack: the robot can audit whether the imagined future is internally consistent rather than comparing actions alone.

Existing test-time scaling methods for robotic policies—RoboMonkey's trained VLM verifier, RoVer's process reward model, MG-Select's model-internal masked-distribution signals, WAV's forward–inverse asymmetry—either spend a fixed sampling budget at every decision point or rely on learned, model-internal, or interface-specific signals. The paper positions itself as orthogonal to work improving WAM backbones themselves and complementary to efficiency methods that optimize architecture, denoising trajectories, or execution horizons. The central question is when extra compute is worth spending at a given control step, and how to choose among sampled futures without task success labels, ground-truth futures, online environment rollouts, or WAM-specific value heads.

Method

The framework audits rollouts along two axes of failure with different costs. The first axis tests whether the generated action agrees with the motion visible in the predicted future; the second tests whether predicted futures from different camera views correspond to a single coherent 3D scene.

Action–future gate. For the initial rollout $\tau_1$, dense optical flow (Farneback) is computed between the current primary-view observation and the predicted future frame. End-effector motion implied by the generated action chunk—via forward kinematics for joint-position actions or accumulated displacement for delta actions—is projected into the primary camera. For each moving arm (idle arms below a 1 cm threshold are filtered), optical flow is averaged inside a capsule-shaped region around the projected end-effector trajectory, and cosine agreement between visual and projected motion is computed. The gate triggers if any moving arm has agreement below $\tau_{\mathrm{gate}} = -0.2$; otherwise the initial chunk executes directly.

Cross-view geometric evaluator. When triggered, $N_{\max}-1$ additional rollouts are sampled. Each candidate's predicted primary- and wrist-view frames are fed to frozen VGGT-$\Omega$, which estimates camera geometry and per-pixel depth. The wrist-view 3D point map is projected into the primary camera frame, and the score is the mean absolute log-ratio between projected depth and directly predicted depth over pixels with valid projection, positive depth, and VGGT confidence above 0.5. The logarithmic ratio reduces sensitivity to absolute depth scale; the candidate with lowest inconsistency is executed. All thresholds are fixed across experiments, and no component requires training.

Fixed-budget results

Across all five benchmark–backbone settings on RoboCasa, LIBERO Long, and RoboTwin 2.0, fixed-budget GeoBoN improves $N{=}8$ success over single-rollout baselines:

Setting	Baseline	GeoBoN ($N{=}8$)	Δ (95% CI)
RoboCasa / Cosmos Policy	66.3%	68.4%	+2.1 [+0.6, +7.3]
RoboCasa / X-WAM	80.8%	82.5%	+1.7 [+0.3, +2.6]
LIBERO Long / Cosmos Policy	97.5%	99.3%	+1.8 [+0.4, +2.2]
LIBERO Long / LingBotVA	97.2%	98.3%	+1.1 [+0.2, +1.5]
RoboTwin 2.0 / Motus	87.8%	89.9%	+2.1 [+0.2, +3.4]

Category-level results are mixed and instructive. With X-WAM on RoboCasa, Door/Drawer Manipulation degrades from 96.7% to 92.1%, while Coffee Making rises sharply from 73.3% to 83.7% at $N{=}8$ and 92.0% at $N{=}16$. Geometric reranking helps most when tasks leave room for rejecting implausible futures but can hurt when the initial rollout is already strong. Notably, several settings saturate or slightly degrade at $\Omega$0: fixed-budget geometric selection is not monotonic in candidate count, a point the paper addresses explicitly rather than treating as incidental.

Gated results

With gating enabled ($\Omega$1), Gated GeoBoN triggers additional sampling on only 14.2–34.5% of decision points while recovering 63.6–85.7% of the always-on success gain, averaging 74.8% recovery at a 26.2% average trigger rate. Latency reductions are substantial: on RoboCasa/X-WAM, gated inference averages 3.11 s versus 9.67 s always-on, reaching 82.1% versus 82.5%. On LIBERO Long/Cosmos Policy it reaches 98.8% with a 14.2% trigger rate and near-baseline latency (0.96 s vs. 0.90 s). Gating thus preserves most of the benefit at roughly one-third of the cost in the most expensive settings.

Diagnostics

Selector ablation. Against two training-free alternatives—a VGGT confidence-based selector and a reimplemented future-consensus selector (the original implementation is not public)—cross-view reprojection achieves the best offline error recovery in three of four settings and is the only selector that improves closed-loop success over baseline in all settings. Confidence-only ranking is unstable and often reduces online success. The paper concedes that offline error recovery is only an approximate proxy for closed-loop success.

Gate ablation. Compared against a random trigger at equal trigger rates, the gate raises the GeoBoN help rate substantially on RoboCasa/X-WAM (67.2% vs. 43.2%), RoboCasa/Cosmos Policy (63.5% vs. 43.2%), and LIBERO/LingBotVA (69.4% vs. 47.8%), but only marginally on RoboTwin/Motus (75.8% vs. 71.7%). The signal is informative but not uniformly strong across domains.

Value-head comparison. Using identical candidate budgets, GeoBoN consistently outperforms the Cosmos Policy value head by +1.3 to +2.3 points on RoboCasa and +0.7 to +1.3 points on near-saturated LIBERO Long, establishing cross-view geometric consistency as a competitive training-free alternative to learned model-internal selectors.

Failure analysis

The paper identifies false low-score selections as the mechanism behind non-monotonic scaling: cases where Best-of-$\Omega$2 selects a candidate with unusually low reprojection error despite being visually near-identical (LPIPS < 0.02) to another candidate with no visible geometric difference. The false-selection rate grows from 7% at $\Omega$3 to 12% at $\Omega$4 and 32% at $\Omega$5, consistent with a multiple-comparisons effect: larger pools increase both the chance of finding a genuinely better rollout and the chance of selecting a spurious evaluator outlier. This explains the saturation observed at large budgets and motivates moderate budgets combined with selective invocation—the design rationale for Gated GeoBoN itself.

Limitations and open questions

Several limitations are stated plainly. The gate's discriminative value varies across domains, being weak on RoboTwin/Motus. Category-level degradation on already-strong tasks (X-WAM door/drawer) shows the selector can actively harm performance when the initial rollout is reliable. The consensus baseline was reimplemented from a paper description rather than evaluated with original code. Offline error recovery correlates imperfectly with closed-loop success. Open questions include how to make the geometric evaluator robust to spurious outliers so that scaling remains monotonic, whether the gate thresholds transfer without retuning to new embodiments or camera configurations, and how the approach extends beyond the bimanual manipulation settings tested here.

Conclusion

Gated GeoBoN demonstrates that the futures already exposed by WAM rollouts suffice for training-free test-time decisions: cheap action–future flow consistency allocates compute, and cross-view depth reprojection via a frozen geometry foundation model ranks candidates. It improves success in every benchmark–backbone setting evaluated, recovers roughly three-quarters of the always-on gain at about a quarter of the trigger rate, outperforms a learned value head under equal budgets, and provides a concrete failure-mode analysis explaining why Best-of-$\Omega$6 saturates at large candidate counts.