- The paper introduces a multi-level framework that couples predictive correction with cross-level alignment across action, subgoal, and trajectory stages.
- It employs Sinkhorn-based optimal transport and InfoNCE losses to minimize semantic drift and error propagation in Vision-Language-Action agents.
- Empirical results show up to a 9% improvement on benchmarks like MineDojo, validating its robust mitigation of cascading failures.
ReCAPA: Hierarchical Predictive Correction to Mitigate Cascading Failures
Introduction and Problem Statement
Vision-Language-Action (VLA) systems have advanced the capabilities of embodied agents in naturalistic, long-horizon environments—ranging from manipulation to navigation—by leveraging LLM planning and multi-modal perception. However, persistent obstacles remain. A core pathology is the propagation and accumulation of local execution errors, whereby a misaligned or erroneous substep in a multi-stage plan compounds over subsequent steps, causing agents to experience severe semantic drift from their original intent. Existing VLA frameworks often employ static task decomposition and post-hoc correction that provide only limited capacity for dynamic, cross-level alignment, leaving execution-goal consistency unchecked across the plan-action-trajectory hierarchy. The empirical consequence is sharply degraded performance on benchmarks like AI2-THOR, VisualAgentBench, and MineDojo as task horizon grows.
Hierarchical Predictive Correction via ReCAPA
The Predictive Alignment and Planning Architecture (ReCAPA) introduces a multi-level framework for proactive mitigation of error propagation in VLA agents. The fundamental deviation from prior architectures is the explicit coupling of predictive correction and global alignment at three granularity levels: action, subgoal, and full trajectory. Central to this design is the Hierarchical Predictive Contrastive Correction (HPCC) module, which enforces compositional alignment via cross-level predictive losses and semantic contrast between predicted and target representations.
At the action level, fine-grained execution steps are embedded and synchronized with higher-level subgoal semantics, correcting for local deviations through predictive and contrastive objectives. The subgoal level forecasts the cumulative effect of recent trajectory segments and aligns their representation with realized rollouts, detecting and repairing misalignments before they escalate. At the trajectory level, prompt conditioning is maintained over the global rollout using a Sinkhorn-based optimal transport divergence, ensuring distributional alignment with the semantic intent specified at initialization.
Alignment and Correction Mechanisms
Sinkhorn-based Prompt-Trajectory Alignment
The trajectory-level semantic alignment is achieved using an entropy-regularized Sinkhorn optimal transport module, which quantifies and minimizes distributional discrepancy between the trajectory and the prompt in latent embedding space. This approach confers robustness, as global task structure is preserved regardless of local ambiguities or action permutations, and it avoids the instability introduced by overly sensitive local correspondence.
Score-Field Alignment for Local Corrections
Complementing global alignment, the Score-field module applies denoising score matching at the state/action and subgoal levels, generating vector fields that drive low-density, off-intent representations toward prompt-defined high-probability regions. This dual-level alignment ensures that both coarse (trajectory) and fine (action, subgoal) semantics remain simultaneously regularized and enables rapid error attenuation.
Hierarchical InfoNCE-based Contrastive Loss
At every level of abstraction, ReCAPA implements a cross-level prediction task using InfoNCE-style losses: a lower-level embedding predicts a higher-level state, which is then regularized using contrastive discrimination against hard negative samples crafted from LLM-generated alternatives. This methodology ensures that each abstraction layer maintains semantic consistency with the layer above, creating a feedback loop that enables both proactive (anticipatory) and reactive (corrective) alignment.
Evaluation Metrics: EPR and PAC
Recognizing the inadequacy of traditional terminal metrics for robustness assessment, ReCAPA proposes two diagnostic measures: Error Propagation Rate (EPR) and Propagation Attenuation Coefficient (PAC). EPR quantifies the increase in post-error risk across future steps—offering a direct lens into the system's vulnerability to error cascades—while PAC measures the exponential decay rate of risk following an initial error, distinguishing recovery capability from persistent drift. These metrics allow comparative models to be distinguished not only by final success rate, but also by their internal error dynamics and stability under long-horizon rollouts.
Across VisualAgentBench, AI2-THOR, MineDojo, and other standardized embodied reasoning environments, ReCAPA consistently achieves higher success rates than strong baselines, including proprietary LMMs such as Claude-4-Sonnet, GPT-4V, and Gemini-2.5, as well as open-source models like CogVLM2, Qwen-VL, and LLaVA-NeXT. Numerically, ReCAPA achieves up to +9% improvement in success rate on MineDojo, a +7% increase on AI2-THOR, and surpasses all competitors in AVG score (58.65) on VisualAgentBench. Notably, the method produces the lowest EPR values and the highest PAC decay rates among all tested models, directly confirming its superior propagation control.
Ablation studies reinforce the necessity of full hierarchical integration: removing any single abstraction (action, subgoal, or trajectory) or decoupling alignment modules invariably leads to substantial performance drops. The synergy of prediction and contrastive correction mechanisms is critical; disabling either results in excess local drift or delayed recovery, validating the theoretical rationale for unified, cross-level optimization.
Theoretical Implications
The alignment-based execution gradient framework underlying ReCAPA is theoretically justified: contrastive and Sinkhorn-based objectives directly minimize population-level divergence between agent rollouts and target plans. Joint hierarchical optimization provides provably stable convergence rates under standard assumptions, with sample efficiency and bounded error accumulation guaranteed by layered, structure-aware supervision.
Practical Implications and Future Directions
By proactively suppressing error cascades at multiple granularity levels, ReCAPA enhances long-horizon stability, compositionality, and recovery in VLA agents—qualities necessary for reliable deployment in open-ended, real-world scenarios. However, discrete hierarchical scoring can introduce latency in error correction between update intervals; the deterministic module also cannot represent alternative continuations under epistemic uncertainty. Addressing these limitations by introducing continuous or uncertainty-aware correction mechanisms represents a promising future research direction.
The two new diagnostic metrics—EPR and PAC—provide a more complete toolkit for characterizing agent robustness and recovery; future benchmarks should adopt them broadly. The architecture's success also motivates further exploration into joint alignment strategies across multi-agent and multi-modal planning systems, especially those that require persistent intent-to-execution consistency under environmental nonstationarity.
Conclusion
ReCAPA advances embodied agent planning through a hierarchical predictive correction and contrastive alignment framework capable of dynamically forecasting and correcting semantic deviations at every level of abstraction. Empirical results across multiple benchmarks demonstrate substantial gains in both terminal and propagation metrics over state-of-the-art baselines. The theoretical and practical advances herein outline a path toward robust, compositional, and self-consistent long-horizon reasoning in AI agents, with broad implications for future research in embodied cognition and VLA systems (2604.21232).