Technical Summary: REIS – Redundancy Elimination for Inference-efficient Systems

Problem Statement

Recent robotic policies leveraging Large Language Models (LLMs) and Vision-Language Models (VLMs) have demonstrated superior semantic planning and generalization through "reason-and-act" paradigms (e.g., CoT-VLA, DiffusionVLA). However, these reasoning-heavy approaches suffer from high inference latency, often conflicting with the strict real-time constraints (10–100 ms) required for robotic control. Existing solutions, such as hierarchical or dual-process architectures, attempt to separate slow semantic reasoning from fast low-level action generation. While effective, these methods treat reasoning latency as an unavoidable cost, creating a fragile trade-off where frequent reasoning updates negate latency benefits, while infrequent updates lead to stale contextual understanding.

The authors identify that the primary bottleneck is not action generation itself, but temporal redundancy in the reasoning process. In dynamic environments, consecutive observations frequently yield identical actions and subgoals, yet standard pipelines re-infer these states from scratch. This redundancy causes observations to become stale before execution, paradoxically reducing reactivity and robustness.

Methodology: REIS Framework

To address this, the authors propose REIS (Redundancy Elimination for Inference-efficient Systems), a human cognition-inspired, dual-process framework that minimizes unnecessary reasoning while preserving semantic adaptability. The architecture separates decision-making into two distinct systems:

System One: Fast Intuitive Decision Making

System One operates asynchronously at the frame rate to bypass heavy inference when the environment is stable. It consists of two components:

1. EMA-HSVS (Ego-Motion Aware Head-Selective Visual Similarity): A lightweight scene gating mechanism. Instead of processing all visual features, it selects specific transformer heads sensitive to geometric variations. It computes cosine similarity between the current frame and the last inference trigger, after applying a "Self-Mask" to filter out robot-arm-induced pixel differences. System Two is only invoked if the similarity falls below a predefined threshold, indicating a structural scene change.
2. KV-Steered Affordance Router: A mechanism to verify subgoal progress and action validity without full autoregressive reasoning. It injects pre-computed steering vectors (derived from the difference between reasoning-rich and label-only KV states) into the cached Key-Value (KV) states of the model. This biases the model toward deterministic affordance decisions (e.g., "keep current action" or "advance subgoal") with zero token generation overhead.

System Two: Deliberative Planning

System Two executes high-level semantic reasoning and multi-step replanning only when escalated by System One (e.g., during ambiguous states, failures, or significant structural changes). It comprises:

• Online Deliberative Reasoning: Generates high-level subgoals and low-level actions.
• Offline Anticipatory Logic Synthesis: Constructs the KV steering tensors used by System One. This involves simulating positive (reasoning-rich) and negative (label-only) trajectory pairs to compute mean element-wise differences in KV states, creating steering vectors that guide System One's routing decisions.

The framework adapts to task modalities: for navigation, EMA-HSVS actively filters macro visual noise; for manipulation, EMA-HSVS is deactivated to maintain sensitivity to local object-level changes, relying instead on the Affordance Router.

Key Contributions

1. Lightweight Vision-Encoder Head Selection: A method to prune redundant visual tokens by identifying and utilizing only the transformer heads most sensitive to geometric variations, enabling efficient perception.
2. Priming-Effect-Inspired Inference Mechanism: A novel approach leveraging historical context via KV-cache steering to bypass redundant reasoning steps, accelerating LM-based robotic control without sacrificing contextual adaptability.
3. Comprehensive Validation: Demonstration of the framework across diverse foundation model architectures (Qwen-3-VL, DeepThink) and pipelines (navigation and manipulation), showing substantial latency reductions with minimal impact on task success in both simulation (ALFRED) and real-world environments.

Experimental Results

The framework was evaluated on an NVIDIA Jetson Orin NX (16GB) using Qwen-3-VL (4B) and DeepThink (3B) models.

• Efficiency vs. Accuracy Trade-off: REIS achieved an average execution speedup of 15.09× across tasks while incurring only a 3.73% degradation in episode accuracy compared to a naive baseline that infers every frame.
  • Navigation: Achieved a 27.35× speedup (0.641 accuracy vs. 0.680 baseline) by suppressing redundancy during long-horizon movements.
  • Manipulation: Achieved a 2.34× speedup (0.729 accuracy vs. 0.760 baseline).
  • ALFRED Benchmark: Achieved a 15.58× speedup (0.512 accuracy vs. 0.554 baseline).
• Latency Reduction: The Affordance Router reduced inference latency from 43.59 seconds (autoregressive reasoning) to 406.21 ms (KV-steered routing) while improving classification accuracy from ~0.75 to 0.913.
• Failure Detection: In manipulation tasks, REIS reduced failure detection latency to 0.933 ± 0.874 s, a 2.21× improvement over periodic architectures that suffer from blocking autoregressive generation.
• Trigger Precision: The EMA-HSVS module achieved over 97% recall for true state transitions and 94.6% precision in filtering redundant frames.

Significance and Claims

The paper claims that REIS demonstrates that temporal continuity can be treated as an asset rather than a computational burden. By directly minimizing unnecessary reasoning rather than decoupling it from acting, robotic systems can maintain an active reasoning loop without sacrificing real-time reactivity.

The authors emphasize that their approach:

• Eliminates Redundancy: Proves that continuous per-frame reasoning is often unnecessary, as meaningful updates occur primarily at structural transitions.
• Enables Edge Deployment: Successfully runs complex reasoning-based policies on resource-constrained hardware (Jetson Orin NX) by reducing the computational footprint to 10–12 GB.
• Zero-Shot Utility: Achieves these gains using off-the-shelf foundation models without requiring task-specific fine-tuning for the gating mechanisms, relying instead on offline synthesis of steering vectors.

The work suggests a shift in design philosophy for Embodied AI: moving from "reasoning at every step" to "reasoning only when necessary," thereby resolving the conflict between semantic richness and real-time control.