MARS-Cache: Caching in MAR Models & Mars Sampling

• MARS-Cache is a dual framework that accelerates MAR generative models by caching token and condition redundancies, achieving up to 2.83× speedup.
• It employs periodic refresh and selective deep-layer computation to balance efficiency with minimal perceptual quality loss in image generation.
• For Mars Sample Return, MARS-Cache refers to a cache of sample tubes, detected via hybrid methods combining template matching and Mask R-CNN for reliable localization.

MARS-Cache refers to two distinct but foundational concepts in contemporary research: (1) a methodology for efficient caching and computational acceleration in masked autoregressive (MAR) generative models, primarily for image token generation ("MARS-Cache" or "LazyMAR" in deep learning); and (2) the physical cache of sample tubes on the Martian surface deployed by the Perseverance rover for collection and return to Earth, together with associated machine vision methods for localization during planetary sample return missions. Both uses of "MARS-Cache" are unified by the theme of enabling reliable retrieval—whether of neural features or physical artifacts—via efficient and robust data management.

1. Token Redundancy and Condition Redundancy in MAR Models

MARS-Cache ("LazyMAR") exploits two central forms of representational redundancy in masked autoregressive generative models (Yan et al., 16 Mar 2025):

• Token Redundancy: For a sequence of tokens $X$ and their representations $h^k_i \in \mathbb{R}^d$ at decoding step $k$, the cosine similarity $s^k_i = \cos\!_{\text{sim}}(h^k_i, h^{k-1}_i)$ is used to identify tokens whose semantic state remains stable over consecutive steps. If $s^k_i \geq \theta_{\text{token}}$ (with $\theta_{\text{token}}\approx 0.98$), the token is deemed redundant and its deep representations can be reused ("cached") in subsequent steps.
• Condition Redundancy: In classifier-free guidance (CFG), the difference $$r^k = F_{\text{uncond}}(X^k) - F_{\text{cond}}(X^k)$$ between unconditional and conditional outputs is observed to be low-variance across decoding steps. If $\|r^k - r^{k-1}\|_2 \leq \epsilon_{\text{cond}}$, the unconditional branch output can be approximated as $$F_{\text{uncond}}(X^k) \approx F_{\text{cond}}(X^k) + r^{k-1}$$, skipping redundant computation.

This dual exploitation of redundancy forms the mathematical and algorithmic foundation of the MARS-Cache approach.

2. Caching Mechanisms: Periodicity, Token Selection, and Algorithmic Flow

LazyMAR partitions inference into intervals of length $\tau$, with each interval starting with a "full-compute" refresh and followed by $(\tau-1)$ "cache-reuse" steps. The token cache algorithm involves:

• Forward propagation of all tokens through the first $L_0$ layers to generate early representations $E^k_j$.
• Cosine similarity testing to split tokens into $I_{\text{cache}}$ (cached, reused) and $I_{\text{compute}}$ (recomputed).
• For tokens in $I_{\text{cache}}$, deep-layer activations are loaded from cache; for $I_{\text{compute}}$, activations are forward-propagated and the cache updated accordingly.

The condition cache operates in parallel, maintaining a residual $r_{\text{cache}}$ to approximate the unconditional branch.

The periodic refresh—after every $\tau$ steps—prevents unbounded cache drift. Effective hyperparameters include $\theta_{\text{token}}$ (cosine threshold), $L_0$ (probe layers, typically 3), and $\tau$ (interval, e.g., 5-9 steps).

The importance of targeted token selection is empirically demonstrated: token skipping based on maximum cosine similarity achieves $2.83\times$ speedup with minor FID degradation, whereas random skipping at the same rate causes unacceptable quality loss (Yan et al., 16 Mar 2025).

3. Complexity Reduction and Empirical Acceleration

Under LazyMAR/MARS-Cache, computational savings are realized by:

• Skipping deep-layer processing for up to $84\%$ of tokens on most inference steps.
• Skipping the entire unconditional CFG branch (for up to $50\%$ further savings).
• Reducing per-layer workload to $\sim20\%$ for probe layers (always computed), and $\sim12.8\%$ for deep layers (on the $16\%$ of uncached tokens).

On ImageNet 256$\times$256 with MAR-H (943M parameters, 64 steps), total FLOPs drop from $69.06$T to $24.38$T, yielding a $2.83\times$ speedup and reducing GPU latency from $1.74$s to $0.75$s. FID degradation is minimal: for MAR-H, FID increases from $1.59$ to $1.69$ and IS remains unchanged ($299.1 \to 299.2$). Ablations confirm that both token and condition caches are necessary for full acceleration at negligible quality cost (Yan et al., 16 Mar 2025).

4. Cache Mechanisms in Masked Autoregressive Transformers

Complementary to LazyMAR, MARché provides a cache-aware attention and selective key/value (KV) refresh strategy for MAR generation (Jiang et al., 22 May 2025):

• Cache-aware Attention: Separates tokens into active and cached sets. Only the active set $A^{(t)}$ (just unmasked, recently generated, or contextually relevant tokens) has $K$/ $V$ recomputed; all other tokens use cached values with cosine similarities exceeding $0.95$.
• Two-path Attention Kernel: The attention kernel is split such that active-active and active-cached computations are handled separately, yielding exactly the same results as standard attention but at reduced cost.
• Selective KV Refresh: For each step, attention scores from newly generated tokens identify the $K$ most contextually important cached tokens for refresh.

This design permits up to $1.7\times$ speedups with $<0.4$ FID degradation on ImageNet-style benchmarks and no change to the underlying MAR architecture (Jiang et al., 22 May 2025). Table 1 summarizes empirical results:

Model	Latency (s/im)	FID	IS	Speedup
MAR-H	0.336	1.62	298.6	1.00
MARché-H	0.195	2.02	281.4	1.72

5. Integration, Application Scope, and Extension

Both LazyMAR and MARché are training-free and require no modification to network weights. Integration involves lightweight wrappers that implement the token and condition cache logic. No retraining is necessary; the system can be used as a plug-and-play acceleration module in any MAR transformer codebase (Yan et al., 16 Mar 2025, Jiang et al., 22 May 2025).

MARS-Cache methods are not limited to image generation. The caching schemes and periodic refresh mechanisms extend directly to:

• Text infilling with masked LLMs in arbitrary order.
• Audio token generation (e.g., SoundStream).
• Video generation and point-cloud synthesis with masked transformers.

This generality is enabled by the abstraction of redundancy and local token stability across sequential generative tasks.

6. Mars Surface Sample Cache: Machine Vision for Sample Localization

MARS-Cache also refers to the on-surface repository of sample tubes deployed by Perseverance for Mars Sample Return (MSR) (Daftry et al., 2021). Robust detection and localization of these sample tubes is a key computer vision challenge for the Sample Fetch Rover.

The two principal methods evaluated are:

• Geometry-Driven Template Matching: Uses a CAD-based library of tube contours and quantized gradient orientation filters for model-based, pose-aware matching. The similarity metric is cross-correlation of gradient orientations, enabling efficient, interpretable detection with explicit pose recovery via back-projection under planar-ground assumptions.
• Data-Driven Mask R-CNN: Employs a ResNet-50 FPN backbone, with RPN and mask heads trained on 824 annotated images from the JPL Mars Yard. The system generalizes across terrains and occlusion conditions, achieving [email protected] values above $0.91$ and F1-scores near $0.87$. Robustness across lighting and dust is superior to the template matcher; AP degrades $10\%$ with light dust vs. $40\%$ for template matching.

A hybrid system—template-matching for initial hypotheses, with CNN-based verification—is advocated to maximize both validation and detection performance under autonomy and reliability constraints.

7. Summary and Comparative Analysis

MARS-Cache-enabled acceleration for masked autoregressive models (both LazyMAR and MARché) yields substantial speed improvements through token-level and conditional redundancy exploitation, with empirical speedups of $1.7$–$2.8\times$ and minimal perceptual or quantitative quality loss (FID, IS). These approaches are architecture-neutral, training-free, and broadly applicable across transformer modalities.

For physical sample caches on Mars, rigorous benchmarking and comparative analysis of machine-vision methods reveal that deep learning (Mask R-CNN) outperforms model-based methods in detection, localization, and environmental robustness, but interpretability and verification favor hybrid systems (Daftry et al., 2021).

MARS-Cache thus represents a convergence of efficient retrieval—whether of data representations or physical artifacts—requiring advanced caching, selection, and detection strategies to optimize system performance and reliability across scientific and engineering domains.