STRIDE: AI Systems & Algorithms

• STRIDE is a suite of integrated frameworks and models in AI, robotics, and computational biology, characterized by modular design and analytic tractability.
• It leverages decomposition, agentic reasoning, and optimization techniques to enhance performance in tasks such as scene text recognition and reward design in RL.
• Empirical benchmarks demonstrate significant speed, efficiency, and accuracy improvements, making STRIDE a pivotal tool for advanced multi-domain AI applications.

STRIDE refers to a diverse set of frameworks, models, datasets, and algorithms across numerous domains in artificial intelligence, robotics, computer vision, scientific computing, interpretability, and computational biology. This entry provides a technical overview of major STRIDE systems as documented in primary arXiv publications, emphasizing their design, mathematical framework, algorithmic innovation, and empirical impact.

1. Core STRIDE Systems and Domains

STRIDE is used as an acronym for distinct systems, each with a specific domain of application:

• Scene Text Recognition In-Device: a resource-constrained scene text recognition system for mobile devices (Munjal et al., 2021).
• Reward Design and RL Optimization for Humanoid Robotics: a framework automating reward design and training for deep reinforcement learning (DRL) agents using agentic LLM-driven workflows (Wu et al., 7 Feb 2025).
• AI Modality Selection: a decision-theoretic evaluator for selecting between LLM calls, assistants, and agents, given task decomposition and risk analysis (Asthana et al., 1 Dec 2025).
• Spatio-Temporal Road Image Dataset and World Model: an open, large-scale graph-structured dataset for navigation, geolocation, and temporal world modeling, supporting sophisticated multi-modal AI agents (Carrión et al., 12 Jun 2025).
• Training Data Attribution in LLMs: a compressive-sensing-based activation-space method for causal training data influence estimation in large models (Dagli et al., 3 Jun 2026).
• Physical System Dynamics Modeling: a structured Lagrangian and flow-matching approach for splitting conservative and stochastic components in robot dynamics (Kotecha et al., 9 Mar 2026).
• Training-Free Diversity in Diffusion Models: a PCA-projected, spatial-noise perturbation for single/few-step vision diffusion models (Yadav et al., 12 May 2026).
• Program Analysis (Decompiler Typing/Naming): an n-gram-based method for variable type/name recovery in reverse engineering (Green et al., 2024).
• Ultrasound Computed Tomography: a Python HPC platform for PDE-constrained imaging problems (ultrasound tomography) (Cueto et al., 2021).
• MRI Interference Suppression: total-variation regularized subtraction for EMI suppression in MR image formation (Mertens et al., 17 Nov 2025).
• Verifiable RL with Strategic Reasoning: a trajectory-level, discriminative n-gram credit assignment for RL post-training in LLMs and agentic reasoning (Zhao et al., 14 Jun 2026).
• Bipedal Robotics Platform: an open-source, modular, and reconfigurable bipedal hardware platform and experimental setup (Huang et al., 2024).
• Occlusion-Robust 3D Human Pose Estimation: a test-time training approach using motion priors to robustify 3D pose estimation under severe occlusion (Lal et al., 2023).
• Functional XAI via Subset-Free Decomposition: a kernel-based, non-enumerative approach to functional model explanation beyond scalar attributions (Ko, 11 Sep 2025).
• Environmental Sensing for Collision Risk: a panoramic dataset and multi-task model for road environment detection and pedestrian collision prediction (González et al., 2023).
• Spatiotemporal Field Reconstruction: a recurrent-implicit decoder for reconstructing continuous fields from sparse sensor data (learning parametric PDE dynamics) (Tong et al., 4 Feb 2026).
• LLM-Driven Equation Discovery: a self-reflective, agentic symbolic regression loop integrating generation, critic-driven repair, and memory (Su et al., 18 May 2026).
• Multi-Hop Question Answering: a meta-planned, supervisor-controlled, and grounded-executor framework for multi-hop QA (Chen et al., 19 Apr 2026).
• Discrete Sequence Refinement via Edit Trajectories: an LLM-based post-training method for discrete biosequence optimization via atomic edit planning (Zhang et al., 3 Mar 2026).

2. Key Principles and Mathematical Frameworks

STRIDE systems vary widely in technical construction, but several share unifying principles:

• Decomposition and Modularization: STRIDE frequently leverages architectural separation—e.g., in world modeling, text recognition, XAI, and multi-hop QA—by explicit partitioning of task, control, execution, or signal pathways.
• Functional Decomposition: In explainability, STRIDE eschews exponential subset enumeration by recursive centering of kernel-based expansions in RKHS, establishing orthogonality of functional components and enabling analytic projections (Ko, 11 Sep 2025).
• Agentic Reasoning and Self-Reflexivity: Both RL reward design (Wu et al., 7 Feb 2025) and equation discovery (Su et al., 18 May 2026) treat generation and evaluation as roles in a positive feedback loop, with policy, critic, executor, and memory modules exchanging functional statistics.
• Statistical and Sparsity-Based Recovery: Training data attribution formulates instance-level influence recovery as a sparse Lasso problem, leveraging expander matrices and activation-space steering (Dagli et al., 3 Jun 2026).
• Optimization Under Constraints: Physical system modeling (robotics, MRI, tomography) integrates convex/structured optimization, Lagrangian mechanics, and flow-matching generative models (Kotecha et al., 9 Mar 2026, Cueto et al., 2021, Mertens et al., 17 Nov 2025).

3. Algorithmic Innovations and Architecture

STRIDE systems deliver domain-optimized architectures:

STRIDE Instance	Core Algorithmic Elements	Empirical Highlights
Scene Text Recognition (Munjal et al., 2021)	Shallow conv backbone + CBAM + Bi-LSTM + CTC	0.88M params, <2.5ms/word, SOTA on-device trade-offs
RL Reward Design (Wu et al., 7 Feb 2025)	LLM-driven agentic pipeline, zero-shot code, reflection loop	3x efficiency vs GPT-4-based EUREKA; >2m/s locomotion
Data Attribution (Dagli et al., 3 Jun 2026)	Compressive-sensing, low-rank basis, activation steering	13x speedup, ~0.167 Spearman, 90%+ code success
Diffusion Diversity (Yadav et al., 12 May 2026)	PCA-projected pink noise injected into early transformer blocks	-7.5% InBSim, Pareto-optimal diversity/fidelity frontier
XAI Functional Decomposition (Ko, 11 Sep 2025)	RKHS-centered kernel, analytic projections, “component surgery”	3x faster than TreeSHAP, 0.81–0.999 R², functional interaction maps
Spatiotemporal Road Dataset (Carrión et al., 12 Jun 2025)	Graph-structured nodes; multi-modal; DFS-based augmentation	6.3M sequences, 131k nodes, 80B tokens, tasks: georef, planning
Multi-hop QA (Chen et al., 19 Apr 2026)	Meta-planner, supervisor, modular executor, self-supervised FT	+0.085 F1 over baselines, 50–80% fewer tokens or time

Certain STRIDE algorithms use distinctive mathematical constructs. For example, in XAI (Ko, 11 Sep 2025), the centered product kernel for subset $S$ is recursively defined as

$$K_S^{(c)}(x_S, t_S) = K_S(x_S, t_S) - \sum_{R \subsetneq S} K_R^{(c)}(x_R, t_R)$$

while in data attribution (Dagli et al., 3 Jun 2026), recovery is performed via

$$\hat w(x) = \arg\min_w\,\tfrac{1}{2} \|y_x - Mw\|_2^2 + \lambda \|w\|_1$$

across an expander-structured binary $M$.

4. Empirical Evaluation and Performance Benchmarks

STRIDE frameworks are consistently accompanied by comprehensive benchmark evaluations matching or surpassing specialist baselines:

• Scene Text Recognition: STRIDE (0.88M params) achieves 88.4% accuracy on ICDAR-13 with 2.44ms/word inference, outperforming Google ML Kit (>15MB, 20ms) and Paddle PP-OCR (1.8M, 5ms) in both metric footprint and accuracy (Munjal et al., 2021).
• RL Reward Design: STRIDE improves Max Success Scores by 250% over template-based EUREKA on humanoid locomotion, attaining near-human speeds under wave/random terrains; code success rate >90% (Wu et al., 7 Feb 2025).
• LLM Data Attribution: STRIDE achieves top linear d-Modeling Scores (0.158–0.167) at 13x speedup vs LoGRA/AirRep, and error recall@100 of 74.2% in contamination detection (Dagli et al., 3 Jun 2026).
• Diversity in Diffusion: On COCO, STRIDE reduces InBatchSim by up to –7.5%, achieving CLIP alignment improvements with no retraining for fast diffusers (Yadav et al., 12 May 2026).
• Biped Robotics: STRIDE has sub-2k hardware BOM, modular terrain, push-force quantification, and supports real-time gait adaptation and robust performance (Huang et al., 2024).
• Spatiotemporal Field Reconstruction: Across PDE benchmarks, STRIDE-FMMNN attains the lowest relative errors (e.g., SWE 2.78%), outperforming SHRED and SHRED-ROM, and supporting super-resolution and sparse sensor regimes (Tong et al., 4 Feb 2026).

5. Limitations, Trade-Offs, and Extensions

Every STRIDE instantiation explicitly articulates domain-specific trade-offs and open issues:

• Scene Text Recognition: Model remains restricted to Latin scripts; larger character sets would further increase parameter/evaluation costs (Munjal et al., 2021).
• RL Automation: Current agentic-LLM pipelines rely on code compatibility and may be gated by simulator fidelity and code-generation accuracy (Wu et al., 7 Feb 2025).
• Training Data Attribution: The additive-influence assumption may not hold under severe distribution shift/strong nonlinear memorization (Dagli et al., 3 Jun 2026).
• PCA Diversity Injection: Without on-manifold projection, noise injection can degrade sample quality or have zero effect due to model invariance (Yadav et al., 12 May 2026).
• Functional XAI: While scalable, STRIDE imposes RKHS kernel choices and focuses on moderate-dimension tabular data; TreeSHAP or LIME may be preferable in small-$d$ or highly optimized codebases (Ko, 11 Sep 2025).
• MRI Interference: TV-regularized STRIDE can still inject sensor noise into the image, unless SVD-based denoising of EMI sensors is applied; L1-TV solvers may outperform under severe under-sampling (Mertens et al., 17 Nov 2025).
• Symbolic Discovery: Parameter-role identification for mixed fitting may misclassify in highly coupled symbolic forms, and semantic equivalence detection remains string-based (Su et al., 18 May 2026).

6. Impact and Broader Significance

STRIDE systems have advanced the capabilities and efficiency of AI across:

• Resource-constrained vision and OCR: Demonstrating sub-megabyte, real-time, high-accuracy text recognition (Munjal et al., 2021).
• Data attribution in massive models: Enabling retraining-level causal analysis and policy interventions at scale (Dagli et al., 3 Jun 2026).
• Automated physical system modeling: Enabling robust, uncertainty-aware control in contact-rich robotic environments (Kotecha et al., 9 Mar 2026).
• Design-time modality selection: Systematically reducing over-engineering and compute costs in enterprise AI deployments (Asthana et al., 1 Dec 2025).
• Scientific discovery: Empowering autonomous, repairable, and memory-augmented pipelines for symbolic regression and scientific law recovery (Su et al., 18 May 2026).
• Biosequence optimization: Realizing compositional, verifiable, reward-aligned edit planning in discrete, highly constrained sequence spaces (Zhang et al., 3 Mar 2026).
• Generative modeling and world exploration: Supporting navigation, planning, and embodied reasoning for generalist agents via large-scale, spatio-temporally indexed datasets (Carrión et al., 12 Jun 2025).

Across domains, the defining properties of STRIDE instances are modularity, analytic tractability, causal grounding, and empirical competitiveness against or beyond state-of-the-art in their respective areas.