Related Work Report: Intent-Predictive Operating Systems (Early)¶

Project context: DiPECS — a cloud-LLM-driven distributed intent operating system prototype targeting Android API 33, with privacy air-gap enforcement, deterministic state machine transitions, and strict mechanism-policy separation.

Date: 2026-03-19

Table of Contents¶

1. LLM-Driven / AI Operating Systems
2. Mobile App & Intent Prediction Systems
3. Privacy-Preserving AI for User Behavior
4. Cloud-Edge Hybrid OS & Proactive Systems
5. Comparative Analysis
6. Benchmark Landscape Summary
7. Differentiation of DiPECS
8. References

1. LLM-Driven / AI Operating Systems¶

1.1 AIOS: LLM Agent Operating System¶

• Authors: Kai Mei, Xi Zhu, Wujiang Xu, Wenyue Hua, Mingyu Jin, Zelong Li, Shuyuan Xu, Ruosong Ye, Yingqiang Ge, Yongfeng Zhang
• Year: 2024 (accepted COLM 2025)
• Paper: arXiv:2403.16971

Core approach: OS kernel architecture that isolates LLM-specific services (scheduling, context management, memory management, storage, access control) from agent application logic. Provides an SDK so agent frameworks (ReAct, Reflexion, Autogen, MetaGPT) can run concurrently atop a shared LLM resource managed by the kernel.

Benchmarks & Evaluation:

Limitations: Modest accuracy improvements (1–2 pp); no cross-machine distributed scheduling; no privacy layer.

1.2 OS-Copilot (FRIDAY): Towards Generalist Computer Agents with Self-Improvement¶

• Authors: Zhiyong Wu, Chengcheng Han, Zichen Ding, Zhenmin Weng, Zhoumianze Liu, Shunyu Yao, Tao Yu, Lingpeng Kong
• Year: 2024
• Paper: arXiv:2402.07456

Core approach: Framework for building generalist agents that interact across all OS elements — web, terminal, filesystem, multimedia, third-party apps. Accumulates reusable skills and self-improves.

Benchmarks & Evaluation:

Limitations: Relies on GPT-4 as backbone (cost/latency). No mobile or embedded OS evaluation.

1.3 UFO / UFO2: UI-Focused Agent for Windows OS Interaction¶

• Authors: Chaoyun Zhang et al. (Microsoft Research)
• Year: 2024 (UFO), 2025 (UFO2)
• Papers: arXiv:2402.07939, arXiv:2504.14603

Core approach: Dual-agent architecture (HostAgent + AppAgent) where GPT-Vision observes Windows GUI via UI Automation APIs. UFO2 extends to a full "Desktop AgentOS" with hybrid GUI-API control and speculative multi-action planning.

Benchmarks & Evaluation:

Limitations: Windows-only; tightly coupled to UI Automation backend; struggles with apps lacking UIA support.

1.4 SWE-agent: Agent-Computer Interfaces for Automated Software Engineering¶

• Authors: John Yang et al. (Princeton University)
• Year: 2024 (NeurIPS 2024)
• Paper: arXiv:2405.15793

Core approach: Introduces Agent-Computer Interfaces (ACI) — curated commands with guardrails for LM agents to explore repositories and fix bugs autonomously.

Benchmarks & Evaluation:

Limitations: 87.5% of issues still unsolved; high token cost; evaluated primarily on Python repos.

1.5 MemGPT: Towards LLMs as Operating Systems¶

• Authors: Charles Packer et al. (UC Berkeley)
• Year: 2023 (revised February 2024)
• Paper: arXiv:2310.08560

Core approach: OS virtual memory analogy for LLM context management — hierarchical memory (main context = RAM, archival = disk) with self-directed paging and interrupt-driven control flow.

Benchmarks & Evaluation:

Limitations: Every memory operation adds LLM inference cost; no formal memory consistency guarantees.

1.6 LLM as OS, Agents as Apps (Vision Paper)¶

• Authors: Yingqiang Ge et al. (Rutgers University)
• Year: 2023
• Paper: arXiv:2312.03815

Core approach: Conceptual taxonomy framing the LLM as an OS kernel providing system calls for agent applications. Distinguishes System Agents (OS services) from User Agents (application tasks).

Benchmarks: None — purely a position/vision paper. Lays the conceptual foundation for the AIOS implementation.

2. Mobile App & Intent Prediction Systems¶

2.1 FALCON: Fast App Launching via Predictive User Context¶

• Authors: Tingxin Yan, David Chu, Deepak Ganesan, Aman Kansal, Jie Liu
• Year: 2012
• Venue: ACM MobiSys

Core approach: Predicts next app launch using contextual signals (location, time-of-day, sensors). Cost-benefit learning algorithm weighs preloading benefit against energy cost.

Benchmarks & Evaluation:

Limitations: Windows Phone only; 16 users; no inter-app sequential modeling.

2.2 AppUsage2Vec: Modeling Smartphone App Usage for Prediction¶

• Authors: Sha Zhao et al.
• Year: 2019
• Venue: IEEE ICDE

Core approach: Doc2Vec-inspired model with app-attention mechanism encoding user personalization, temporal context, and app-transition patterns into a unified embedding space.

Benchmarks & Evaluation:

Limitations: Non-public carrier-side dataset; no rich multi-modal context.

2.3 DeepAPP: Deep Reinforcement Learning for App Usage Prediction¶

• Authors: Zhihao Shen et al.
• Year: 2019
• Venue: ACM SenSys

Core approach: Model-free deep RL agent predicting next app from complex environmental context, with online adaptation and lightweight personalized agents.

Benchmarks & Evaluation:

Limitations: Only 29-user field study; cold-start problem unaddressed.

2.4 Atten-Transformer: Progressive Temporal Attention for App Prediction¶

• Authors: Longlong Li, Cunquan Qu, Guanghui Wang (Shandong University)
• Year: 2025
• Paper: arXiv:2502.16957

Core approach: Integrates sinusoidal temporal positional encoding with time-aware attention inside a Transformer network, dynamically weighting critical usage moments.

Benchmarks & Evaluation:

Limitations: Performance plateaus after 6 days of history; overfitting risk; no multi-modal context.

2.5 GNN + Self-Attention Enhancement for Next App Prediction¶

• Authors: Junxin Chen et al.
• Year: 2025
• Venue: Scientific Reports (Nature)

Core approach: Combines gated graph neural networks (GGNN) for session dynamics with Transformer self-attention for long-term patterns. Dual graph structures for sequential and personalized interactions.

Benchmarks & Evaluation:

Limitations: Only 7 days of data; no seasonal modeling; cold-start unaddressed.

2.6 Appformer: Progressive Multi-Modal Data Fusion¶

• Authors: Chuike Sun et al.
• Year: 2024
• Paper: arXiv:2407.19414

Core approach: Two-module Transformer architecture with cross-modal attention fusing app sequences, POI vectors, user IDs, and temporal features.

Benchmarks & Evaluation:

Limitations: Single-city, single-week dataset; POI data requires base station association.

2.7 Anticipatory Mobile Computing (Survey)¶

• Authors: Veljko Pejovic, Mirco Musolesi
• Year: 2015
• Venue: ACM Computing Surveys, Vol. 47, No. 3

Core approach: Taxonomizes predictable phenomena (mobility, activity, health, social dynamics) and ML techniques (HMMs, Bayesian Networks, Markov processes). Defines the anticipatory pipeline: sense → infer context → predict → act proactively.

Key challenges identified: Energy efficiency, privacy/anonymity, non-deterministic behavior modeling, notification timing, distributed coordination, context drift adaptation.

3. Privacy-Preserving AI for User Behavior¶

3.1 SeqMF: Federated Privacy-Preserving Collaborative Filtering for Next App Prediction¶

• Authors: Albert Sayapin et al.
• Year: 2023 (published in User Modeling and User-Adapted Interaction, 2024)
• Paper: arXiv:2303.04744

Core approach: Sequential matrix factorization adapted for federated learning — raw usage data never leaves the device. Custom privacy mechanism protects gradient transmissions.

Benchmarks & Evaluation:

Limitations: Not superior in static settings; non-standard privacy mechanism complicates formal comparison.

3.2 Federated Learning for Mobile App Privacy Preferences¶

• Authors: Andre Brandao, Ricardo Mendes, Joao P. Vilela
• Year: 2022
• Venue: ACM CODASPY

Core approach: Privacy profiles via privacy-preserving clustering, then federated neural networks predict permission decisions (allow/deny).

Benchmarks & Evaluation:

Limitations: Narrow scope (permission preferences only); communication overhead not quantified.

3.3 Apple On-Device Intelligence (Siri Suggestions / Apple Intelligence)¶

• Year: 2016–present (Private Cloud Compute announced 2024)
• Type: Industry system (proprietary)

Core approach: On-device neural networks learn from local signals (Safari, emails, messages, contacts) for app prediction and contextual shortcuts. Neural Engine inference at <1W. When cloud LLM inference is needed, Apple routes through Private Cloud Compute with cryptographic guarantees — data is not stored or accessible to Apple. Federated learning for speaker recognition; differential privacy noise injected into local training.

Benchmarks: Internal A/B testing across hundreds of millions of devices. No public precision/recall numbers or epsilon values.

Limitations: Entirely proprietary; no reproducibility; tightly coupled to Apple hardware (Neural Engine, Secure Enclave).

4. Cloud-Edge Hybrid OS & Proactive Systems¶

4.1 Android App Standby Buckets & Adaptive Battery¶

• Year: 2018–present (Android 9+; Restricted bucket in Android 12)
• Type: Industry system (Google / DeepMind)

Core approach: ML-based classification of apps into five priority buckets (Active, Working Set, Frequent, Rare, Restricted) using an on-device TensorFlow Lite model predicting future usage probability. Based on bucket assignment, the OS throttles jobs, alarms, network access, and FCM message priority.

Benchmarks & Evaluation:

Limitations: OEM fragmentation (Samsung, Xiaomi customize behavior); no formal DP; aggressive bucketing can break background-dependent apps; no public benchmark.

4.2 RAMOS: Reference Architecture for Cloud-Edge Meta-Operating Systems¶

• Authors: Panagiotis Trakadas et al.
• Year: 2022
• Venue: Sensors (PMC 9692311)

Core approach: Peer-to-peer meta-OS transforming hierarchical cloud-edge-IoT into a dynamic distributed continuum. Data-centric orchestration with federated and swarm learning across heterogeneous devices.

Benchmarks: No empirical evaluation. Validated through five theoretical use-case scenarios only.

Limitations: No prototype, no implementation, no measured performance data.

4.3 Edge-Cloud Collaborative Computing Survey¶

• Authors: Jing Liu et al.
• Year: 2025
• Paper: arXiv:2505.01821

Core approach: Comprehensive survey across five layers: architectures, model optimization (compression, pruning, distillation), resource management, privacy/security, and deployments.

Evaluation dimensions catalogued: Latency, throughput, bandwidth efficiency, inference speed, model size reduction, accuracy preservation, power consumption, data protection effectiveness.

Open challenges: Accuracy-compression trade-offs; heterogeneous device management; federated learning with non-IID data; privacy mechanism overhead.

5. Comparative Analysis¶

5.1 System Feature Comparison¶

5.2 Benchmark & Evaluation Method Comparison¶

6. Benchmark Landscape Summary¶

6.1 Common Evaluation Metrics by Domain¶

LLM OS / Agent Systems:

• Success Rate (SR%) on standardized benchmarks (GAIA, SWE-bench, HumanEval)
• Throughput (system calls/sec, tasks/min)
• Latency (agent wait time, end-to-end task time)
• Token consumption (cost efficiency)

App / Intent Prediction:

• HR@K (Hit Rate at K) — the dominant metric
• MRR@K (Mean Reciprocal Rank)
• NDCG@K (Normalized Discounted Cumulative Gain)
• Precision and Recall (older works)
• F1-score (macro/micro)

Proactive Systems:

• Latency savings (seconds reduced per interaction)
• Energy overhead (% battery per day)
• Content staleness (minutes)
• User satisfaction scores (field studies)

Privacy:

• Privacy-utility trade-off curves
• Differential privacy epsilon (ε) values (rarely reported)
• Accuracy loss from privacy mechanisms (%)
• F1 under obfuscation

6.2 Standard Datasets¶

7. Differentiation of DiPECS¶

Based on this survey, DiPECS occupies a unique position at the intersection of several research threads. Key differentiators:

1. Deterministic State Machine with Replay: No existing LLM-OS or intent prediction system enforces fully observable, replayable state transitions. Golden trace validation (data/traces/) is architecturally unique. The closest analog is formal verification in traditional OS research, not in any AI-OS system.

2. Privacy Air-Gap as Architecture, Not Add-On: Apple's Private Cloud Compute is the closest analog, but it operates as infrastructure (hardware enclaves + cryptographic routing), not as an explicit software architectural layer. DiPECS's aios-core privacy scrubbing before any cloud transmission is a distinct pattern — comparable to a data diode in classified systems.

3. Mechanism-Policy Separation: While cloud-edge computing papers discuss workload partitioning, none frame it as an explicit mechanism (local, deterministic) vs. policy (cloud, stochastic) separation. This maps to the Lampson/Levin principle from classic OS theory, applied to LLM-era systems.

4. Android-First with Cross-Compile: All LLM-OS works target desktop or platform-agnostic backends. App prediction works evaluate on Android traces but do not build OS-level infrastructure. DiPECS targets Android API 33 with NDK cross-compilation as a first-class concern.

5. LLM for Prediction, Not Interaction: Existing LLM-OS systems use LLMs to interpret and execute user commands (task automation). DiPECS uses the cloud LLM for predictive intent inference — acting before explicit user request — which aligns more with the anticipatory computing tradition (FALCON, DeepAPP) but using modern LLM capabilities.

Suggested Benchmark Strategy for DiPECS¶

Based on the evaluation landscape above, a comprehensive DiPECS evaluation should cover:

8. References¶

1. Mei, K. et al. "AIOS: LLM Agent Operating System." arXiv:2403.16971, 2024.
2. Wu, Z. et al. "OS-Copilot: Towards Generalist Computer Agents with Self-Improvement." arXiv:2402.07456, 2024.
3. Zhang, C. et al. "UFO: A UI-Focused Agent for Windows OS Interaction." arXiv:2402.07939, 2024.
4. Zhang, C. et al. "UFO2: The Desktop AgentOS." arXiv:2504.14603, 2025.
5. Yang, J. et al. "SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering." NeurIPS, 2024.
6. Packer, C. et al. "MemGPT: Towards LLMs as Operating Systems." arXiv:2310.08560, 2023.
7. Ge, Y. et al. "LLM as OS, Agents as Apps." arXiv:2312.03815, 2023.
8. Yan, T. et al. "Fast App Launching for Mobile Devices Using Predictive User Context." ACM MobiSys, 2012.
9. Zhao, S. et al. "AppUsage2Vec: Modeling Smartphone App Usage for Prediction." IEEE ICDE, 2019.
10. Shen, Z. et al. "DeepAPP: A Deep Reinforcement Learning Framework for Mobile Application Usage Prediction." ACM SenSys, 2019.
11. Li, L. et al. "Atten-Transformer: A Deep Learning Framework for User App Usage Prediction." arXiv:2502.16957, 2025.
12. Chen, J. et al. "Next App Prediction Based on Graph Neural Networks and Self-Attention Enhancement." Scientific Reports, 2025.
13. Sun, C. et al. "Appformer: Mobile App Usage Prediction via Progressive Multi-Modal Data Fusion." arXiv:2407.19414, 2024.
14. Pejovic, V. & Musolesi, M. "Anticipatory Mobile Computing: A Survey." ACM Computing Surveys 47(3), 2015.
15. Sayapin, A. et al. "SeqMF: Federated Privacy-Preserving Collaborative Filtering for On-Device Next App Prediction." arXiv:2303.04744, 2023.
16. Brandao, A. et al. "Prediction of Mobile App Privacy Preferences with User Profiles via Federated Learning." ACM CODASPY, 2022.
17. Trakadas, P. et al. "RAMOS: A Reference Architecture for Cloud-Edge Meta-Operating Systems." Sensors, 2022.
18. Liu, J. et al. "Edge-Cloud Collaborative Computing on Distributed Intelligence and Model Optimization: A Survey." arXiv:2505.01821, 2025.