References

All references cited in the keyword matrix and the two discovery reports, organized by theme.

MSR Methodology Bug Taxonomy (Direct Precedents) On-Device Performance Measurement Hardware & SoC Research Memory & KV Cache Energy & Thermal Quantization Deployment & On-Device Pipeline Background & Survey

MSR Methodology

CONF
Islam, M. J.; Nguyen, G.; Pan, R.; Rajan, H. A Comprehensive Study on Deep Learning Bug Characteristics. ESEC/FSE 2019, pp. 510–520.
DOI 10.1145/3338906.3338955

Foundational symptom-seeded vocabulary protocol. Establishes that MSR search terms should be derived from observable symptoms (crash, hang, wrong output), not root-cause names.
CONF
Islam, M. J.; Pan, R.; Nguyen, G.; Rajan, H. Repairing Deep Neural Networks: Fix Patterns and Challenges. ICSE 2020.
DOI 10.1145/3377811.3380378

Fix-pattern taxonomy; motivates the memory_leak and regression terms.
arXiv
Codabux, Z.; Fard, F.; Verdecchia, R.; Palomba, F.; Di Nucci, D.; Recupito, G. Teaching Mining Software Repositories. arXiv:2501.01903, 2025.
arXiv:2501.01903

MSR study protocol including the 20-result pre-test exclusion rule and the importance of adhering to the mined community's native jargon (tok/s, KV cache).
arXiv
Thomas, D.-G.; Biagiola, M.; Humbatova, N.; Wardat, M.; Jahangirova, G.; Rajan, H.; Tonella, P. muPRL: A Mutation Testing Pipeline for Deep Reinforcement Learning based on Real Faults. arXiv:2408.15150, 2024.
arXiv:2408.15150

Two-axis strategy for mining real repositories for a failure taxonomy; motivates the device-axis precision filter.
CONF
Cao, R. et al. A Comprehensive Study on DNN Compilation Bugs. ESEC/FSE 2022. arXiv:2112.01771.
arXiv:2112.01771

Establishes standard RQ structure for DNN bug studies: (1) symptom taxonomy, (2) root-cause analysis, (3) fix-effort characterization.
CONF
Long, T.; Chen, L. An Empirical Study of Deep Learning Frameworks Bugs. EASE 2022. arXiv:2204.07893.
arXiv:2204.07893

Bug Taxonomy in LLM Inference (Direct Precedents)

CONF
Chen, Z.; Yao, H.; Lou, Y.; Cao, Y.; Liu, Y.; Wang, H.; Liu, X. An Empirical Study on Deployment Faults of Deep Learning Based Mobile Applications. ICSE 2021.
PDF

Closest on-device precedent: 304 deployment faults of TF Lite / Core ML apps from GitHub and Stack Overflow, 23-category symptom taxonomy. Method mirrors ours: label-based exclusion of feature/build/docs issues, manual refining before labelling, 50% pilot, κ = 0.819 / 0.743. Memory and speed issues are 25.5% of their inference-stage faults; 20.1% of deployment faults do not crash and surface only through developer experience, which motivates predicting cost. Cited by the criteria validation.
arXiv
Liu, X. et al. An Empirical Study of Bugs in LLM Inference Frameworks. arXiv:2506.09713, 2025.
arXiv:2506.09713

Closest direct precedent: 929 bugs across llama.cpp, MLC-LLM, vLLM, DeepSpeed, TensorRT-LLM. Cohen's κ = 0.927. Two repositories shared with this study. S4 Abnormal Performance symptom category maps directly to RQ1.
arXiv
Cai, X. et al. An Empirical Study on Bug Characteristics of LLM Open-source Projects. arXiv:2409.16559, 2024.
arXiv:2409.16559

994 closed issues from 15 LLM open-source projects including Ollama. 15-category taxonomy with Performance Issue (7.14%) and Memory Issue (3.92%) as separate top-level categories — direct precedent for resource_category field separation.
ICSE
Chen, X. et al. An Empirical Study on Deployment Faults of Deep Learning Based Mobile Applications. ICSE 2021. arXiv:2101.04930.
arXiv:2101.04930

Methodological baseline for on-device mobile inference faults: 304 faults from TF Lite and Core ML, 23-category taxonomy. Baseline from which this study advances by covering the post-2022 generation of on-device LLM runtimes.
CONF
Dantas Cruz, B.; Paul, A. K.; Song, Z.; Tilevich, E. STARGAZER: A Deep Learning Approach for Estimating the Performance of Edge-Based Clustering Applications. IEEE SMDS 2020.
DOI 10.1109/SMDS49396.2020.00009

Stargazer v1: motivates the energy/memory regression targets and the Raspberry Pi / edge hardware cohort.

On-Device LLM Performance Measurement

WS
Imai, S.; Nakazawa, R.; Amaral, M.; Choochotkaew, S.; Chiba, T. Predicting LLM Inference Latency: A Roofline-Driven ML Method. ML for Systems @ NeurIPS 2024.
Roofline model for on-device latency prediction; motivates KV cache, latency, tokens/s, and prefill terms.
arXiv
Tummalapalli, P. et al. LLM Inference at the Edge: Mobile, NPU, and GPU Performance Efficiency Trade-offs Under Sustained Load. arXiv:2603.23640, 2026.
arXiv:2603.23640

Documents iPhone 16 Pro thermal throttling: 40.35 tok/s → 22.56 tok/s within two inference iterations. Key validation for thermal, throttling, throttled terms.
arXiv
MobiBench team. MobiBench: Benchmarking LLM Inference on Mobile Devices. arXiv:2609.13159, 2026.
arXiv:2609.13159

Reports TTFT and TPOT as the two canonical latency decomposition metrics. Gap: neither appears in the current keyword grid.
arXiv
lm-Meter team. lm-Meter: A Profiling Benchmark for On-Device LLM Inference. arXiv:2510.06126, 2025.
arXiv:2510.06126

Profiling paper covering TTFT, TPOT/ITL, prefill latency, decode latency, and KV cache saturation at phase level.
arXiv
Characterizing Mobile SoC for Accelerating Heterogeneous LLM Inference. arXiv:2501.14794, 2026.
arXiv:2501.14794

SoC as primary unit of analysis; identifies 68 GB/s memory bandwidth ceiling on Snapdragon 8 Gen 3. Key paper for SoC, memory bandwidth, and snapdragon terms.
arXiv
Phase-Aware Workload Characterization of LLM Inference (Hydra). arXiv:2608.25053, 2026.
arXiv:2608.25053

Title named around prefill/decode distinction as the dominant structuring principle of on-device performance analysis.
arXiv
A Systematic Evaluation of On-Device LLMs for Diverse Tasks and Devices. arXiv:2505.15030, 2025.
arXiv:2505.15030

Systematic evaluation varying quantization and context length simultaneously; reports peak memory footprint from 392 MB (0.5B) to 7297 MB (7B, Q8).
arXiv
Tu, X. et al. Transferable Latency Prediction for Fast LLM Screening on Heterogeneous Edge Devices. arXiv:2607.21602, 2026.
arXiv:2607.21602

Heterogeneous edge device screening; motivates ARM, Jetson, energy, and INT4 terms.
arXiv
Large Language Model Inference on Edge Devices: A Survey. arXiv:2608.15693, 2026.
arXiv:2608.15693

Survey covering Jetson, Raspberry Pi, ARM, NPU, and on-device deployment vocabulary.

Hardware & SoC Research

arXiv
Cai, G. et al. Is Your NPU Ready for LLMs? Dissecting the Hidden Efficiency Bottlenecks in Mobile LLM Inference. arXiv:2607.05475, 2026.
arXiv:2607.05475

NPU bottleneck analysis; most-cited paper in the keyword matrix for the NPU device term. Validates ios, iphone, ipad, snapdragon.
CONF
When NPUs Are Not Always Faster: Understanding the Bottlenecks of NPU Inference for LLMs. ASPLOS 2025.
arXiv:2605.27435 ACM DL

NPU in conference paper title; validates NPU as a primary research subject at the highest level (ASPLOS).
CONF
Benchmarking LLM Inference on Apple Silicon. ACM DL, 2024.
DOI 10.1145/3771563

Validates "Apple Silicon" as a recognized hardware category in paper titles.
arXiv
Hexagon-Accelerated LLM Inference: From PyTorch to the NPU. arXiv:2509.23324, 2026.
arXiv:2509.23324

Hexagon named as "a representative mobile NPU due to its typical architecture, widespread adoption, and accessible SDK." Also covers model conversion vocabulary.
arXiv
Toward Efficient Inference: LLMs on Apple Neural Engine. arXiv:2603.06728, 2026.
arXiv:2603.06728

ANE "remains largely unused for LLM workloads"; over two billion devices ship with the ANE. Gap term candidate.
arXiv
Apple Neural Engine for LLM Inference: Challenges and Opportunities. arXiv:2604.18788, 2026.
arXiv:2604.18788
arXiv
DVFS-Aware LLM Inference on Mobile SoCs. arXiv:2512.01644, 2025.
arXiv:2512.01644

DVFS (Dynamic Voltage and Frequency Scaling) as the hardware mechanism linking thermal/power events to compute degradation.

Memory & KV Cache

arXiv
PowerInfer-2: Fast Large Language Model Inference on a Smartphone. arXiv:2406.06282, 2024.
arXiv:2406.06282

Reports 11.68 tokens/s for a 47B LLM on a smartphone. Key example of tokens/s as the universal on-device throughput metric.
arXiv
MiniKV: Pushing the Limits of LLM Inference via 2-Bit Layer-Discriminative KV Cache. arXiv:2411.18077, 2024.
arXiv:2411.18077

KV cache as both a symptom label and an active area of optimization research.
arXiv
mzCache: A Memory-Efficient KV Cache for On-Device LLM Inference. arXiv:2609.01338, 2026.
arXiv:2609.01338
arXiv
Less Is More: Efficient On-Device LLM Inference via Session Management. arXiv:2604.24636, 2026.
arXiv:2604.24636

Documents session rotation at fixed intervals (3–5 generations) as a named mitigation for KV cache saturation on-device.
arXiv
FlexInfer: Flexible Layer Offloading for On-Device LLM Inference. arXiv:2503.03777, 2026.
arXiv:2503.03777

Model offloading — swapping model layers between DRAM and Flash storage — as a named edge-specific memory management strategy.
arXiv
Sustainable LLM Inference: Memory Footprint Characterization. arXiv:2504.03360, 2025.
arXiv:2504.03360

RAM as the mobile-community label for physical memory; documents peak RSS as a more precise academic form.

Energy & Thermal

arXiv
Zou, B. et al. EnerInfer: Energy-Aware On-Device LLM Inference. arXiv:2606.23001, 2026.
arXiv:2606.23001

Primary reference for battery, energy, and power terms. Energy per token / Joules per token as the canonical on-device efficiency metric.
arXiv
PELM: Power-Efficient LLM Inference on Mobile Devices. arXiv:2609.09662, 2026.
arXiv:2609.09662

Instrumented-measurement study reporting steady-state Watts during inference.
arXiv
The Battery Price of Edge AI: An Empirical Study. arXiv:2609.11940, 2026.
arXiv:2609.11940

"20 conversations consume up to 6–25% of battery in under 15 minutes" — battery as a user-facing metric.

Quantization

CONF
AWQ: Activation-aware Weight Quantization for LLM Compression and Acceleration. MLSys 2024. Best Paper Award.
MLSys 2024

3–4× performance boost vs. FP16 via TinyChat on edge platforms. Motivates INT4 and AWQ as deployment format terms.
arXiv
GGUF Quantization Quality and Deployment Scenarios. arXiv:2601.14277, 2026.
arXiv:2601.14277

Full GGUF format hierarchy: Q3_K_S through Q8_0. Quality degrades "gradually from Q8_0 to Q4_K_M (0.51% avg loss), then precipitously from Q4_K_M to Q2_K (7.94%)". Validates INT4 as a deployment descriptor.
arXiv
GPTQ: Accurate Post-Training Quantization for Generative Pre-trained Transformers. arXiv:2306.00978, 2023.
arXiv:2306.00978

GPTQ as the standard GPU-serving baseline; less discriminative of on-device edge contexts.
arXiv
Quantization Accuracy Under Progressive Compression: Fidelity Loss and Distributional Drift. arXiv:2609.07664, 2026.
arXiv:2609.07664

Introduces "fidelity loss" and "distributional drift" as terms for accuracy-beyond-metrics degradation under progressive quantization.
arXiv
A Standard for Portable On-Device Deployment: GGUF and Its Ecosystem. arXiv:2604.24785, 2025.
arXiv:2604.24785

GGUF confirmed as "a practical standard for portable deployment across ARM CPUs and NVIDIA GPUs" in academic papers.

Deployment & On-Device Pipeline

arXiv
ExecuTorch: An On-Device AI Framework for PyTorch. arXiv:2605.08195, 2025.
arXiv:2605.08195

Introduces torch.export, Export IR, Edge Dialect, and backend lowering vocabulary. Motivates model conversion as QUERY-4 subject.
arXiv
Cross-Runtime LLM Deployment: PyTorch to ONNX to TFLite Pipeline. arXiv:2511.13453, 2025.
arXiv:2511.13453

PyTorch → ONNX → TFLite as an established "multi-step pipeline for on-device inference".
arXiv
From PyTorch to the NPU: LLM-Agent-Driven Model Conversion Across Heterogeneous Inference Runtimes. arXiv:2509.23324, 2026.
arXiv:2509.23324

Positions "model conversion" as a primary research term covering operator compatibility handling, runtime integration, and precision verification. Justifies QUERY-4.

Background & Survey

arXiv
A Survey of Small Language Models. arXiv:2410.20011, 2024.
arXiv:2410.20011

SLM (Small Language Model) as the class label for edge/mobile-deployable LLMs.
CONF
Landis, J. R.; Koch, G. G. The Measurement of Observer Agreement for Categorical Data. Biometrics 33(1):159–174, 1977.
Cohen's Kappa interpretation scale used for inter-annotator agreement: κ = 0.795 (LLM pilot), threshold 0.61 for human round.
CONF
Amershi, S. et al. Software Engineering for Machine Learning: A Case Study. ICSE-SEIP 2019.
Motivated the deployment candidate (rejected in §5.1).
CONF
Sculley, D. et al. Hidden Technical Debt in Machine Learning Systems. NeurIPS 2015.
Motivated the production candidate (rejected in §5.1).