{"generatedAt":"2026-10-29","newPapersCount":9,"intersectionCount":0,"topPapers":[{"title":"Software Aging in LLM-Generated Applications: Runtime Evidence, Static Analysis, and Human-Written Comparisons","authors":["Cesar Santos","Michele Vitagliano","Roberto Natella","Ermeson Andrade"],"year":2026,"url":"http://arxiv.org/abs/2608.26391v2","venue":"arXiv","abstractSnippet":"Large Language Models (LLMs) are increasingly used to generate executable software systems from natural language specifications, accelerating development and reducing manual implementation effort. Although recent studies have investigated the functional correctness, security, maintainability, and robustness of LLM-gene","source":"arXiv","axes":{"horizon":"long","aging":true,"energy":false,"outOfBand":"unverified"},"mappedAgenda":["A1","A2"],"citeKey":"santos2026software","discoveredAt":"2026-10-29"},{"title":"Characterizing Software Aging in GPU-Based LLM Serving Systems","authors":["Domenico Cotroneo","Bojan Cukic"],"year":2026,"url":"http://arxiv.org/abs/2606.11916v1","venue":"arXiv","abstractSnippet":"This paper proposes an empirical methodology to study software aging in GPU-based LLM serving systems. Traditional aging studies focus on CPU-centric software with relatively regular workloads; LLM serving is different, spanning a Python host and a CUDA device, handling requests whose cost varies by orders of magnitude","source":"arXiv","axes":{"horizon":"long","aging":true,"energy":false,"outOfBand":"unverified"},"mappedAgenda":["A1","A2","A3","A4"],"citeKey":"cotroneo2026characterizing","discoveredAt":"2026-10-29"},{"title":"Diagnosing and Mitigating Long-Session Degradation in Full-Duplex Speech Models","authors":["Junho Kim","CheHyun Lee","Donghun Kang","Youngho Cho","Dongwook Lee","Heeseung Kim","Sungroh Yoon"],"year":2026,"url":"http://arxiv.org/abs/2610.34991v2","venue":"arXiv","abstractSnippet":"End-to-end full-duplex spoken dialogue models enable natural, real-time spoken interactions with interruptions, backchannels, and overlapping speech. Prior evaluations have assessed diverse interaction capabilities and identified some degradation patterns across conversational turns. However, how model behavior evolves","source":"arXiv","axes":{"horizon":"unknown","aging":true,"energy":false,"outOfBand":"unverified"},"mappedAgenda":["A1","A2"],"citeKey":"kim2026diagnosing","discoveredAt":"2026-10-29"},{"title":"Investigating Software Aging in LLM-Generated Software Systems","authors":["César Santos","Ermeson Andrade","Roberto Natella"],"year":2025,"url":"http://arxiv.org/abs/2510.24188v1","venue":"arXiv","abstractSnippet":"Automatically generated software, especially code produced by Large Language Models (LLMs), is increasingly adopted to accelerate development and reduce manual effort. However, little is known about the long-term reliability of such systems under sustained execution. In this paper, we experimentally investigate the phe","source":"arXiv","axes":{"horizon":"long","aging":true,"energy":false,"outOfBand":"unverified"},"mappedAgenda":["A1","A2"],"citeKey":"santos2025investigating","discoveredAt":"2026-10-29"},{"title":"Understanding Container-based Services under Software Aging: Dependability and Performance Views","authors":["Jing Bai","Xiaolin Chang","Fumio Machida","Kishor S. Trivedi"],"year":2023,"url":"http://arxiv.org/abs/2308.12784v1","venue":"arXiv","abstractSnippet":"Container technology, as the key enabler behind microservice architectures, is widely applied in Cloud and Edge Computing. A long and continuous running of operating system (OS) host-ing container-based services can encounter software aging that leads to performance deterioration and even causes system fail-ures. OS re","source":"arXiv","axes":{"horizon":"long","aging":true,"energy":false,"outOfBand":"unverified"},"mappedAgenda":["A1","A2","A3"],"citeKey":"bai2023understanding","discoveredAt":"2026-10-29"},{"title":"BellowsKV: One Latent for Many Layers, Expanded on Demand","authors":["Jiamu Zhang","Hanjie Chen"],"year":2026,"url":"http://arxiv.org/abs/2610.35023v1","venue":"arXiv","abstractSnippet":"Large language models (LLMs) now serve as the backbone of agents that reason over many turns, call tools, and pursue long-horizon tasks. To avoid recomputation, a model stores the attention keys and values of all past tokens in a key-value (KV) cache. Each layer maintains its own cache, so KV-cache memory can become th","source":"arXiv","axes":{"horizon":"unknown","aging":false,"energy":false,"outOfBand":"unverified"},"mappedAgenda":["A2"],"citeKey":"zhang2026bellowskv","discoveredAt":"2026-10-29"},{"title":"RIFT: KV Cache Compression for Draft Models with Complementary Compressibility Training","authors":["Dor Tsur","Maor Ashkenazi","Maximilian Golub","Peter Hicks","Yaniv Galron","Talor Abramovich","Yoav Miron","Tiyasa Mitra","Bita Darvish Rouhani","Ran Zilberstein"],"year":2026,"url":"http://arxiv.org/abs/2610.31098v1","venue":"arXiv","abstractSnippet":"Speculative decoding is a popular decode acceleration method. As such, it is particularly relevant to long-output generation, which requires an increasingly large KV cache. KV cache compression offers a way to constrain costs in long-context drafting. However, compression methods that focus on prefill information can l","source":"arXiv","axes":{"horizon":"short","aging":false,"energy":false,"outOfBand":"unverified"},"mappedAgenda":["A2"],"citeKey":"tsur2026rift","discoveredAt":"2026-10-29"},{"title":"MoEZip: Routing-Aware KV Cache Compression for Sparse Mixture-of-Experts LLMs","authors":["Minsang Kim","Seungjun Baek"],"year":2026,"url":"http://arxiv.org/abs/2610.30206v1","venue":"arXiv","abstractSnippet":"Mixture-of-Experts (MoE) language models scale model capacity efficiently by activating a small subset of parameters per token. However, in long-context inference, memory remains a major bottleneck because the KV cache grows linearly with sequence length. Existing KV cache compression methods reduce this cost for dense","source":"arXiv","axes":{"horizon":"short","aging":false,"energy":false,"outOfBand":"unverified"},"mappedAgenda":["A2"],"citeKey":"kim2026moezip","discoveredAt":"2026-10-29"},{"title":"TideKV: Delayed KV Eviction via Layer-Wise Chunked Prefill for Memory-Efficient Long-Context LLM Inference","authors":["Yu Wang","Zhu Wang","Ning Hu","Yifan Guo","Xiu Ma","Bin Guo","Zhiwen Yu"],"year":2026,"url":"http://arxiv.org/abs/2610.24499v1","venue":"arXiv","abstractSnippet":"Long-context large language model (LLM) inference is important for modern intelligent applications but remains constrained by peak memory during prompt prefill. Existing KV cache compression reduces decoding memory but typically requires full-prompt prefill to preserve global context, retaining a high construction peak","source":"arXiv","axes":{"horizon":"unknown","aging":false,"energy":false,"outOfBand":"unverified"},"mappedAgenda":["A2"],"citeKey":"wang2026tidekv","discoveredAt":"2026-10-29"}],"researchQuestions":[{"id":"RQ1-SEPARABILITY","pillar":"A1","targetPaper":"paper2-aging-position / paper3-mechanism","question":"How can progressive software aging in long-running LLM inference runtimes (e.g., Ollama / llama.cpp) be statistically isolated from thermal throttling and DVFS governor drift under variable arrival workloads?","hypothesis":"Under fixed hardware frequency caps (e.g., 2.5 GHz) that eliminate turbo thermal saturation, any residual monotonic drift in per-token latency or RSS memory over >24h is attributable to runtime/OS state accumulation rather than thermal decay.","experimentalDesign":"Compare 48h uncapped vs 2.5 GHz capped runs under Poisson arrival vs closed-loop saturation while logging RSS, minor/major page faults, slab/page-cache size, and CPU residency."},{"id":"RQ2-LAYER-LOCALIZATION","pillar":"A2","targetPaper":"paper3-mechanism","question":"Which intervention (container/process restart, KV-cache flush, or OS page-cache drop via /proc/sys/vm/drop_caches) restores throughput and Joules-per-token efficiency without requiring a full node reboot?","hypothesis":"In MoE and dense GGUF models on CPU-only edge nodes, memory allocator fragmentation and OS page-cache eviction pressure account for the majority of non-thermal throughput drift, recoverable via lightweight process restart without OS reboot.","experimentalDesign":"Run a 3-treatment intervention experiment at t=24h of a 48h campaign: (T1) drop_caches=3 only, (T2) Ollama service restart (model reload from page cache), (T3) full container restart with cold disk reload."},{"id":"RQ3-JOULE-REJUVENATION","pillar":"A3","targetPaper":"paper4-framework","question":"When is the energy and carbon cost of a rejuvenation action (e.g., reloading a 4.7 GB–8.9 GB GGUF model into RAM) amortized by the post-rejuvenation reduction in Joules per token?","hypothesis":"Energy-aware rejuvenation is only net-positive in Joules when the cumulative throughput degradation slope exceeds E_rejuv / (P_active * T_horizon); under high-carbon grid hours, deferring rejuvenation to low-carbon windows minimizes gCO2eq.","experimentalDesign":"Measure exact wall-plug energy E_rejuv (in Joules) of cold and warm model reloads on the TP-Link P115 meter and formulate the break-even renewal reward equation in Joules and gCO2eq."},{"id":"RQ4-VERIFIABLE-PROTOCOL","pillar":"A4","targetPaper":"paper2-aging-position / paper4-framework","question":"What minimal continuous out-of-band telemetry invariants are required to certify that a multi-day edge AI experiment actually maintained its declared hardware and workload state?","hypothesis":"Boundary-only configuration checks fail silently in >15% of multi-day edge campaigns (as demonstrated by the two uncapped 2.5 GHz runs and the 40.1h power telemetry drop in Paper 1); continuous in-band + out-of-band cross-auditing detects 100% of silent cap violations and sensor dropouts.","experimentalDesign":"Formalize the 4-invariant verification protocol from llm-energy-tests (frequency residency adherence, wall-plug sample continuity, prompt-cycle completeness, and thermal steady-state)."},{"id":"RQ5-WORKLOAD-NONSTATIONARITY","pillar":"A1","targetPaper":"paper3-mechanism","question":"Does bursty/diurnal prompt arrival accelerate KV-cache allocator fragmentation compared to closed-loop saturation at equal total token volume?","hypothesis":"Variable prompt lengths and bursty inter-arrival times trigger frequent KV-cache reallocations, increasing heap fragmentation compared to uniform saturation.","experimentalDesign":"Replay trace-driven diurnal arrival patterns vs steady saturation at 2.5 GHz cap."},{"id":"RQ6-MOE-EXPERT-HOTSPOTS","pillar":"A2","targetPaper":"paper3-mechanism","question":"Do Mixture-of-Experts (MoE) models (e.g., DeepSeek-v2 Lite) exhibit different page-cache and RSS aging trajectories than dense models due to sparse expert routing?","hypothesis":"Sparse activation across experts creates non-uniform page access patterns that interact differently with kernel page-cache eviction over multi-day runs.","experimentalDesign":"Compare per-hour RSS, minor page faults, and LLC miss rates between DeepSeek-v2 Lite (MoE) and Llama 3.1 8B (Dense) over 48 hours."},{"id":"RQ7-CARBON-AWARE-REJUVENATION-WINDOW","pillar":"A3","targetPaper":"paper4-framework","question":"Can rejuvenation actions (process restart / cold model reload) be co-scheduled with low-carbon grid windows (e.g., solar midday in PT/ES) to reduce net rejuvenation carbon cost by >40%?","hypothesis":"Aligning proactive rejuvenation with diurnal grid carbon troughs minimizes the gCO2eq penalty of model reload spikes while preventing late-stage aging degradation.","experimentalDesign":"Simulate and evaluate carbon-aware rejuvenation policies across the 5 European electricity grids (NO, FR, PT, ES, PL) from Paper 1."}],"markdown":"# Relatório Semanal de Survey: Software Aging & Rejuvenation em Edge LLMs\n**Data:** 2026-10-29 | **Doutorando:** Magnus Alencar da Cruz (CISUC/DEI - UC)\n\n## 1. Resumo Executivo da Semana\n- **Artigos novos detetados:** 9\n- **Artigos na interseção direta (Aging + Energy):** 0\n- **Research Questions (RQs) ativas no banco de tese:** 7\n\n## 2. Top Artigos Recentes Classificados nos Eixos da Tese\n| CiteKey Sugerida | Título | Ano / Fonte | Horizon | Aging | Energy | Pilares (A1–A4) |\n|---|---|---|---|---|---|---|\n| `santos2026software` | Software Aging in LLM-Generated Applications: Runtime Evidence, Static Analysis, and Human-Written Comparisons (http://arxiv.org/abs/2608.26391v2) | 2026 (arXiv) | long | ✓ | --- | A1, A2 |\n| `cotroneo2026characterizing` | Characterizing Software Aging in GPU-Based LLM Serving Systems (http://arxiv.org/abs/2606.11916v1) | 2026 (arXiv) | long | ✓ | --- | A1, A2, A3, A4 |\n| `kim2026diagnosing` | Diagnosing and Mitigating Long-Session Degradation in Full-Duplex Speech Models (http://arxiv.org/abs/2610.34991v2) | 2026 (arXiv) | unknown | ✓ | --- | A1, A2 |\n| `santos2025investigating` | Investigating Software Aging in LLM-Generated Software Systems (http://arxiv.org/abs/2510.24188v1) | 2025 (arXiv) | long | ✓ | --- | A1, A2 |\n| `bai2023understanding` | Understanding Container-based Services under Software Aging: Dependability and Performance Views (http://arxiv.org/abs/2308.12784v1) | 2023 (arXiv) | long | ✓ | --- | A1, A2, A3 |\n| `zhang2026bellowskv` | BellowsKV: One Latent for Many Layers, Expanded on Demand (http://arxiv.org/abs/2610.35023v1) | 2026 (arXiv) | unknown | --- | --- | A2 |\n| `tsur2026rift` | RIFT: KV Cache Compression for Draft Models with Complementary Compressibility Training (http://arxiv.org/abs/2610.31098v1) | 2026 (arXiv) | short | --- | --- | A2 |\n| `kim2026moezip` | MoEZip: Routing-Aware KV Cache Compression for Sparse Mixture-of-Experts LLMs (http://arxiv.org/abs/2610.30206v1) | 2026 (arXiv) | short | --- | --- | A2 |\n| `wang2026tidekv` | TideKV: Delayed KV Eviction via Layer-Wise Chunked Prefill for Memory-Efficient Long-Context LLM Inference (http://arxiv.org/abs/2610.24499v1) | 2026 (arXiv) | unknown | --- | --- | A2 |\n\n## 3. Alimentação de Research Questions (RQs) para os Papers 2, 3 e 4\n- **RQ1-SEPARABILITY [A1 — paper2-aging-position / paper3-mechanism]**: How can progressive software aging in long-running LLM inference runtimes (e.g., Ollama / llama.cpp) be statistically isolated from thermal throttling and DVFS governor drift under variable arrival workloads?\n  - *Hipótese*: Under fixed hardware frequency caps (e.g., 2.5 GHz) that eliminate turbo thermal saturation, any residual monotonic drift in per-token latency or RSS memory over >24h is attributable to runtime/OS state accumulation rather than thermal decay.\n- **RQ2-LAYER-LOCALIZATION [A2 — paper3-mechanism]**: Which intervention (container/process restart, KV-cache flush, or OS page-cache drop via /proc/sys/vm/drop_caches) restores throughput and Joules-per-token efficiency without requiring a full node reboot?\n  - *Hipótese*: In MoE and dense GGUF models on CPU-only edge nodes, memory allocator fragmentation and OS page-cache eviction pressure account for the majority of non-thermal throughput drift, recoverable via lightweight process restart without OS reboot.\n- **RQ3-JOULE-REJUVENATION [A3 — paper4-framework]**: When is the energy and carbon cost of a rejuvenation action (e.g., reloading a 4.7 GB–8.9 GB GGUF model into RAM) amortized by the post-rejuvenation reduction in Joules per token?\n  - *Hipótese*: Energy-aware rejuvenation is only net-positive in Joules when the cumulative throughput degradation slope exceeds E_rejuv / (P_active * T_horizon); under high-carbon grid hours, deferring rejuvenation to low-carbon windows minimizes gCO2eq.\n- **RQ4-VERIFIABLE-PROTOCOL [A4 — paper2-aging-position / paper4-framework]**: What minimal continuous out-of-band telemetry invariants are required to certify that a multi-day edge AI experiment actually maintained its declared hardware and workload state?\n  - *Hipótese*: Boundary-only configuration checks fail silently in >15% of multi-day edge campaigns (as demonstrated by the two uncapped 2.5 GHz runs and the 40.1h power telemetry drop in Paper 1); continuous in-band + out-of-band cross-auditing detects 100% of silent cap violations and sensor dropouts.\n- **RQ5-WORKLOAD-NONSTATIONARITY [A1 — paper3-mechanism]**: Does bursty/diurnal prompt arrival accelerate KV-cache allocator fragmentation compared to closed-loop saturation at equal total token volume?\n  - *Hipótese*: Variable prompt lengths and bursty inter-arrival times trigger frequent KV-cache reallocations, increasing heap fragmentation compared to uniform saturation.\n- **RQ6-MOE-EXPERT-HOTSPOTS [A2 — paper3-mechanism]**: Do Mixture-of-Experts (MoE) models (e.g., DeepSeek-v2 Lite) exhibit different page-cache and RSS aging trajectories than dense models due to sparse expert routing?\n  - *Hipótese*: Sparse activation across experts creates non-uniform page access patterns that interact differently with kernel page-cache eviction over multi-day runs.\n- **RQ7-CARBON-AWARE-REJUVENATION-WINDOW [A3 — paper4-framework]**: Can rejuvenation actions (process restart / cold model reload) be co-scheduled with low-carbon grid windows (e.g., solar midday in PT/ES) to reduce net rejuvenation carbon cost by >40%?\n  - *Hipótese*: Aligning proactive rejuvenation with diurnal grid carbon troughs minimizes the gCO2eq penalty of model reload spikes while preventing late-stage aging degradation."}