arXiv subjects

Anubhab Banerjee

Publications and source records attributed to Anubhab Banerjee.

7 recordsLinked to original sources

arXiv2026

Anatomy of a Quantized Agent: VRAM Stability and Forecasting in Code-Synthesis Agentic Workloads

Analytical models of peak VRAM consumption for LLM inference decompose memory into weight-storage, KV-cache, and activation terms parameterized by step count, tool invocations, and context expansion. We evaluate this decomposition empirically within a strictly scoped measurement study: a LangGraph-based CUDA-kernel-synthesis agent (AgentK), a 4-bit quantization family (Q4 K M), a single NVIDIA H100 GPU, and four LLM backbones across 1,920 trajectories. Focusing on peak-memory forecasting behavior, we report two primary observations. First, closed-form analytical models achieve competitive accuracy when provided with two empirical constants: loaded-weight VRAM and a fixed activation-memory overhead. Supplied with live GPU readings and ground-truth trajectory parameters, the closed-form model matches or outperforms the best learned baseline on three of the four backbones (test MAPE 2.2-4.4% vs. 3.4-6.5%, p = 0.76). The exception is the smallest backbone (Phi-4-mini), where minimal VRAM variance (CV 0.3%) causes dynamic modeling to underperform simple regression. Second, compile success strictly bifurcates by backbone capacity (from 5.7% for Phi-4-mini to 62.0% for Qwen2.5-Coder-14B), demonstrating that functional code synthesis remains constrained by intrinsic LLM capabilities rather than available memory. Furthermore, because overall peak-memory variance is remarkably low across all backbones (CV 0.3-9.4%), learned prompt-feature regression offers statistically insignificant improvements over a constant-mean baseline. Consequently, we find no justification for deploying complex predictive VRAM models in highly quantized, weight-dominated regimes. We release the evaluated corpus and anonymized framework to support replication.

cs.AI↗

arXiv2026

When Words Predict Workload

Standard distributed \ac{llm} schedulers rely on static token counts or rolling latency averages, making them susceptible to failures on statutorily constrained text. On \ac{epo} claims governed by Article 84 \ac{epc}, linguistic rigidity makes human and machine authorship statistically indistinguishable. Resolving this ambiguity mid-flight forces dynamic multi-model ensemble expansion, triggering unpredictable KV-cache and weight-allocation spikes that saturate consumer-grade edge GPU VRAM and cause severe \ac{oom} crashes. To prevent hardware collapse, we propose a CPU-side Linguistic Resource Forecasting (LRF) gateway. The gateway extracts a 16-dimensional text-structure vector and applies an XGBoost predictor to forecast trap-band membership. The resulting escalation probability ($\Pesc$) is evaluated against a dynamic, closed-form routing threshold ($\Tauroute(t)$) computed via real-time latency telemetry. Requests are safely routed to either a local Qwen2.5-7B edge worker or a remote contrastive ensemble (Qwen2.5 7B + 32B) on an NVIDIA H100 \emph{before} any edge GPU memory is allocated. In a 6,000-request live trial, the LRF gateway reduced the operational misroute fraction ($R_{\mathrm{mis}}$) to $0.087$--$0.095$, an order of magnitude below the token-count baseline ($0.849$). Peak edge VRAM remained safely bounded at $\SI{4.82}{\gibi\byte}$ (under the $\SI{8}{\gibi\byte}$ ceiling) across a $27\times$ variation in \ac{wan} delay. The predictor achieved a live-trial AUROC of $0.84$, and the dynamic $\Tauroute(t)$ controller yielded an $8.2\%$ relative reduction in misroutes compared to an equivalent static threshold.

cs.DC↗

arXiv2026

The Perplexity Trap: When Patent Law Makes Human Writing Look Like AI

The European Patent Office (EPO) reported record filings in 2025, and the 2026 EPO Guidelines hold applicants strictly responsible for LLM-assisted content under Article 83 and Rule 42, creating pressure to triage suspected AI-generated patent text. Two constraints make this hard. First, realistic prosecution settings often have only consumer GPUs with about 8 GB VRAM, not datacenter-class scoring stacks. Second, Article 84 of the European Patent Convention requires claims to be clear and concise, pushing human drafting onto the same low-perplexity, low-burstiness manifold that LLMs occupy. We benchmark three open-source zero-shot detectors on 500 granted EPO H04 telecom patents versus 500 LLM-generated counterparts using five prompting strategies, all under the consumer hardware envelope. At claim level, all detectors exceed 60 percent false-positive rate: Binoculars 78.3 percent, Fast-DetectGPT 61.3 percent, DetectGPT 80.5 percent. The failure persists under Qwen2.5-3B-Instruct regeneration, LoRA-adapted Pythia-2.8B scoring heads, cross-IPC replication on A61K, C07D, and F03D (mean FPR 84.6 percent), and H100 re-evaluation with published Falcon-7B and GPT-J-6B heads, arguing the issue is structural rather than substitute-model capacity. A seven-feature linguistic-complexity logistic regression reaches 74.0 percent accuracy at 28.1 percent FPR, a 13 percentage-point gain over a perplexity-only baseline at a comparable operating point, without using likelihood at inference and within the same hardware budget.

cs.CL↗

arXiv2026

Inductive Latent Context Persistence: Closing the Post-Handover Cold Start in 6G Radio Access Networks

In modern radio access networks (RANs), rule-based handover (HO) decisions (e.g., A3/A5) depend on user equipment (UE) measurements only, so UEs at the same location can receive inconsistent HO outcomes. GNN-based methods improve HO KPIs using richer context than measurements alone. However, recurrent or graph models discard the per-UE recurrent state at HO and reinitialize at the target next-generation Node B (gNB), losing mobility history and forcing the target model to rebuild from post-HO measurements only. We address this post-HO cold start with Inductive Latent Context Persistence (ILCP), compressing the source recurrent state, transporting it on the 3GPP Xn as a 128-byte payload, and adapting it at the target gNB. We model the RAN as a dynamic heterogeneous graph over UE nodes, gNB nodes, measurement edges, and Xn edges. On a Vienna 4G/5G drive-test, ILCP achieves 0.0% ping-pong HOs versus 6.5% for an identical no-transfer baseline and 22.6% for a Transformer baseline; post-HO accuracy improves by +5.1 pp on average (peak +13.3 pp) in the 50-250 ms window. On one NVIDIA GTX 1080 (8 GB), ILCP runs end-to-end at 7.7 ms p99 per handover decision. Under perturbations (shadow fading, NLOS blockage, SSB-burst sparsity), robustly trained ILCP keeps handover failure (HOF) in the 10-13% range. Under the same fixed-reference-label setting, A3/A5 rises from 1.1% to 57-65% HOF when measurements are perturbed, exposing limits of measurement-only rules.

cs.NI↗

arXiv2020

RAN Cognitive Controller

Cognitive Autonomous Networks (CAN) deploys learning based Cognitive Functions (CF) instead of conventional rule-based SON Functions (SF) as Network Automation Functions (NAF) to increase the system autonomy. These CFs work in parallel sharing the same resources which give rise to conflicts among them which cannot be resolved using conventional rule based approach. Our main target is to design a Controller which can resolve any type of conflicts among the CFs in a dynamic way.

cs.NI↗

arXiv2020

On the Necessity and Design of Coordination Mechanism for Cognitive Autonomous Networks

Cognitive Autonomous Networks (CAN) are promoted to advance Self Organizing Network (SON), replacing rule-based SON Functions (SFs) with Cognitive Functions (CFs), which learn optimal behavior by interacting with the network. As in SON, CFs do encounter conflicts due to overlap in parameters or objectives. However, owing to the non-deterministic behavior of CFs, these conflicts cannot be resolved using rulebased methods and new solutions are required. This paper investigates the CF deployments with and without a coordination mechanism, and proves both heuristically and mathematically that a coordination mechanism is required. Using a two-CF Multi-Agent-System model with the possible types of conflicts, we show that the challenge is a typical bargaining problem, for which the optimal response is the Nash bargaining Solution (NBS). We use NBS to propose a coordination mechanism design that is capable of resolving the conflicts and show via simulations how implementation of the proposed solution is feasible in real life scenario.

cs.NI↗

arXiv2019

Game theoretic pricing policy in online media delivery platforms

Residential users get most of their preferred content (e.g., films, news) from their wireless access point, which is connected to content providers through one or several network providers. Nowadays, these access points have enough storage and wireless transmission capacities to distribute the content to neighboring access points if required and allowed by the user. Based on this concept, this paper considers a hybrid content distribution model in peer-to-peer content deliver networks. A user or peer in this network can get its content either directly from the content provider (CP), or from a neighboring user who has this content stored and is willing to share it. Two types of users are considered: premium users are the users getting always the content directly from the CP, and standard user are the users getting the content from other standard users, when possible, otherwise from the CP. Furthermore, standard user share their content to other users if required. This paper studies the pricing problem of the CP such that there are enough standard users to distribute the content (using less network resources and reducing significantly the delay) while maximizing the CP benefits.We address the problem from a game theoretic approach and evaluate when the Nash Equilibrium can be attained. The obtained case have been applied to a real life study using a BBC iPlayer dataset and an analysis on the CP benefits dependence on different parameters. Finally we extend this pricing model when considering also the reduction of load on its content custodian (server) and discuss how the pricing policy is affected.

cs.NI↗