Searcharxiv⌕ Search

arXiv subjects

Suman Banerjee

Publications and source records attributed to Suman Banerjee.

At least 19 recordsLinked to original sources

Shared Context, Separate Engines: Stateful Streaming Inference for Multimodal LLMs on Edge SoCs

Streaming multimodal large language models (MLLMs) must take in new video frames while answering questions about what they have already seen. On an edge SoC, if one engine does both, ingestion and answering compete for its time while the other engines sit idle. To take over a query, another engine must first import or rebuild the accumulated history, because each runtime keeps it as KV state in its own format. \name lets one engine answer from history that another engine built. Each query reads the history as it stood when the query was asked, so one engine can keep ingesting while another answers. The CPU and GPU can also decode answers to different queries simultaneously. Phase-level state continuation lets an engine prefill a query, pass its KV state to another engine for decoding, and return to ingestion. State-Driven Execution Planning (SDEP) chooses where each query is prefilled and decoded, based on the state each engine holds and the work queued on it. Across M4 Max, RK3588, and QCS6490, \name outperforms single-engine execution. On a live M4 Max stream, SDEP cuts p95 time to first token by $76\%$, and every query receives its first token within one second. Under backlog, it raises request throughput by up to $35\%$. For two concurrent video streams on RK3588, SDEP delivers $19.6\%$ higher request throughput than single-engine execution and $27.2\%$ more than running both engines in parallel without shared state. \name delivers up to $2.48\times$ the throughput of the same engines run serially. Every answer matches serial execution token for token.

cs.LG↗

Koa-decision: Fast and Consistent Structured Decision Making with Generative LLMs

Industry applications often demand low-latency classification, yet current large language model (LLM) approaches remain poorly suited for latency-critical applications. Existing prompting and constrained decoding produce verbose, multi-token outputs that require expensive token-by-token generation, while encoder-based models achieve faster inference but sacrifice task flexibility. We propose Koa-decision, a framework for low-latency atomic actions -- fast, single-step decisions such as classification, semantic endpointing, Boolean checks, and scoring -- formulated as constrained generation with single-token outputs. By introducing atomic label tokens and applying supervised fine-tuning, our method reduces classification to a deterministic one-step decoding problem. Across standard benchmarks, Koa-decision delivers competitive accuracy with consistently low and stable latency. On a production intent-routing benchmark, Koa-decision reaches 85.5% accuracy -- competitive with the strongest frontier models (Claude-4.8-Opus, Gemini-Pro-3.1) and ahead of GPT-5 and Gemini-2.5-Pro -- while answering in about half a second, several-fold faster than every frontier model (up to ~7.5x at the median) under identical serving conditions. Against the dedicated single-token system Jev/TypeSafe, Koa-decision is competitive on accuracy and faster at the median, while also handling multimodal inputs and multi-label outputs that single-label text systems do not.

cs.LG↗

Knowing Is Not Choosing: What Explicit Verification Adds Beyond Generative Preference

Generating a correct answer does not mean that a language model will select it. We separate factual recall into three steps: generating a correct candidate, ranking the available candidates, and selecting the final answer. Pre-generation readouts predict factual recall and which questions sampling will cover across three model families, but say little about whether an available correct answer will ultimately be selected. Explicit verification with $P(\mathrm{True})$ improves within-question ranking over mean log-likelihood in Gemma, Qwen3, and Llama, with AUROC gains of $0.08$--$0.12$. In a prospectively defined Gemma cohort, verification raises plurality accuracy by about $5$ points, and still gains about $2$ points over chat-template likelihood, a stronger generative baseline. The advantage is strongest for relations with common-answer priors and depends on access to the entity; masking the entity removes the ranking advantage in larger Qwen models. Finally, the measured benefit depends on how correctness is defined: recall-oriented reference matching can credit option lists favored by likelihood and substantially understate the improvement seen under human semantic judgments. Prior work shows that models can carry latent factual knowledge and judge candidate answers; we show that these capabilities do not collapse into a single notion of ``knowing,'' and trace where information is gained, lost, or mismeasured between availability, ranking, and final choice.

cs.CL↗

Pricing IoT Data Delivered via LEO Satellites

IoT terminals served by LEO satellite constellations transmit data to passing satellites in discrete uplink windows. That data is delivered to buyers only when the satellite reaches a ground station. This store-and-forward structure makes the achievable price a discontinuous function of the delivery completeness threshold (a phenomenon we call a pricing cliff) and creates a geographic pricing asymmetry between polar and equatorial terminals that depends on constellation size. We present an integrated pricing model combining a full SGP4 orbital simulation of the deployed Kinéis constellation with a buffer flow model and a value decay framework with decay constants derived from operational decision-window timescales in the application literature and anchored to observed satellite imagery market prices. We show that polar and equatorial terminals are dominated by structurally different latency components requiring different infrastructure interventions, formally characterize the pricing cliff phenomenon and its commercial significance, and derive a geographic pricing premium that is exactly independent of the residual value assumption. Going from 1 to 25 satellites reduces the polar--equatorial premium by 25x, with diminishing returns beyond n=10; at full constellation size the premium falls below 1% of peak data value and remains below 1.5% across the full plausible range of buyer urgency.

cs.NI↗

Statistical Convergence of Transformer Encoder-Accelerated Robust Reinforcement Learning

Obtaining the optimal action-value function in Markov decision processes is computationally intensive in large state--action spaces. In this study, we present statistically rigorous convergence results for a robust reinforcement learning algorithm warm-started by a transformer-based action-value function prediction, where natural language prompts encode task specifications. Our framework adopts the R-contamination model to characterize uncertainty in the state transition kernel, and employs conformal prediction to certify convergence via trajectory-level nonconformity scores constructed from the contracting Bellman residual. The resulting conformal quantile bounds the gap between the running and optimal action-value functions simultaneously over all iterations, thereby yielding a pre-certified stopping rule that requires little knowledge of the true transition kernel. Numerical case studies on perturbed maze environments of varying size and contamination level confirm that the transformer-based warm start measurably reduces the initial error and accelerates convergence, while the proposed conformal bounds track the true error trajectory more tightly than existing guarantees.

cs.LG↗

Distributed Balanced Butterfly Counting in Signed Bipartite Graphs

The balanced butterfly is a fundamental primitive for analyzing signed bipartite graphs and provides a basis for studying higher-order structural properties, such as clustering coefficients and community structure. Despite its importance, existing approaches primarily rely on serial algorithms for balanced butterfly counting, which become inefficient on large-scale graphs. To address this limitation, we propose a distributed algorithm, D-BBC, based on a hybrid MPI+TBB framework that exploits MPI for inter-process communication and Intel TBB for intra-node parallelism. We conduct an experimental assessment of the proposed approach across 15 real-world datasets. Experimental results demonstrate that, on a single-node distributed system, D-BBC achieves average speedups of 1321x and 16.2x over the serial BB2K and multi-core M-BBC implementations, respectively. Furthermore, D-BBC achieves a maximum speedup of 23.58x over the distributed baseline S-Monarch in terms of end-to-end execution time. These results demonstrate the efficiency of the proposed distributed approach and its potential to enable high- performance signed motif analysis on large-scale bipartite graphs.

cs.DC↗

QEmbed: A Deep Learning Based Cardinality Estimator for Efficient Query Processing

Cardinality estimation is at the core of any commercial database system for efficient query processing. Over the decades, non-learning-based estimation techniques (e.g., histogram-based, sampling-based) have been widely used in both commercial and open-source database platforms. However, these techniques are only effective when the number of columns in a table is small, as they cannot properly capture dependencies between multiple attributes. Recently, learning-based approaches have been shown to perform significantly better than the heuristic methods that have been used for the past three decades. Despite this success, existing learned models often struggle to balance memory efficiency and accuracy when dealing with datasets that mix high and low cardinality attributes. In this paper, we propose a deep learning model formally called QEmbed. Our model is built upon the Masked Autoencoder for Distribution Estimation (MADE) auto-regressive framework to learn joint data distributions for selectivity estimation. To improve data representation and overcome the limitations of using a single encoding method, we design a hybrid encoding scheme that combines one-hot and embedding encodings. This hybrid design enables QEmbed to retain fine-grained attribute information for smaller domains while capturing compact semantic patterns for large, sparse domains. We capture attribute correlations by factoring the joint data distribution into a series of conditional probabilities. This approach naturally accommodates both point and range queries. Through extensive experiments, we show that while QEmbed faces a latency trade-off on extremely wide schemas, it provides highly reliable cardinality estimates overall. A key advantage of our model is that it reduces extreme tail errors (maximum Q-errors), avoiding catastrophic estimation failures on complex, highly correlated workloads.

cs.DB↗

Trajectory-Driven Multi-Product Influence Maximization in Billboard Advertising

Billboard Advertising has emerged as an effective out-of-home advertising technique, where the goal is to select a limited number of slots and play advertisement content there, with the hope that it will be observed by many people and, effectively, a significant number of them will be influenced towards the brand. Given a trajectory and a billboard database and a positive integer $k$, how can we select $k$ highly influential slots to maximize influence? In this paper, we study a variant of this problem where a commercial house wants to make a promotion of multiple products, and there is an influence demand for each product. We have studied two variants of the problem. In the first variant, our goal is to select $k$ slots such that the respective influence demand of each product is satisfied. In the other variant of the problem, we are given with $\ell$ integers $k_1,k_2, \ldots, k_{\ell}$, the goal here is to search for $\ell$ many set of slots $S_1, S_2, \ldots, S_{\ell}$ such that for all $i \in [\ell]$, $|S_{i}| \leq k_i$ and for all $i \neq j$, $S_i \cap S_j=\emptyset$ and the influence demand of each of the products gets satisfied. We model the first variant of the problem as a multi-submodular cover problem and the second variant as its generalization. To solve the common-slot variant, we formulate the problem as a multi-submodular cover problem and design a bi-criteria approximation algorithm based on the continuous greedy framework and randomized rounding. For the disjoint-slot variant, we proposed a sampling-based approximation approach along with an efficient primal-dual greedy algorithm that enforces disjointness naturally. Extensive experiments with real-world trajectory and billboard datasets highlight the effectiveness and efficiency of the proposed solution approaches.

cs.DB↗

Cast a Wider Net: Coordinated Pass@K Policy Optimization for Code Reasoning

Repeated sampling with a verifier is the standard way to allocate test-time compute for code generation, with pass@$K$ as the canonical metric. Yet the standard policy class draws $K$ independent samples from a single answer distribution, so attempts often collapse onto near-duplicate reasoning paths and waste the budget on redundant rollouts. This failure is costly in competitive programming, where many problems admit multiple distinct algorithmic strategies and pass@$K$ requires only one correct attempt. We propose Coordinated Pass@$K$ Policy Optimization (CPPO), which turns pass@$K$ generation into joint exploration over strategies: a planner emits a tuple of $K{=}4$ alternative high-level methods, and a shared solver attempts one solution per method. CPPO trains this joint policy with a multiplicative planner reward, $R_{\mathrm{plan}} = J_ψ\cdot R_{\mathrm{out}}$, assigning credit only to valid strategy tuples that lead to verifier-confirmed pass@$K$ success. Across APPS, CodeContests, and LiveCodeBench-v6, CPPO improves pass@$4$ over direct sampling, planning baselines, planner-only SFT, and pass@$K$-oriented RL under the same $K{=}4$ solver-attempt budget, with statistically significant gains on six of nine model--benchmark cells. The largest single gain is $+0.16$ on Qwen3.5-9B LiveCodeBench-v6 over the strongest baseline, PKPO ($0.588 \rightarrow 0.748$; paired bootstrap, $p < 0.05$).

cs.CL↗

EMBER: Efficient Memory via Budgeted Evidence Retention for Long-Horizon Agents

Long-horizon agents can archive large histories, but future answers still incur retrieval, rereading, and context costs. When retained memory misses answer-relevant evidence, the system must return to larger portions of the raw history. We study budgeted evidence survival: before the query is known, which source evidence should be retained so that it remains recoverable and usable under a fixed retained source-evidence token budget? We instantiate this setting as Budgeted Pre-Query Retention, where memory is written during ingestion and later read without access to the full raw stream. We introduce EMBER, a learned retention policy that constructs a compact, source-backed evidence state. EMBER stores evidence capsules: verbatim source excerpts paired with retrieval keys and update metadata, preserving both grounding and read-time access. Post-query outcome feedback trains the writer to preserve evidence across the ingestion-retrieval-answer chain. On LongMemEval-RR, our LongMemEval-derived retained-evidence protocol, EMBER-14B reaches 0.3017 F1 at the 8192-token retained-evidence comparison point, compared with 0.1765 for the strongest non-EMBER budgeted baseline. Across retained source-evidence budgets, EMBER improves F1, Retain-Recall, and Read-Recall, indicating that long-horizon memory depends on retaining evidence within the budget rather than rereading larger histories.

cs.CL↗

MEDUSA: Scalable Biometric Sensing in the Wild through Distributed MIMO Radars

Radar-based techniques for detecting vital signs have shown promise for continuous contactless vital sign sensing and healthcare applications. However, real-world indoor environments face significant challenges for existing vital sign monitoring systems. These include signal blockage in non-line-of-sight (NLOS) situations, movement of human subjects, and alterations in location and orientation. Additionally, these existing systems failed to address the challenge of tracking multiple targets simultaneously. To overcome these challenges, we present MEDUSA, a novel coherent ultra-wideband (UWB) based distributed multiple-input multiple-output (MIMO) radar system, especially it allows users to customize and disperse the $16 \times 16$ into sub-arrays. MEDUSA takes advantage of the diversity benefits of distributed yet wirelessly synchronized MIMO arrays to enable robust vital sign monitoring in real-world and daily living environments where human targets are moving and surrounded by obstacles. We've developed a scalable, self-supervised contrastive learning model which integrates seamlessly with our hardware platform. Each attention weight within the model corresponds to a specific antenna pair of Tx and Rx. The model proficiently recovers accurate vital sign waveforms by decomposing and correlating the mixed received signals, including comprising human motion, mobility, noise, and vital signs. Through extensive evaluations involving 21 participants and over 200 hours of collected data (3.75 TB in total, with 1.89 TB for static subjects and 1.86 TB for moving subjects), MEDUSA's performance has been validated, showing an average gain of 20% compared to existing systems employing COTS radar sensors. This demonstrates MEDUSA's spatial diversity gain for real-world vital sign monitoring, encompassing target and environmental dynamics in familiar and unfamiliar indoor environments.

cs.AR↗

NPUsper: Eliminating Redundant Computation for Real-Time Whisper on Mobile NPUs

We present NPUsper, a live transcription system that makes Whisper efficient on mobile NPUs by eliminating redundant computation. To avoid the heavy padding used by prior streaming systems, NPUsper detects hallucinated tokens online from temporal patterns in decoder cross-attention, allowing each inference round to process short audio inputs with minimal carryover. For efficient mobile-NPU execution, we propose controlled unrolling, which executes autoregressive decoding as K-step chunk graphs, removing unnecessary KV-cache computation and reducing graph-dispatch overhead. NPUsper achieves up to 4.84x lower per-word latency, up to 33.2x lower time-to-first-token (TTFT), and up to 88.64% lower average power consumption compared with baselines, while maintaining comparable transcription accuracy. The code is available at https://github.com/npusper/NPUsper.

cs.SD↗

Exploiting RIS Optimization Limits for Multi-User Beamforming and Signal Suppression

This paper presents a unified framework for exploiting the boundaries of reconfigurable intelligent surfaces (RIS) joint optimization in multi-user wireless systems, where a single RIS accommodates diverse objectives.We first propose an adaptive gradient-scaling mechanism that accelerates the convergence of the underlying optimization algorithm while maintaining stable performance across varying channel and system parameters. The proposed mechanism enables the solver to reach a reasonably good solution rapidly without requiring manual tuning of step sizes or algorithmic hyperparameters when system inputs change. We then propose a low-complexity beamformer recovery method tailored for single-user scenarios, which circumvents the full matrix decomposition required by traditional approaches, thereby significantly reducing computational overhead. Building on these foundations, we develop an element allocation strategy that enables user-specific prioritization through assignment of RIS subsets. This is further extended by a modular add-drop mechanism that supports partial-panel optimization in general multi-user settings. The framework is evaluated across three representative scenarios: (i) signal amplification for all users, (ii) signal suppression for all users, and (iii) selective amplification and suppression. To characterize performance limits, we derive power trade-off boundaries using scalarized joint optimization, which closely align with Monte Carlo simulations. Our unified joint optimization method consistently yield solutions near these boundaries, confirming its near-optimality. Extensive simulations under realistic channel models demonstrate that the proposed approach outperforms conventional semidefinite relaxation techniques, offering a scalable and effective RIS control strategy for cooperative and competitive multi-user environments.

eess.SP↗

Fairness-Aware Profit Maximization using Deep Reinforcement Learning

Given a social network represented as a graph where the nodes are the users and the edges represent the social relations, and a positive integer k, how to select k nodes to maximize the influence in the network remains an active area of research. In this paper, we consider a variant of the problem in which network users are associated with two parameters: a benefit value and a cost. A fixed budget is given, and the network is partitioned into communities. The task is to select a subset of users (the seed set) within the budget so that their initial activation maximizes the earned profit, while ensuring that each community realizes at least a minimum fraction of its total benefit under a maximin fairness criterion. For any seed set, the earned benefit is defined as the sum of the benefit values of the users influenced by the seed set, and the profit is defined as the difference between the earned benefit and the total cost. Formally, we call this the Fairness-Aware Profit Maximization Problem. We propose a Deep Reinforcement Learning-based approach for solving it: we first model the problem as a Markov Decision Process and subsequently propose a Deep Q-Learning Algorithm. The proposed solution has been implemented and tested on real-world social network datasets. From the reported results, we observed that the proposed approach yields a seed set whose initial activation produces up to 10 times more profit than the baseline methods. The implementation of our methodology is available at https://github.com/PoonamSharma-PY/DRL_FPM.git.

cs.SI↗

Counting Small Balanced (p,q)-bicliques in Signed Bipartite Graphs

Two disjoint sets of entities and their relationship can be modelled as a bipartite graph. Real-life examples include drug-target interaction in biological networks, user-item relationships in e-commerce networks, etc. Motif-based analysis is essential for understanding the structure of large-scale networks, and bipartite graphs are no exception. In contrast to unsigned graphs, motif analysis in signed bipartite graphs has received limited attention. The smallest non-trivial motif in a signed bipartite graph is a balanced (2,2)-biclique, often called a balanced butterfly, which captures only local patterns and cannot reveal higher-order relationships. Bipartite motifs have been studied in the literature in the context of signed bipartite graphs, such as maximal biclique, bitruss, and so on. None of these works addresses bipartite motifs with fixed-sized vertex sets, which are often relevant in practical situations. In this work, we study the balanced (p,q)-biclique counting problem for small values of p and q. As a baseline, we first adapt and extend the state-of-the-art BCList++ algorithm for unsigned bipartite graphs to incorporate edge signs, which we call SBCList++. We then propose two efficient algorithms: BBWC, a wedge-centric approach that enforces balance constraints during enumeration, and BBVP, a vertex-based pruning approach that directly enumerates feasible vertex sets. Extensive experiments on large real-world datasets demonstrate that the vertex-based pruning algorithm, BBVP, significantly outperforms the baseline, achieving an average speedup of 636$\times$ over SBCList++ (where p=q=3).

cs.DS↗

Tiny but Mighty: A Software-Hardware Co-Design Approach for Efficient Multimodal Inference on Battery-Powered Small Devices

Large Multimodal Models (LMMs) are inherently modular, comprising vision and audio encoders, a projector, and a language backbone. Yet existing systems execute them monolithically, underutilizing the heterogeneous accelerators (NPUs, GPUs, DSPs) on modern SoCs and inflating end-to-end latency. We present Nanomind, a hardware-software co-design inference framework that decomposes each LMM into modular "bricks"--vision, projector, language, and audio--and maps each brick to its best-suited compute units. A Token-Aware Buffer Manager (TABM) enables zero-copy embedding transfer across accelerators on unified-memory SoCs, bypassing CPU bottlenecks. Combined with customized hardware, a battery-aware scheduler, and fused low-bit GEMM kernels, Nanomind runs entirely on a compact, battery-powered prototype that operates fully offline. Nanomind reduces end-to-end energy by 42.3% against mainstream edge frameworks and devkits; in its on-demand low-power mode, the prototype runs LLaVA-OneVision-Qwen2-0.5B with a camera for nearly 18.8 hours on a single 2,000 mAh battery.

cs.DC↗

Accelerating Transformer-Based Monocular SLAM via Geometric Utility Scoring

Geometric Foundation Models (GFMs) have recently advanced monocular SLAM by providing robust, calibration-free 3D priors. However, deploying these models on dense video streams introduces significant computational redundancy. Current GFM-based SLAM systems typically rely on post hoc keyframe selection. Because of this, they must perform expensive dense geometric decoding simply to determine whether a frame contains novel geometry, resulting in late rejection and wasted computation. To mitigate this inefficiency, we propose LeanGate, a lightweight feed-forward frame-gating network. LeanGate predicts a geometric utility score to assess a frame's mapping value prior to the heavy GFM feature extraction and matching stages. As a predictive plug-and-play module, our approach bypasses over 90% of redundant frames. Evaluations on standard SLAM benchmarks demonstrate that LeanGate reduces tracking FLOPs by more than 85% and achieves a 5x end-to-end throughput speedup. Furthermore, it maintains the tracking and mapping accuracy of dense baselines. Project page: https://lean-gate.github.io/

cs.CV↗

Approximation Algorithms for Budget Splitting in Multi-Channel Influence Maximization

How to utilize an allocated budget effectively for branding and promotion of a commercial house is an important problem, particularly when multiple advertising media are available. There exist multiple such media, and among them, two popular ones are billboards and social media advertisements. In this context, the question naturally arises: how should a budget be allocated to maximize total influence? Although there is significant literature on the effective use of budgets in individual advertising media, there are hardly any studies examining budget allocation across multiple advertising media. To bridge this gap, this paper introduces the \textsc{Budget Splitting Problem in Billboard and Social Network Advertisement}. We introduce the notion of \emph{interaction effect} to capture the additional influence due to triggers from multiple media of advertising. Using this notion, we propose a noble influence function $Φ(,)$ that captures the total influence and shows that this function is non-negative, monotone, and non-bisubmodular. We introduce \emph{bi-submodularity ratio $(γ)$} and \emph{generalized curvature $(α)$} to measure how close a function is to being bi-submodular and how far a function is from being modular, respectively. We propose the Randomized Greedy and Two-Phase Adaptive Greedy approach, where the influence function is non-bisubmodular and achieves an approximation guarantee of $\frac{1}α\left(1-e^ {-γα} \right)$. We conducted several experiments using real-world datasets and observed that the proposed solution approach's budget splitting leads to a greater influence than existing approaches.

cs.DS↗