arXiv subjects

Nima Safaei

Publications and source records attributed to Nima Safaei.

9 recordsLinked to original sources

arXiv2026

LiteSearch-VL: Small Multimodal Search Agents via Trajectory Distillation and Synthetic Step-DPO

Multimodal search agents answer visual questions by interleaving image understanding, web retrieval, tool use, and evidence synthesis. Strong systems exist, but in two expensive regimes: proprietary frontier models such as GPT-5 and Gemini, or large open vision-language backbones trained with substantial agentic data and reinforcement learning. We ask a different question: when released agent trajectories are distilled into much smaller backbones under a single-node budget, what is actually transferred? We study this with LiteSearch-VL, a low-compute recipe for Qwen3-VL-2B and Qwen3-VL-4B that uses only released OpenSearch-VL trajectories, parameter-efficient LoRA adapters, and synthetic step-level preferences: DPO on GPT-5-generated hard negatives targeting five local failure modes (premature answer, wrong tool, weak query, repeated query, ignored image). Across 12,400 GPT-5-judged rollouts on SimpleVQA, FVQA, LiveVQA, and VDR-Bench-testmini, the dominant effect is behavioral rather than a uniform accuracy lift: full-trajectory supervised fine-tuning transfers the agent contract, taking the 2B model from almost never emitting a usable answer (1,237/1,240 no_answer rollouts) to 28.4% macro Pass@1, matching or slightly exceeding the off-the-shelf 4B base (25.6%). Synthetic preference learning and compact tool distillation act as refinements rather than phase transitions (best 4B configuration: 30.8% macro Pass@1). Finally, a controlled VDR step-budget ablation shows that extra search turns convert abstentions into wrong_entity errors rather than correct answers, identifying answer verification, not search depth, as the next bottleneck for small multimodal agents.

cs.AI↗

arXiv2026

Understanding Pruning Regimes in Vision-Language Models Through Domain-Aware Layer Selection

Transformer-based vision-language models (VLMs) contain substantial depth redundancy, yet the effect of removing specific decoder layers remains poorly understood, especially for domains that require tight coupling between perception and multi-step reasoning. We study structured decoder layer pruning through the lens of domain-aware activation similarity, measuring how strongly each layer transforms representations for math versus non-math inputs. This yields simple math-aware, non-math-aware, and mixed ranking criteria that identify layers whose input-output activations change least within a target domain. Across two state-of-the-art VLMs and a broad suite of math and general multimodal benchmarks, we uncover a consistent three-regime structure: at low pruning budgets, performance is highly sensitive to which layers are removed; at moderate budgets, methods converge as structural damage accumulates; and at high budgets, structural continuity dominates, favoring spacing-aware strategies. Our domain-aware rankings achieve the strongest stability in the ranking-sensitive regime, while matching or exceeding structure-aware baselines at larger budgets. These results provide a clearer picture of how depth contributes to domain-specific behavior in VLMs and offer a practical, interpretable approach to reducing model depth without sacrificing essential mathematical or general vision-language capabilities.

cs.CV↗

arXiv2026

VisTIRA: Closing the Image-Text Modality Gap in Visual Math Reasoning via Structured Tool Integration

Vision-language models (VLMs) lag behind text-only language models on mathematical reasoning when the same problems are presented as images rather than text. We empirically characterize this as a modality gap: the same question in text form yields markedly higher accuracy than its visually typeset counterpart, due to compounded failures in reading dense formulas, layout, and mixed symbolic-diagrammatic context. First, we introduce VisTIRA (Vision and Tool-Integrated Reasoning Agent), a tool-integrated reasoning framework that enables structured problem solving by iteratively decomposing a given math problem (as an image) into natural language rationales and executable Python steps to determine the final answer. Second, we build a framework to measure and improve visual math reasoning: a LaTeX-based pipeline that converts chain-of-thought math corpora (e.g., NuminaMath) into challenging image counterparts, and a large set of synthetic tool-use trajectories derived from a real-world, homework-style image dataset (called SnapAsk) for fine-tuning VLMs. Our experiments show that tool-integrated supervision improves image-based reasoning, and OCR grounding can further narrow the gap for smaller models, although its benefit diminishes at scale. These findings highlight that modality gap severity inversely correlates with model size, and that structured reasoning and OCR-based grounding are complementary strategies for advancing visual mathematical reasoning.

cs.AI↗

arXiv2022

Winter wheat yield prediction using convolutional neural networks from environmental and phenological data

Crop yield forecasting depends on many interactive factors, including crop genotype, weather, soil, and management practices. This study analyzes the performance of machine learning and deep learning methods for winter wheat yield prediction using an extensive dataset of weather, soil, and crop phenology variables in 271 counties across Germany from 1999 to 2019. We proposed a Convolutional Neural Network (CNN) model, which uses a 1-dimensional convolution operation to capture the time dependencies of environmental variables. We used eight supervised machine learning models as baselines and evaluated their predictive performance using RMSE, MAE, and correlation coefficient metrics to benchmark the yield prediction results. Our findings suggested that nonlinear models such as the proposed CNN, Deep Neural Network (DNN), and XGBoost were more effective in understanding the relationship between the crop yield and input data compared to the linear models. Our proposed CNN model outperformed all other baseline models used for winter wheat yield prediction (7 to 14% lower RMSE, 3 to 15% lower MAE, and 4 to 50% higher correlation coefficient than the best performing baseline across test data). We aggregated soil moisture and meteorological features at the weekly resolution to address the seasonality of the data. We also moved beyond prediction and interpreted the outputs of our proposed CNN model using SHAP and force plots which provided key insights in explaining the yield prediction results (importance of variables by time). We found DUL, wind speed at week ten, and radiation amount at week seven as the most critical features in winter wheat yield prediction.

cs.LG↗

arXiv2021

WheatNet: A Lightweight Convolutional Neural Network for High-throughput Image-based Wheat Head Detection and Counting

For a globally recognized planting breeding organization, manually-recorded field observation data is crucial for plant breeding decision making. However, certain phenotypic traits such as plant color, height, kernel counts, etc. can only be collected during a specific time-window of a crop's growth cycle. Due to labor-intensive requirements, only a small subset of possible field observations are recorded each season. To help mitigate this data collection bottleneck in wheat breeding, we propose a novel deep learning framework to accurately and efficiently count wheat heads to aid in the gathering of real-time data for decision making. We call our model WheatNet and show that our approach is robust and accurate for a wide range of environmental conditions of the wheat field. WheatNet uses a truncated MobileNetV2 as a lightweight backbone feature extractor which merges feature maps with different scales to counter image scale variations. Then, extracted multi-scale features go to two parallel sub-networks for simultaneous density-based counting and localization tasks. Our proposed method achieves an MAE and RMSE of 3.85 and 5.19 in our wheat head counting task, respectively, while having significantly fewer parameters when compared to other state-of-the-art methods. Our experiments and comparisons with other state-of-the-art methods demonstrate the superiority and effectiveness of our proposed method.

cs.CV↗

arXiv2020

Regularization and False Alarms Quantification: Two Sides of the Explainability Coin

Regularization is a well-established technique in machine learning (ML) to achieve an optimal bias-variance trade-off which in turn reduces model complexity and enhances explainability. To this end, some hyper-parameters must be tuned, enabling the ML model to accurately fit the unseen data as well as the seen data. In this article, the authors argue that the regularization of hyper-parameters and quantification of costs and risks of false alarms are in reality two sides of the same coin, explainability. Incorrect or non-existent estimation of either quantities undermines the measurability of the economic value of using ML, to the extent that might make it practically useless.

cs.LG↗

arXiv2020

Gasoline Pricing Policies for Transportation Safety

Economic factors can have substantial effects on transportation crash trends. This study makes a comprehensive examination of the relationship between the retail gasoline price (including state and federal fuel taxes) and transportation fatal crashes from 2007 to 2016 in the US. Data on motor vehicle, bicycle and pedestrian fatal crashes come from Fatality Analysis Reporting System (FARS) provided by the National Highway Safety Administration (NHTSA) and the gasoline price data is from U.S. Energy Information Administration (EIA). Random effect negative binomial regression models are used to estimate the impact of inflation-adjusted gasoline prices on trends of transportation fatal crashes. Initial results combined with results of previous studies showed that gender and transportation mean type (motorcycle, non-motorcycle, bicycle and pedestrian) play prominent roles in interpreting the final model, so by using random effect negative binomial regression, seven models are developed to evaluate the effects of gasoline price changes on total population, male, female, motorcyclists, non-motorcyclists, bicyclists and pedestrians separately. Our findings suggest that increasing the gasoline prices will not significantly alter the number of total fatal crashes. However, by looking at different vehicle types, it is estimated that one dollar increase in adjusted gasoline price is associated with 24.2% increase in the number of motorcycle fatal crashes, 1.9% decrease in the number of non-motorcycle fatal crashes, and 0.7% decrease in the number of pedestrian fatal crashes. Also, there is no noticeable difference between male and female in response to the gasoline price changes.

stat.AP↗

arXiv2019

Correspondent Banking Networks: Theory and Experiment

We employ the mathematical programming approach in conjunction with the graph theory to study the structure of correspondent banking networks. Optimizing the network requires decisions to be made to onboard, terminate or restrict the bank relationships to optimize the size and overall risk of the network. This study provides theoretical foundation to detect the components, the removal of which does not affect some key properties of the network such as connectivity and diameter. We find that the correspondent banking networks have a feature we call k-accessibility, which helps to drastically reduce the computational burden required for finding the above mentioned components. We prove a number of fundamental theorems related to k-accessible directed graphs, which should be also applicable beyond the particular problem of financial networks. The theoretical findings are verified through the data from a large international bank.

cs.LG↗

arXiv2018

A Swift Heuristic Method for Work Order Scheduling under the Skilled-Workforce Constraint

The considered problem is how to optimally allocate a set of jobs to technicians of different skills such that the number of technicians of each skill does not exceed the number of persons with that skill designation. The key motivation is the quick sensitivity analysis in terms of the workforce size which is quite necessary in many industries in the presence of unexpected work orders. A time-indexed mathematical model is proposed to minimize the total weighted completion time of the jobs. The proposed model is decomposed into a number of single-skill sub-problems so that each one is a combination of a series of nested binary Knapsack problems. A heuristic procedure is proposed to solve the problem. Our experimental results, based on a real-world case study, reveal that the proposed method quickly produces a schedule statistically close to the optimal one while the classical optimal procedure is very time-consuming.

cs.AI↗