Alireza Hashemi

Alireza Hashemi, Hernan Makse

We extend the graph convolutional network method for deep learning on graph data to higher order in terms of neighboring nodes. In order to construct representations for a node in a graph, in addition to the features of the node and its immediate neighboring nodes, we also include more distant nodes in the calculations. In experimenting with a number of publicly available citation graph datasets, we show that this higher order neighbor visiting pays off by outperforming the original model especially when we have a limited number of available labeled data points for the training of the model.

Hamed Mahdavi, Pouria Mahdavinia, Alireza Farhadi et al.

State-of-the-art (SOTA) LLMs have progressed from struggling on proof-based Olympiad problems to solving most of the IMO 2025 problems, with leading systems reportedly handling 5 of 6 problems. Given this progress, we assess how well these models can grade proofs: detecting errors, judging their severity, and assigning fair scores beyond binary correctness. We study proof-analysis capabilities using a corpus of 90 Gemini 2.5 Pro-generated solutions that we grade on a 1-4 scale with detailed error annotations, and on MathArena solution sets for IMO/USAMO 2025 scored on a 0-7 scale. Our analysis shows that models can reliably flag incorrect (including subtly incorrect) solutions but exhibit calibration gaps in how partial credit is assigned. To address this, we introduce agentic workflows that extract and analyze reference solutions and automatically derive problem-specific rubrics for a multi-step grading process. We instantiate and compare different design choices for the grading workflows, and evaluate their trade-offs. Across our annotated corpus and MathArena, our proposed workflows achieve higher agreement with human grades and more consistent handling of partial credit across metrics. We release all code, data, and prompts/logs to facilitate future research.

Forough Mehralian, Ryan Shar, James R. Rae et al.

As large language models become increasingly capable of generating code, evaluating their performance remains a complex and evolving challenge. Existing benchmarks primarily focus on functional correctness, overlooking the diversity of real-world coding tasks and developer expectations. To this end, we introduce a multi-language benchmark that evaluates LLM instruction-following capabilities and is extensible to operate on any set of standalone coding problems. Our benchmark evaluates instruction following in two key settings: adherence to pre-defined constraints specified with the initial problem, and the ability to perform refinements based on follow-up instructions. For this paper's analysis, we empirically evaluated our benchmarking pipeline with programming tasks from LiveBench, that are also automatically translated from Python into Java and JavaScript. Our automated benchmark reveals that models exhibit differing levels of performance across multiple dimensions of instruction-following. Our benchmarking pipeline provides a more comprehensive evaluation of code generation models, highlighting their strengths and limitations across languages and generation goals.

Heinrich Peters, Alireza Hashemi, James Rae

Machine learning (ML) and artificial intelligence (AI) systems rely heavily on human-annotated data for training and evaluation. A major challenge in this context is the occurrence of annotation errors, as their effects can degrade model performance. This paper presents a predictive error model trained to detect potential errors in search relevance annotation tasks for three industry-scale ML applications (music streaming, video streaming, and mobile apps). Drawing on real-world data from an extensive search relevance annotation program, we demonstrate that errors can be predicted with moderate model performance (AUC=0.65-0.75) and that model performance generalizes well across applications (i.e., a global, task-agnostic model performs on par with task-specific models). In contrast to past research, which has often focused on predicting annotation labels from task-specific features, our model is trained to predict errors directly from a combination of task features and behavioral features derived from the annotation process, in order to achieve a high degree of generalizability. We demonstrate the usefulness of the model in the context of auditing, where prioritizing tasks with high predicted error probabilities considerably increases the amount of corrected annotation errors (e.g., 40% efficiency gains for the music streaming application). These results highlight that behavioral error detection models can yield considerable improvements in the efficiency and quality of data annotation processes. Our findings reveal critical insights into effective error management in the data annotation process, thereby contributing to the broader field of human-in-the-loop ML.

Hamed Mahdavi, Alireza Hashemi, Majid Daliri et al.

Recent advances in large language models (LLMs) have shown impressive progress in mathematical reasoning tasks. However, current evaluation benchmarks predominantly focus on the accuracy of final answers, often overlooking the crucial logical rigor for mathematical problem solving. The claim that state-of-the-art LLMs can solve Math Olympiad-level problems requires closer examination. To explore this, we conducted both qualitative and quantitative human evaluations of proofs generated by LLMs, and developed a schema for automatically assessing their reasoning capabilities. Our study reveals that current LLMs fall significantly short of solving challenging Olympiad-level problems and frequently fail to distinguish correct mathematical reasoning from clearly flawed solutions. Our analyses demonstrate that the occasional correct final answers provided by LLMs often result from pattern recognition or heuristic shortcuts rather than genuine mathematical reasoning. These findings underscore the substantial gap between LLM performance and human expertise in advanced mathematical reasoning and highlight the importance of developing benchmarks that prioritize the soundness of the reasoning used to arrive at an answer rather than the mere correctness of the final answers.

Hamed Mahdavi, Pouria Mahdavinia, Samira Malek et al.

State-of-the-art (SOTA) LLMs have progressed from struggling on proof-based Olympiad problems to solving most of the IMO 2025 problems, with leading systems reportedly handling 5 of 6 problems. Given this progress, we assess how well these models can grade proofs: detecting errors, judging their severity, and assigning fair scores beyond binary correctness. We study proof-analysis capabilities using a corpus of 90 Gemini 2.5 Pro-generated solutions that we grade on a 1-4 scale with detailed error annotations, and on MathArena solution sets for IMO/USAMO 2025 scored on a 0-7 scale. Our analysis shows that models can reliably flag incorrect (including subtly incorrect) solutions but exhibit calibration gaps in how partial credit is assigned. To address this, we introduce agentic workflows that extract and analyze reference solutions and automatically derive problem-specific rubrics for a multi-step grading process. We instantiate and compare different design choices for the grading workflows, and evaluate their trade-offs. Across our annotated corpus and MathArena, our proposed workflows achieve higher agreement with human grades and more consistent handling of partial credit across metrics. We release all code, data, and prompts/logs to facilitate future research.

Majid Ashouri, Alireza Hashemi

In this paper, we employed a transfer learning technique to predict the Nusselt number for natural convection flows in enclosures. Specifically, we considered the benchmark problem of a two-dimensional square enclosure with isolated horizontal walls and vertical walls at constant temperatures. The Rayleigh and Prandtl numbers are sufficient parameters to simulate this problem numerically. We adopted two approaches to this problem: Firstly, we made use of a multi-grid dataset in order to train our artificial neural network in a cost-effective manner. By monitoring the training losses for this dataset, we detected any significant anomalies that stemmed from an insufficient grid size, which we further corrected by altering the grid size or adding more data. Secondly, we sought to endow our metamodel with the ability to account for additional input features by performing transfer learning using deep neural networks. We trained a neural network with a single input feature (Rayleigh) and extended it to incorporate the effects of a second feature (Prandtl). We also considered the case of hollow enclosures, demonstrating that our learning framework can be applied to systems with higher physical complexity, while bringing the computational and training costs down.