Generalization
Learning systems that remain reliable beyond their training distributions.
About the lab
Building machine learning systems that understand, reason, and generalize.
The Machine Learning Lab (MLL), directed by Dr. Mahdieh Soleymani, is a research group at Sharif University of Technology. We investigate fundamental and applied questions across modern machine learning, from robust generalization to compositional intelligence.
Compositional Learning
Understanding and combining concepts, attributes, and relations.
Multimodal Intelligence
Studying vision-language models, LVLMs, and reinforcement learning agents.
Generative Models
Improving control, faithfulness, and reasoning in generative systems.
Research
Questions that drive our work
We study how learning systems represent knowledge, reason over complex inputs, and remain dependable when the world differs from their training data.
Out-of-Distribution GeneralizationBuilding models that generalize beyond training distributions.
Robustness to Spurious Correlation
We develop frameworks that mitigate shortcut learning and improve robustness under subpopulation shifts, including realistic settings where spurious correlations are unknown or unlabeled.
Language-Informed Reinforcement Learning
We study how language abstractions and compositional representations can guide agents toward robust performance in unseen environments.
Compositional and Complex OoD
We create benchmarks and learning methods for shifts formed by novel compositions of previously seen concepts.
Compositional Text-to-Image GenerationFaithfully composing objects, attributes, and relations.
We design test-time strategies that improve compositional control in text-to-image generation. Our work addresses missing objects, attribute binding, and relational failures by enforcing stronger alignment between textual descriptions and visual elements.
Vision-Language ModelsCompositional understanding and visual reasoning.
Visual Reasoning
We develop benchmarks and methods that probe multi-step reasoning and relational understanding in large vision-language models.
Representational Analysis
We examine how models such as CLIP represent objects, attributes, and their interactions to identify where compositionality emerges—and where it breaks.
InterpretabilityTracing and explaining computations inside modern models.
Mechanistic Interpretability
We develop intervention-based tools to trace how semantic functions form and transfer across layers in language and vision-language models.
Vision Transformer Attribution
We introduce more stable and faithful attribution techniques that address known failure modes in transformer explanations.
Large Language ModelsReasoning, agency, safety, and scientific discovery.
Persian Benchmarks
High-quality evaluation resources for reasoning and generation in a low-resource language.
Reasoning
Methods for more systematic, compositional, and logically consistent multi-step reasoning.
LLM Agents
Planning, memory, tool use, and coordination in long-horizon environments.
Scientific Discovery
Language models as tools for idea generation, pattern discovery, and scientific reasoning.
Chain-of-Thought Faithfulness
Evaluation of deceptive or misleading reasoning behavior in language models.
Our community
People
Researchers and students working across the lab's core areas.
Director
Teaching
Courses
Selected courses taught by members of the lab at Sharif University of Technology.
Artificial Intelligence
Visit course websiteSystem II
Visit course websiteModern Information Retrieval
Website unavailable
Deep Learning
Website unavailable
Large Language Models
Visit course websiteMachine Learning
Website unavailable
Publications
Recent work from the lab.
Selected recent publications spanning mechanistic interpretability, visual reasoning, reinforcement learning, and robust generalization.
Mechanistic Interpretability of Large-Scale Counting in LLMs through a System-2 Strategy
Reveals how a System-2 strategy decomposes large counting problems, transfers partial counts through dedicated attention heads, and aggregates them reliably.
Bridging Reasoning to Learning: Unmasking Illusions using Complexity Out-of-Distribution Generalization
Frames reasoning as generalization to problems whose solutions require greater representational or computational complexity than the training examples.
Efficient Adversarial Attacks on High-Dimensional Offline Bandits
Demonstrates how small targeted perturbations to reward models can hijack offline bandits and studies a practical partial defense.
SUSD: Structured Unsupervised Skill Discovery through State Factorization
Factorizes environment states and skill variables to learn richer, more diverse, and compositionally reusable behaviors without rewards.
Visual Structures Help Visual Reasoning: Addressing the Binding Problem in LVLMs
Introduces VISER, which adds lightweight spatial structure and sequential-scanning prompts to improve counting, visual search, and spatial reasoning.
CAREL: Instruction-Guided Reinforcement Learning with Cross-Modal Auxiliary Objectives
Uses cross-modal auxiliary objectives and instruction tracking to improve sample efficiency and systematic generalization in language-guided reinforcement learning.
Open positions
Interested in working with us?
No specific vacancies are currently listed. Strong prospective students whose interests align with our research are welcome to introduce themselves by email.
Contact
Start a conversation.
For research inquiries, prospective student introductions, and collaboration opportunities, contact the Machine Learning Lab at Sharif University of Technology.
soleymani@sharif.edu