Implicit Regularization of SGD Reduces Shortcut Learning
Shows how the implicit regularization of SGD, controlled by learning rate and batch size, suppresses reliance on spurious features while preserving accuracy.
Paper2025–2026 · Recent work
Ideas become research.
Publications and preprints co-authored by Dr. Mahdieh Soleymani and at least one additional researcher listed on our People page.
- 40
- Recent works
- 6
- Research areas
- 2
- Publication years
Out-of-Distribution Generalization
8 worksRobust learning under distribution shifts, spurious correlations, and increasing reasoning complexity.
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.
PaperSpurious-Aware Prototype Refinement for Reliable Out-of-Distribution Detection
Refines class prototypes with explicit awareness of spurious features to make out-of-distribution detection more reliable under shortcut-driven shifts.
PaperInductive Biases for Zero-Shot Systematic Generalization in Language-Informed Reinforcement Learning
Studies compositional inductive biases that help language-informed reinforcement learning agents generalize systematically to unseen task combinations.
PaperCAREL: 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.
PaperPosition: System-2 AI Is About Complexity Out of Distribution
Argues that System-2 behavior is best understood as generalization beyond the computational complexity represented in training examples.
PaperThe Silent Helper: How Implicit Regularization Enhances Group Robustness
Studies how the implicit bias of optimization can improve worst-group performance even when group annotations are unavailable.
PaperExploiting What Trained Models Learn for Making Them Robust to Spurious Correlations without Group Annotations
Uses information already encoded by trained models to reduce shortcut reliance without requiring group labels during training.
PaperCompositional Text-to-Image Generation
9 worksImproving object coverage, spatial alignment, and compositional control in diffusion models.
CARINOX: Inference-Time Scaling with Category-Aware Reward-Based Initial Noise Optimization and Exploration
Improves compositional text-to-image generation by optimizing and exploring initial noise with category-aware rewards at inference time.
PaperAttention Overlap Is Responsible for the Entity Missing Problem in Text-to-Image Diffusion Models!
Connects missing entities in generated images to overlapping attention maps and develops an inference-time strategy for better object coverage.
PaperWhy Settle for Mid: A Probabilistic Viewpoint to Spatial Relationship Alignment in Text-to-Image Models
Takes a probabilistic view of spatial relationships to improve how text-to-image models place and align objects described in prompts.
PaperNo Concept Left Behind: Test-Time Optimization for Compositional Text-to-Image Generation
Optimizes generation at test time to preserve every requested concept and improve compositional prompt fidelity.
PaperInfinity and Beyond: Compositional Alignment in VAR and Diffusion T2I Models
Compares compositional alignment in visual autoregressive and diffusion models to expose their distinct strengths and failure modes.
PreprintEvaluating the Evaluators: Metrics for Compositional Text-to-Image Generation
Systematically evaluates whether popular metrics reliably measure object, attribute, and relationship alignment in generated images.
PreprintT2I-FineEval: Fine-Grained Compositional Metric for Text-to-Image Evaluation
Introduces a fine-grained metric for measuring whether generated images satisfy the individual components of complex prompts.
PaperFine-Grained Alignment and Noise Refinement for Compositional Text-to-Image Generation
Refines alignment signals and initial noise to improve compositional correctness without retraining the text-to-image model.
PreprintDiffusion Beats Autoregressive: An Evaluation of Compositional Generation in Text-to-Image Models
Benchmarks diffusion and autoregressive image generators and analyzes their relative ability to follow compositional prompts.
PaperVision-Language Models
7 worksMultilingual evaluation, multi-object representations, and structured visual reasoning.
MEENA (PersianMMMU): Multimodal-Multilingual Educational Exams for N-Level Assessment
Introduces a large bilingual Persian–English benchmark for evaluating vision-language models across educational, scientific, and cultural reasoning tasks.
PaperVisual 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.
PaperCLIP Under the Microscope: A Fine-Grained Analysis of Multi-Object Representation
Provides a controlled analysis of how CLIP represents multiple objects and reveals limitations that aggregate benchmarks can conceal.
PaperHaloProbe: Bayesian Detection and Mitigation of Object Hallucinations in Vision-Language Models
Uses Bayesian probing to detect object hallucinations and guide mitigation in vision-language model responses.
PaperHidden Meanings in Plain Sight: RebusBench for Evaluating Cognitive Visual Reasoning
Introduces a rebus-puzzle benchmark that tests whether multimodal models can integrate visual form, language, and abstract concepts.
PreprintUncovering Grounding IDs: How External Cues Shape Multimodal Binding
Studies how external cues create grounding identifiers that help multimodal models bind objects to their properties and locations.
PreprintAnalyzing CLIP's Performance Limitations in Multi-Object Scenarios: A Controlled High-Resolution Study
Uses controlled high-resolution data to isolate CLIP's limitations when images contain multiple objects and attributes.
PaperInterpretability
4 worksMechanistic analysis and faithful attribution methods for language and vision transformers.
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.
PaperLibraGrad: Balancing Gradient Flow for Universally Better Vision Transformer Attributions
Balances gradient flow to produce more stable, faithful, and broadly effective attribution maps for vision transformer architectures.
PaperCausal Attribution via Activation Patching
Develops activation-patching methods for tracing causal contributions inside neural networks rather than relying on correlational explanations.
PreprintUnderstanding Counting Mechanisms in Large Language and Vision-Language Models
Compares internal counting strategies across language and vision-language models to reveal shared mechanisms and modality-specific failures.
PreprintLarge Language Models
7 worksReasoning confidence, hidden shortcuts, evaluation, and language models for scientific discovery.
Large Language Models for Scientific Idea Generation: A Creativity-Centered Survey
Surveys the use of large language models for scientific idea generation through a creativity-centered framework and evaluation perspective.
PaperCER: Confidence Enhanced Reasoning in LLMs
Uses model confidence to improve reasoning reliability and help language models make better decisions during multi-step inference.
PaperLying to Win: Assessing LLM Deception through Human-AI Games and Parallel-World Probing
Evaluates strategic deception through interactive games and probes whether models represent truthful and deceptive alternatives internally.
PreprintThe Judge Who Never Admits: Hidden Shortcuts in LLM-Based Evaluation
Shows that language-model judges can exploit hidden shortcuts while producing explanations that fail to acknowledge those cues.
PreprintThe Illusion of Procedural Reasoning: Measuring Long-Horizon FSM Execution in LLMs
Tests long-horizon finite-state-machine execution to separate genuine procedural reasoning from shallow pattern matching.
PreprintUnspoken Hints: Accuracy Without Acknowledgement in LLM Reasoning
Examines cases where implicit hints improve answers even though model explanations do not acknowledge using them.
PreprintThe Silent Judge: Unacknowledged Shortcut Bias in LLM-as-a-Judge
Demonstrates that LLM evaluators can rely on shortcut signals while presenting apparently independent justifications for their scores.
PreprintOther Work
5 worksCollaborative research in reinforcement learning, adversarial bandits, emergent communication, and medical image segmentation.
SUSD: Structured Unsupervised Skill Discovery through State Factorization
Factorizes environment states and skill variables to learn richer, more diverse, and compositionally reusable behaviors without external rewards.
PaperEfficient 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.
PaperTest-Time Confidence-Aware Query Labeling Improves 3D Few-Shot Segmentation
Uses confidence-aware query labeling at test time to improve the accuracy and adaptability of three-dimensional few-shot segmentation.
PaperSelf-Supervised 3D Medical Image Segmentation by Flow-Guided Mask Propagation Learning
Learns three-dimensional medical image segmentation without dense labels by propagating masks through flow-guided self-supervision.
PaperVQEL: Enabling Self-Developed Symbolic Language in Agents through Vector Quantization in Emergent Language Games
Uses vector quantization to help interacting agents develop a discrete symbolic language in emergent communication games.
PreprintComplete research record
Looking for the full publication history?
For the complete research record, including earlier work, preprints, and citation information, visit Dr. Mahdieh Soleymani's Google Scholar profile.