B1#262026
Tractable and Expressive Generative Modeling with Probabilistic Flow Circuits
Neurosymbolic AI: Foundations and Applications, pages 183–222. Wiley Online Library, 2026
Research Agenda
Expressive, tractable, and trustworthy probabilistic AI.
I develop probabilistic AI systems that combine expressive generative modeling, tractable inference, and trustworthy reasoning. My work centers on probabilistic circuits and related generative models that represent uncertainty explicitly, answer probabilistic queries efficiently, and support reliable decisions under missing, noisy, or conflicting evidence.
Model
complex structured distributions
Infer
probabilistic queries efficiently
Evaluate
learning dynamics and reliability
Deploy
with robustness and human guidance
Research Map
One agenda across modeling, learning, deployment, and reasoning.
The sections are connected by a single question: how can uncertainty-aware models become expressive enough to use, reliable enough to deploy, and structured enough to inspect?
Theme 017 connected papers
Expressive Tractable Probabilistic Generative Models
Probabilistic CircuitsNormalizing FlowsVoronoi RoutingRepresentation LearningExact Inference
01
Core Question
How can generative models capture complex data distributions while preserving useful probabilistic inference?
02
Approach
I develop probabilistic generative models that combine the semantics of tractable probabilistic circuits with the representation power of flows, geometry-aware routing, and learned representations.
03
Visual Model
Tractable composition with expressive structure
04
Research Problems
Increase expressivity without silently breaking tractable inference.
Align neural transformations and local geometry with circuit factorization.
Learn probabilistic representations that remain useful when evidence is incomplete.
05
What This Enables
Models that can generate, score, marginalize, condition, and reason about missing or partial evidence in a single probabilistic framework.
06
Representative Papers
Selected work connected to this theme.
Theme 023 connected papers
Understanding Learning Dynamics in Generative Models
Duality GapSharpness-Aware LearningOptimization GeometryHessian StructureScore-Based Learning
01
Core Question
How do generative models learn, converge, overfit, and generalize?
02
Approach
I study training behavior through measurable quantities such as duality gaps, curvature, sharpness, and tractable second-order structure, connecting adversarial training diagnostics with PC optimization.
03
Visual Model
Sharp and flat optima in generative learning
04
Research Problems
Diagnose when adversarial generative training is converging or cycling.
Understand when expressive PCs overfit through sharp likelihood landscapes.
Use tractable model structure to design reliable generative learning objectives.
05
What This Enables
Training procedures whose behavior can be monitored, analyzed, and improved instead of treated as a black box.
06
Representative Papers
Selected work connected to this theme.
C14#252026
Tractable Sharpness-Aware Learning of Probabilistic Circuits
The 40th Annual AAAI Conference on Artificial Intelligence (AAAI), 2026
C1#12021
On Characterizing GAN Convergence Through Proximal Duality Gap
Proceedings of the 38th International Conference on Machine Learning (ICML), PMLR 139, 9660–9670, 2021
Theme 036 connected papers
Robust and Reliable Deployability
Few-Shot LearningTask ShiftRobust EvaluationMultimodal FusionCredibility Estimation
01
Core Question
How can AI systems remain reliable when evidence is limited, shifted, missing, noisy, or conflicting?
02
Approach
I design evaluation protocols and probabilistic models for deployment stressors: few-shot adaptation, task shift, missing modalities, corrupted sources, and conflicting information.
03
Visual Model
Reliability under missing, noisy, and conflicting evidence
Robust multimodal fusion
Reliability analysis for late multimodal fusion with probabilistic circuits.
04
Research Problems
Stress-test adaptation when tasks differ from the training regime.
Model reliability under missing, corrupted, or contradictory modalities.
Expose deployment failures that average benchmark accuracy can hide.
05
What This Enables
Systems that know when adaptation helps, when evidence is unreliable, and how to make decisions under realistic uncertainty.
06
Representative Papers
Selected work connected to this theme.
C13#242026
Context-Specific Credibility-Aware Multimodal Fusion with Conditional Probabilistic Circuits
The 29th International Conference on Information Fusion (FUSION), 2026
C6#142024
On the Robustness and Reliability of Late Multi-Modal Fusion using Probabilistic Circuits
The 27th International Conference on Information Fusion (FUSION), 2024
J3#102023
Leveraging Task Variability in Meta-Learning
SN Computer Science, 4(5), 539. Springer, 2023
Theme 046 connected papers
Human-Allied Learning and Reasoning
Human-Allied LearningDomain ConstraintsNeuro-Symbolic AIKnowledge GraphsProbabilistic Reasoning
01
Core Question
How can probabilistic models incorporate human guidance, domain knowledge, constraints, and credibility-aware reasoning?
02
Approach
I use tractable probabilistic inference as a substrate for constraint-aware learning, human feedback, knowledge-guided modeling, and auditable reasoning about source credibility.
03
Visual Model
Human guidance and probabilistic reasoning
Human-allied probabilistic circuits
A unified interface for incorporating feedback and constraints into PC learning.
04
Research Problems
Express domain knowledge as probabilistic constraints during learning.
Support human feedback without losing calibrated uncertainty.
Use inference queries to make credibility and reasoning inspectable.
05
What This Enables
AI systems that collaborate with human expertise, respect structured knowledge, and expose the probabilistic assumptions behind their decisions.
06
Representative Papers
Selected work connected to this theme.
C11#202025
Human-Allied Relational Reinforcement Learning
The Twelfth Annual Conference on Advances in Cognitive Systems (ACS), 2025
C10#192025
Scalable Knowledge Graph Construction from Unstructured Text: A Case Study on Artisanal and Small-Scale Gold Mining
The 29th Pacific-Asia Conference on Knowledge Discovery and Data Mining (PAKDD), 2025
C9#182025
A Unified Framework for Human-Allied Learning of Probabilistic Circuits
The 39th Annual AAAI Conference on Artificial Intelligence (AAAI), 2025
Cross-Cutting Methods
Common tools and evaluation lenses that connect the research program.
Probabilistic CircuitsNormalizing FlowsTractable InferenceGenerative ModelingOptimization GeometryCredibility ModelingDomain ConstraintsHuman GuidanceRobust EvaluationMultimodal Fusion