publications

Research publications in reinforcement learning, generative AI, and interpretable machine learning.

2026

ICML

Expected Return Causes Outcome-Level Mode Collapse in Reinforcement Learning and How to Fix It with Inverse Probability Scaling

Abhijeet Sinha, Sundari Elango, and Dianbo Liu

In International Conference on Machine Learning, 2026

Abs HTML

We explain outcome-level mode collapse induced by expected-return objectives and introduce inverse-probability scaling to improve coverage of diverse high-reward outcomes.
CVPR

How Does My Model Fail? Automatic Identification and Interpretation of Physical Plausibility Failure Modes with Matryoshka Transcoders

Yiming Tang, Abhijeet Sinha, and Dianbo Liu

In CVPR Workshop on Explainable AI for Computer VisionSpotlight, CVPR XAI4CV Workshop 2026 , 2026

Spotlight Abs HTML

Spotlight presentation

Matryoshka Transcoders learn hierarchical sparse features for coarse-to-fine analysis of physical-plausibility failures in generative models.

2023

ICONIP

Brain-Inspired Attention Model for Object Counting

Abhijeet Sinha, Sweta Kumari, and V. Srinivasa Chakravarthy

In International Conference on Neural Information Processing, 2023

Abs DOI HTML

A Q-learning-based sequential attention model with recurrent visual glimpses achieved 92.1 percent object-counting accuracy.

AAAI

Density-Aware Reward Scaling: A Reinforcement Learning Framework for Reward-Proportional Sampling

Abhijeet Sinha, Ria Shekhawat, and Dianbo Liu

Submitted to AAAI 2027

Abs

DARS is a critic-free reward-processing method that inversely scales rewards by local outcome density while retaining the standard policy-gradient objective.