Publications

COLM 2026

Can LLM Agents Discover? Evaluating Creativity on ML Engineering Tasks

Shitanshu Bhushan, Yunxiang Zhang, Lu Wang

OpenReview

ICML 2026

LiveOIBench: Can Large Language Models Outperform Human Contestants in Informatics Olympiads?

Kaijian Zou, Aaron Xiong, Yunxiang Zhang, Frederick Zhang, Yueqi Ren, Jirong Yang, Ayoung Lee, Shitanshu Bhushan, Lu Wang

arXiv

NeurIPS Datasets and Benchmarks Track 2025

MLRC-Bench: Can Language Agents Solve Machine Learning Research Challenges?

Yunxiang Zhang, Muhammad Khalifa, Shitanshu Bhushan, Grant D Murphy, Lajanugen Logeswaran, Jaekyeom Kim, Moontae Lee, Honglak Lee, Lu Wang

arXiv

Blog Posts

Loading posts...

Projects

LLM Catan Battle

Feb 2025

  • Implemented a system to have different LLMs play Catan against each other using the Catanatron framework, pitting GPT-4o-mini against Gemini 2.0 Flash in strategic gameplay scenarios
  • Observed and analyzed LLM performance in complex game-theoretic situations, finding that current LLMs struggle with the strategic depth of Catan despite their general capabilities
LLMPythonGame TheoryAI

LMAOCaT: Low-rank Mamba and gated Attention Optimization

Oct 2024 - Dec 2024

  • Recreated LoLCATs and developed a hybrid attention framework integrating Gated Linear Attention and Mamba blocks into Llama 3.2 1B, achieving O(n) inference scaling while maintaining 34.5% accuracy on HellaSwag
  • Implemented attention transfer and Low-Rank Adaptation (LoRA) for efficient fine-tuning, demonstrating superior performance of hybrid configurations over linear-only substitutions
Machine LearningNLPPyTorch

DSA Cognitive Tutor for Binary Trees

Sep 2024 - Dec 2024

  • Engineered a Flask-based intelligent tutoring system for data structures, implementing cognitive models derived from CTA to deliver adaptive learning paths and personalized feedback using GPT.
  • Designed question generation system spanning multiple knowledge types, with dynamic difficulty adjustment based on Bloom's taxonomy proficiency tracking and spaced repetition algorithms.
  • Built interactive visualizations and code analysis tools for tree traversal concepts, achieving significant improvements with 4/5 users reaching 100% post-test scores.
FlaskPythonGPTData Structures

Offline Reinforcement Learning for Autonomous Driving

Feb 2024 - May 2024

  • Reinforcement learning project to use soft-actor critic conservative Q-learning implementation to clone human driving behavior and predict vehicle trajectories in complex roundabout scenarios.
  • Leveraged the INTERACTION dataset to preprocess real-world driving data into {State, Action, Reward, Next State} tuples for training reinforcement learning agents. Developed a custom reward function that combines positional, velocity, and collision metrics to simulate human-like driving behavior.
Reinforcement LearningPythonAutonomous Driving

Multilabel chest X-ray classification using PCA+FCN and EfficientNeXtV2

Oct 2023 - Dec 2023

  • Proposed two machine learning methods to classify chest X-rays for pneumonia detection to help doctors make quicker judgement. Dataset used was ChestX-ray14, comprising of 112,120 chest X-rays.
  • Introduced a new deep-learning efficientNet architecture called EfficientNeXtV2 - a hybrid of EfficientNet and EfficientNetV2 with a Gaussian error linear unit (GELU) as an activation function. For training, we used SGD with momentum as our optimizer and as we had an imbalanced data ratio, we used Focal loss. The model was implemented using PyTorch.
Deep LearningPyTorchComputer VisionHealthcare

Resume

GitHub Contributions

Loading contributions...