HELIX: Hierarchical Evolution via LLM-Informed eXploration 2026
Agentic evolutionary optimization for whole repositories: each candidate is a git worktree and each mutation is a tool-enabled coding session that can read, edit, run tests, and score end-to-end.
Open Multi-Agent Runtime (OMAR)
2026
A tmux-based TUI runtime for coordinating multi-agent teams: spawn deep hierarchies, mix heterogeneous backends, and debug agent behavior in one shared environment. Capable of supporting hundreds of AI agents.
LIBERO-Infinity 2026
Open-ended evaluation for robotic manipulation: Scenic-based generation creates statistically diverse scenes and perturbations to test lifelong/compositional skill generalization. Left: distractor objects are introduced. Right: the robot’s initial configuration is perturbed.
Knowing Is Not Seeing
ICLR 2026: Workshop on I Can’t Believe It’s Not Better! (ICBINB)
Benchmarks physical/spatial problem solving where models encode information they fail to surface at inference time, revealing gaps between internal knowledge and output behavior.
GRAID ICLR 2026: Workshop on Multimodal Intelligence: Next Token Prediction & Beyond; The First Workshop on Efficient Spatial Reasoning
High-fidelity data generation for spatial reasoning tasks that better train VLMs.
ScenicNL: Generating Probabilistic Scenario Programs from Natural Language
COLM 2024: Conference Paper
A deep dive into how to get LLMs to write code in languages they have never seen. As a case study, we apply it to generate scenario programs from the last 5 years of autonomous vehicle crashes in California.
Tensor Trust (Read our paper’s contribution statement)
ICLR 2024:
Spotlight Paper (Top 5%)
NeurIPS 2023:
Instruction Tuning and Instruction Following Workshop (Honorable Mention)
Robustness of Few-shot and Zero-shot Learning in Foundation Models Workshop
Adaptive Model-Based Control of Under-Actuated Legged Millirobots
This is one of the three modified kamigami robots that we built.
Implemented Online MPC, Offline MPC, and Nagabandi’s GrBAL
Email me for a copy of the unpublished paper
In person collaborations
I am extremely open to working with new research collaborators or mentoring very motivated undergraduate students. If you are doing well in most of your classes and have at least 20 hours a week to dedicate to research, feel free to email me with [Research Intern] included as a prefix in the subject line.
Karim’s Thoughts 