Scholar
Akarsh Kumar
Google Scholar ID: FaM3qWEAAAAJ
Massachusetts Institute of Technology
Citations & Impact
Citations
43
H-index
3
i10-index
2
Publications
9
Co-authors
10
list available
Contact
Publications
Resume (English only)
Academic Achievements
- FER: Fractured Entangled Representation Hypothesis (arXiv 2025): Proposes that modern deep networks have disorganized internal representations; Picbreeder demonstrates more structured representations via open-ended evolution
- ASAL: Automating the Search for Artificial Life with Foundation Models (ALife Journal 2025, Best Oral at ALIFE 2025): Uses VLMs to search for target, open-ended, and diverse artificial life simulations
- Learning In-Context Decision Making with Synthetic MDPs (AutoRL @ ICML 2024): Shows generalist in-context RL agents trained solely on synthetic MDPs generalize to real-world MDPs
- GESMR: Effective Mutation Rate Adaptation through Group Elite Selection (GECCO 2022): Addresses mutation rate self-adaptation failure in genetic algorithms via group elite selection
- Physically Plausible Pose Refinement using Fully Differentiable Forces (EPIC @ CVPR 2021): Improves pose estimation accuracy through differentiable physics modeling
- Invited speaker at multiple venues including ALIFE 2025, MIT Embodied Intelligence Seminar, Detection and Emergence of Complexity Conference, etc.
Background
- Ph.D. student at MIT CSAIL, advised by Phillip Isola
- Research intern at Sakana AI, working with Yujin Tang and David Ha
- Collaborates with Ken Stanley, Jeff Clune, and Joel Lehman
- Research supported by the NSF GRFP
- Research interests include: applying principles from natural evolution and artificial life to build better AI systems, open-ended processes that indefinitely generate 'interesting' artifacts, evolving intelligence from scratch without the internet, meta-learning, reinforcement learning, automatic environment generation, and multi-agent self-play
