Jason Liang

About Me

I am an AI research scientist at Cognizant AI Lab, based in San Francisco. I received my Ph.D. in Computer Science from The University of Texas at Austin in 2018, advised by Risto Miikkulainen, and my B.S. in EECS from UC Berkeley in 2013. During my Ph.D. I developed CoDeepNEAT, an evolutionary neural-architecture-search algorithm introduced in work that has been cited over 1,500 times, and later scaled it to GPU clusters at Sentient Technologies.

My research is focused on open-ended, self-improving agentic systems and large-scale neuroevolution as a path toward artificial superintelligence, guided by two convictions: a system's founding values compound over its lifetime, and open-endedness and creativity must continue at inference time, not just during training. I architected Caesar (demo here), an autonomous deep-research agent that uses adversarial self-refinement over a dynamic knowledge graph for creative answer synthesis.

I am currently studying how starting conditions and self-modification interact and compound to shape what an agentic system builds, in a loop where the agent rewrites both its own solver and its own values against a scorer held outside the loop. On the variable radius circle packing task the agent wrote a unique solver that matched the best-known value at N=26 for under $20 of compute to first tie, with no web assistance; at N=27 the same solver beat the existing state of the art and now holds the best-known result.

Interests

• Open-Ended Search & Self-Improving Agents
• Automated AI Research & Meta-Learning
• Neuroevolution & Neural Architecture Search
• Agentic Reasoning & Inference-Time Compute
• Multi-Agent Systems & Large Language Models
• Emergent Capabilities in Foundation Models

Education

• Ph.D. in Computer Science, 2018

  The University of Texas at Austin
• B.S. in Electrical Engineering & Computer Science, 2013

  University of California, Berkeley

Featured Publications

See all publications →