Doctoral Thesis Proposal - Zhuoming Chen
November 16, 2026 4:30PM—6:00PM
Location:
6115 & Zoom
-
Gates and Hillman Centers
Speaker:
ZHUOMING CHEN,
Ph.D. Student, Computer Science Department, Carnegie Mellon University
https://dreaming-panda.github.io/
AI agents are becoming the interface by which people write software, learn about the world, and make decisions, yet both access to capable agents and trust in their outputs remain concentrated among a few organizations. Serving and training agents is dominated by memory-bound decoding, attention over long contexts, and RL rollouts, and verifying that an agent's output is correct, and that it serves its user rather than whoever controls the agent, often costs more than generating it.
This thesis argues that AI agents can benefit most people only if both access and trust are democratized, which requires making generation and verification efficient without weakening their guarantees. Five completed projects democratize access. Sequoia makes speculative decoding scalable and hardware-aware, serving exact Llama2-70B inference at 0.56 s/token on a single L40 GPU. MagicPIG replaces TopK attention with locality-sensitive-hashing sampling that carries guarantees, reaching 54 ms decoding latency at a 96K-token context on one RTX 4090. Vortex makes sparse attention programmable, enabling AI agents to generate algorithms with up to 3.46x the throughput of full attention. Kinetics shows that attention dominates the cost of test-time scaling, and that sparse attention improves AIME accuracy by over 60 points in low-cost regimes. Jackpot trains from cheaper rollout models with budgeted rejection sampling while matching on-policy RL.
Benchmarks establish trust in capability: in ServerBench, agent-built servers reach up to 4.87x the value per GPU dollar of vLLM and SGLang, yet only 22 of 33 runs pass independent verification. The proposed work will make generative verifiers, in which agents generate proofs for programs and evidence for factual claims that a checker and independent reviewers decide, efficient and scalable, by drafting with efficient models before verification and by reusing indexed verified results.
Thesis Committee
Beidi Chen (Co-chair)
Zhihao Jia (Co-chair)
Limin Jia
Niklas Stefan Nolte (Anthropic)
In-person & Zoom
Contact
Matt Stewart