Zhichao Xu

Zhichao Xu

About me

I am currently a ML Researcher at Apple Foundation Model. I try to understand language models' mechanisms, behaviors and how to better utilize them.

My research focuses on evaluating and improving language models and agents, especially on long-horizon and hard-to-verify tasks.

I completed my Ph.D. study from UtahNLP, advised by Prof. Vivek Srikumar and Prof. Bei Wang.

I am open to collaborations, feel free to reach out to me at .

@Twitter @BlueSky @Google Scholar

Research

Two current questions guide my work:

  1. How can we construct reliable and challenging benchmarks that measure the capabilities of LLMs and agents?
  2. How can we improve learning using reinforcement learning, across long-horizon agentic tasks and RL from human feedback (RLHF)?

Evaluation

How can we construct reliable and challenging benchmarks that measure the capabilities of LLMs and agents?

Beyond Perplexity: Multi-dimensional Safety Evaluation of LLM Compression

Multi-dimensional Evaluation of Empathetic Dialogue Utterances

Reinforcement Learning

How can we improve learning using RL? I study two connected directions:

Long-horizon agentic tasks

How can agents learn effective strategies over long horizons? My work studies credit assignment, tool use, and learning reusable skills.

SALT: Step-level Advantage Assignment for Long-horizon Agents via Trajectory Graph

Reinforcement Learning for Self-Improving Agent with Skill Library

Learning to Ideate for Machine Learning Engineering Agents

RLHF

How can we improve learning from human feedback as models become more capable? I am exploring how evaluation, data collection, and reward learning can evolve together for hard-to-verify tasks.

Beyond Correctness: Rewarding Faithful Reasoning in Retrieval-Augmented Generation

PERSONAJUDGE: Simulating Individual Human Preference Judgments with Human-Centered Demonstration Data

Efficient LLM Systems

How can we improve the efficiency of LLM systems? My work explores speculative decoding and prompt routing to balance quality, latency, and cost.

APEX: Speculate Smarter, not Deeper

IPR: Intelligent Prompt Routing with User-Controlled Quality-Cost Trade-offs

Retrieval Modeling & Efficiency

How can we build effective retrievers and rerankers with efficient training and inference? My work spans sparse retrieval, model architectures, and scalable training systems.

LACONIC: Dense-Level Effectiveness for Scalable Sparse Retrieval via a Two-Phase Training Curriculum

CSPLADE: Learned Sparse Retrieval with Causal Language Models

Tevatron Meets Megatron: Expert-Parallel LLM Reranker Training on an Academic Budget

Tevatron-Elastic: A Unified Abstraction for Training Elastic Retrievers and Rerankers

All publications, by year ↗

Recent News

Education

Internships

Academic Services

Mentoring

Teaching