00 /

News

Availability. I am applying for PhD positions in the 2026–2027 admissions cycle and am interested in research internships. Feel free to reach out; I would be glad to chat.

01 /

About

I'm an undergraduate at Johns Hopkins University majoring in Applied Mathematics and Statistics with a minor in Computer Science. I work with Prof. Daniel Khashabi in the Intelligence Amplification Lab at the Center for Language and Speech Processing, and am co-supervised by Prof. Tianmin Shu, director of the Social Cognitive AI (SCAI) Lab. I'm currently also interning with Prof. Andrea Zanette.

My research lies at the intersection of continual learning and self-improvement. I want to understand how AI agents can keep learning from experience: improving their reasoning and behavior while retaining what they have learned and remaining able to learn new tasks. I view these goals as inseparable: self-improvement supplies new experience, while continual learning makes that progress cumulative.

I study this question through reinforcement learning, reliable use of feedback, and adaptation in both an agent's context and its model parameters. My recent work explores learning to revise an agent's executable workflow, composing mechanisms for long-horizon memorization, and unifying supervised and reinforcement learning within a shared training algorithm. Together, these projects inform my goal of building agents capable of sustained online self-improvement.

02 /

Research Interests

How can language-model agents improve continually through experience? I approach this question through:

  • 01

    Continual learning: composing mechanisms to retain knowledge across sequential updates while preserving the ability to learn.

  • 02

    Self-improvement and RL: learning from feedback and experience to improve reasoning and decision-making.

  • 03

    Adaptive agents: learning to revise executable workflows and coordinating adaptation in context and model parameters.

  • 04

    Unified training: combining static data and model-generated experience within a common learning algorithm.

  • 05

    Reliable supervision: deciding which feedback to trust, including weak-to-strong generalization and self-correction.

  • 06

    Evaluation and efficiency: diverse reasoning, compute-aware inference, and learning in long-horizon environments.

03 /

Publications

* indicates equal contribution.

In submission / under review

Training Foundation Models with a Single Algorithm

Yuda Song*, Alvin Zhang*, Fahim Tajwar*, Daman Arora, Jeff Schneider, Daniel Khashabi, Ruslan Salakhutdinov, J. Andrew Bagnell, Aarti Singh, Andrea Zanette

A shared likelihood-based algorithm connects pretraining, supervised fine-tuning, and suitable RL objectives, mixing static data with model-generated experience to support retention and further learning.

arXiv

In submission / under review

Harness Learning Enables Generalizable Test-Time Adaptation

Alvin Zhang, Xuecheng Liu, Zixuan Wang, Fahim Tajwar, Daman Arora, Ruslan Salakhutdinov, Daniel Khashabi, Yuda Song, Andrea Zanette

COLM 2026 LLA Workshop · Oral

Learning to adapt an agent's executable workflow from feedback: an RL-trained proposer revises the harness on unseen tasks without updating model parameters at test time.

arXiv

In submission / under review

Continual Learning Mechanisms Compose for Long-Horizon Memorization

Zheyuan Zhang*, Alvin Zhang*, Daniel Khashabi, Tianmin Shu

Composing complementary mechanisms for preserving prior information and allocating updates improves memorization across long sequences of tasks.

arXiv · Project Page

NeurIPS 2026 · Accepted

AgentOdyssey: Open-Ended Long-Horizon Text Game Generation for Test-Time Continual Learning Agents

Zheyuan Zhang*, Zehao Wen*, Alvin Zhang, Andrew Wang, Jianwen Xie, Daniel Khashabi, Tianmin Shu

Workshop version: COLM 2026 LLA Workshop

Open-ended text-game environments for studying how agents learn and improve their strategies through extended interaction.

arXiv · Project Page

ICML 2026

Trust Functions: Near-Lossless Weak-to-Strong Generalization by Learning When to Trust the Weak Teacher

Arda Uzunoglu*, Alvin Zhang*, Daniel Khashabi

Learning which weak supervision to trust helps a stronger model learn from a weaker teacher while limiting the propagation of its errors.

arXiv · Project Page

ICML 2026

Conformal Thinking: Risk Control for Reasoning on a Compute Budget

Xi Wang*, Anushri Suresh*, Alvin Zhang*, Rishi More*, William Jurayj, Benjamin Van Durme, Mehrdad Farajtabar, Daniel Khashabi, Eric Nalisnick

Earlier version: NeurIPS 2025 Workshop on Efficient Reasoning

Risk-controlled early stopping balances reasoning quality with inference cost, asking when additional computation is worth using.

arXiv

TMLR 2026

CreativityPrism: A Cross-Domain Evaluation Framework for Large Language Model Creativity

Zhaoyi Joey Hou, Alvin Zhang, Yining Lu, Bhiman Kumar Baghel, Anneliese Brei, Ximing Lu, Meng Jiang, Faeze Brahman, Snigdha Chaturvedi, Haw-Shiuan Chang, Daniel Khashabi, Xiang Lorraine Li

A cross-domain framework for evaluating the creativity of language-model outputs alongside their correctness.

arXiv · OpenReview

NeurIPS 2025

Feedback Friction: LLMs Struggle to Fully Incorporate External Feedback

Dongwei Jiang*, Alvin Zhang*, Andrew Wang, Nicholas Andrews, Daniel Khashabi

Examining why language models fail to fully use external feedback, even when it provides much of the information needed to correct an answer.

arXiv

04 /

Blog

  • Jul 2026 The completeness cliff: language models escape wrong reasoning until it’s finished. Splice an incorrect chain of thought into a model and it recovers easily, until the trace is complete, at which point recovery collapses (pass@16 0.91→0.04). The disposition to re-examine is trained rather than a property of having a thinking mode, and a paired ablation rules out the scaffold: only the model’s own finished derivation does the damage.
  • Jul 2026 Jacobian counterexample playground. An interactive note on Alpöge’s counterexample to the Jacobian conjecture: compose shears and rescalings over the det-1 map, track an exact determinant ledger, and watch the three-point collision no composition can separate.
  • Jul 2026 Loop deeper, or adapt? Test-time training in looped transformers. Test-time training in looped / recurrent-depth transformers (Ouro): where and when a tiny RMSNorm-scale update helps, and why most of the four-shot GSM8K gain is reusable prompt calibration.
05 /

Education

Aug 2023 — Present

Johns Hopkins University  ·  Baltimore, MD

B.S. in Applied Mathematics and Statistics  ·  Minor in Computer Science
GPA: 4.0 / 4.0  ·  Dean's List every semester

Selected coursework: Honors Mathematical Statistics, Bayesian Statistics, Optimization, Time Series Analysis, Machine Learning: Deep Learning, Natural Language Processing, Self-Supervised Models, AI Agents, Honors Discrete Mathematics, Computational Cognitive Neuroscience of Vision, AI and Democracy.

06 /

Experience

May 2026 — Present

Research Intern

Carnegie Mellon University  ·  Advisor: Prof. Andrea Zanette

Leading Harness Learning, which learns to adapt agent workflows from execution feedback. Co-first author of unified training, connecting static data and reward-selected generations under a shared likelihood-based algorithm.

Aug 2024 — Present

Undergraduate Researcher

Johns Hopkins University  ·  Advisors: Prof. Daniel Khashabi (Intelligence Amplification Lab, CLSP) and Prof. Tianmin Shu (Social Cognitive AI Lab)

Studying continual learning and self-improvement through compositional mechanisms for long-horizon memorization, feedback incorporation, weak-to-strong generalization, efficient reasoning, creativity evaluation, and agent learning in open-ended environments.

Jun — Jul 2024

Software Development Engineer Intern

Caterpillar Inc.  ·  Wuxi, Jiangsu, China

Developed recurrent neural networks to simulate current and voltage in electrical motors, exploring neural-network replacements for traditional simulation pipelines.

07 /

Teaching

Aug — Dec 2026

Course Assistant  ·  EN 601.471/671 NLP Self-Supervised Models

Instructor: Prof. Daniel Khashabi

Aug — Dec 2025

Course Assistant  ·  EN 601.465 Natural Language Processing

Instructor: Prof. Jason Eisner

Aug — Dec 2024

Undergraduate TA  ·  EN 553.361 Introduction to Optimization

Instructor: Prof. Donniell Fishkind

Led TA sessions, ran problem-solving recitations, and graded homework and exams.

08 /

Skills

Languages
English  ·  Chinese (Mandarin)
Programming
Python  ·  Java  ·  MATLAB  ·  R  ·  Mathematica
ML / NLP
PyTorch  ·  Transformers  ·  VeRL  ·  vLLM  ·  SGLang
Systems
Linux  ·  HPC clusters
Areas
Machine Learning  ·  Deep Learning  ·  Reinforcement Learning