|
Albert Gong I'm a fourth-year PhD student in computer science at Cornell, where I'm fortunate to be advised by Raaz Dwivedi and Kilian Q. Weinberger. I'm currently working on using distribution compression (a.k.a. "thinning") to speed up inference and training of large-language models and benchmarking AI agents. Previously, I was an undergrad at Yale, where I had the privilege of working with Andre Wibisono, Zhong Shao, and Cormac O'Dea. Email / Google Scholar / LinkedIn / Github |
|
Research* = equal contribution |
|
PhantomEnvironments: Training LLM Agents in Fictional Worlds
Anmol Kabra, Swathi Saravana Selvam, Albert Gong, Chao Wan, Christian Belardi, Dongyoung Go, Katie Z. Luo, Kilian Q. Weinberger arXiv preprint, 2026 (Oral presentation at COLM 2026 Workshop on Learning from Situated and Embodied Interaction) Code / arXiv Tl;dr—LLMs trained with RL in rule-generated fictional worlds, built at zero marginal cost and without any LLM, become search agents that transfer to real-world multi-hop search benchmarks. |
|
ORCA-bench: How Ready Are Language Model Agents for Oncall?
Albert Gong, Kyuseong Choi, Abhineet Agarwal, Jason Schechner, Ryan Huang, Raj Agrawal, Anish Agarwal, Raaz Dwivedi arXiv preprint, 2026 (Oral presentation at COLM 2026 Workshop on Context Beyond the Window) Website / Dataset / Code / arXiv / Poster Tl;dr—Introduced a benchmark of 1,079 root cause analysis tasks over a live OpenTelemetry-instrumented microservice system, where the best frontier agent reaches only 25.3% accuracy on realistic incident reports. |
|
Express Language Modeling
Albert Gong, Annabelle Michael Carrell, Raaz Dwivedi, Lester Mackey NeurIPS, 2026 (Oral presentations at COLM 2026 workshops: MOSS, Context Beyond the Window, and Efficient Reasoning; Best Paper Award at Efficient Reasoning) Code / arXiv / Slides Tl;dr—Express provides state-of-the-art causal attention guarantees, an efficient I/O-aware Triton implementation, and practical improvements for prefill, cache compression, and decoding. |
|
Learning from Synthetic Data Improves Multi-hop Reasoning
Anmol Kabra, Yilun Yin, Albert Gong, Kamilė Stankevičiūtė, Dongyoung Go, Johann Lee, Katie Z. Luo, Carla P. Gomes, Kilian Q. Weinberger ICLR, 2026 Code / arXiv Tl;dr—RL fine-tuning LLMs on synthetic data improves real-world multi-hop reasoning by teaching knowledge composition skills. |
|
π-CoT: Prolog-Initialized Chain-of-Thought Prompting for Multi-Hop Question-Answering
Chao Wan*, Albert Gong*, Mihir Mishra, Carl-Leander Henneking, Claas Beger, Kilian Q. Weinberger TMLR, 2026 Code (coming soon) / arXiv Tl;dr—π-CoT translates multi-hop questions into Prolog queries, resolves them as single-hop sub-queries, and uses the results to initialize Chain-of-Thought, outperforming standard RAG and in-context CoT. |
|
N2: A Unified Python Package and Test Bench for Nearest Neighbor-Based Matrix Completion
Caleb Chin, Aashish Khubchandani, Harshvardhan Maskara, Kyuseong Choi, Jacob Feitelberg, Albert Gong, Manit Paul, Tathagata Sadhukhan, Anish Agarwal, Raaz Dwivedi arXiv preprint, 2025 Code / arXiv / Poster (CODEML Workshop) Tl;dr—Introduced the N2 package and N2-Bench test bench for nearest neighbor-based matrix completion. |
|
PhantomWiki: On-Demand Datasets for Reasoning and Retrieval Evaluation
Albert Gong*, Kamilė Stankevičiūtė*, Chao Wan*, Anmol Kabra*, Raphael Thesmar, Johann Lee, JT Klenke, Carla P. Gomes, Kilian Q. Weinberger ICML, 2025 (Oral presentation at ICML Workshop on Long Context Foundation Models) Code / arXiv / Poster / Slides Tl;dr—Created a framework to automatically generate both the document corpus and question-answer pairs for benchmarking RAG and agentic workflows. |
|
Low-Rank Thinning
Annabelle Michael Carrell, Albert Gong, Abhishek Shetty, Raaz Dwivedi, Lester Mackey ICML, 2025 Code (see below) / arXiv Tl;dr—Developed new analysis of thinning algorithms that adapts to low-rank structures, enabling faster dot-product attention in Transformers (Thinformer), stochastic gradient descent (KH-SGD), and deep kernel hypothesis testing (DeepCTT). |
|
Supervised Kernel Thinning
Albert Gong, Kyuseong Choi, Raaz Dwivedi NeurIPS, 2024 Code / arXiv / Poster / Slides / Video Tl;dr—Used distribution compression to speed up kernel smoothing and kernel ridge regression. |
|
Source code adapted from Jon Barron's website. |
