Elamur Research⌕
☰
home / authors / J. Zico Kolter

J. Zico Kolter

J. Zico Kolter works on machine learning, and their papers here span three topics: Language models & pretraining, Vision & language, and Bias, fairness & safety. A recurring thread is how neural sequence models behave under demanding conditions. One such condition is input streams from different senses that do not line up in time. Others are deliberate attempts to make a safety-tuned chatbot misbehave, and requests that a trained model drop specific knowledge. The methods include new attention architectures, gradient-guided search over tokens, and carefully controlled benchmarks.

The most cited of these works is the 2019 Multimodal Transformer for Unaligned Multimodal Language Sequences. It extends the Transformer to handle text, facial and vocal signals recorded at different rates. Rather than forcing these signals onto word boundaries, it uses directional crossmodal attention between each pair of modalities. Next comes the 2023 paper Universal and Transferable Adversarial Attacks on Aligned Language Models. It introduced Greedy Coordinate Gradient, an automated method that learns a short suffix to append to a harmful request. The suffix pushes a safety-trained model to begin its reply in agreement, after which it tends to comply.

The newest work here turns to privacy and forgetting in large language models. TOFU: A Task of Fictitious Unlearning for LLMs appeared in 2024. It builds a benchmark from invented biographies of 200 authors who do not exist, so the only source of the knowledge to be erased is known exactly. Models are fine-tuned on these profiles and then asked to forget part of them while retaining the rest. Together with the adversarial-suffix study, it moves the work from multimodal modelling toward the safety and reliability of language models.

Kolter is currently at Carnegie Mellon University in Pittsburgh. Earlier bylines include Stanford University and the Massachusetts Institute of Technology, and from 2019 onward also Robert Bosch in the United States. Co-authors on the papers here include Yao-Hung Hubert Tsai, Ruslan Salakhutdinov, Nicholas Carlini, and Zachary C. Lipton.

3 papers on this site · 2019–2024

current affiliationCarnegie Mellon University, Pittsburgh, United States

J. Zico Kolter has 3 papers on this site, in 2019–2024, most of them on Language models & pretraining, Bias, fairness & safety, Vision & language; often with Andy Zou, Avi Schwarzschild, Louis-Philippe Morency.

OpenAlex · ORCID · h-index 52 · 318 works · 18 912 citations OpenAlex, 2026-10-14

3 works on this site · 1 545 citations in our library · 5 more works in our library

The profile is the library's match of the name; two people with one name are not always told apart.

Elsewhere: Google Scholar · Semantic Scholar · dblp · ACL Anthology

Affiliations

From the bylines of their papers (OpenAlex, 2026-10-14), ORCID and Wikidata; institutions on papers are matched by machine. 4 more institutions on single bylines are not listed.

Topics

Language models & pretraining2Bias, fairness & safety1Vision & language1

Works by year

2020s

20241 paper20231 paper

2010s

20191 paper

Papers

The most cited first (OpenAlex and our library).

  1. Multimodal Transformer for Unaligned Multimodal Language Sequences (2019 · 2 170 citations)
    People communicate through words, facial expressions and tone of voice, but these streams are recorded at different rates.
  2. Universal and Transferable Adversarial Attacks on Aligned Language Models (2023 · 951 citations)
    Chatbots fine-tuned for safety are trained to refuse harmful requests. Before this paper, most ways around that refusal were handwritten jailbreaks that broke easily, and automatic prompt search had mostly failed. The authors present Greedy Coordinate Gradient (GCG), an automated attack that learns a short…
  3. TOFU: A Task of Fictitious Unlearning for LLMs (2024 · 133 citations)
    LLMs can memorize and reveal private details about people. "Unlearning" means adjusting a trained model so it forgets particular data. However, it has been hard to tell whether existing unlearning methods produce a model that behaves like one that never saw that data. TOFU (Task of Fictitious Unlearning) addresses…

Also in our library

5 more works of theirs are in our library without a page here yet; they will appear as we publish them.

Co-authors here

10 co-authors with a page, the most papers together first.

Cited by

Papers of our library cite their works 1 545 times (of the 140 884 papers whose bibliographies we read); 60 citing papers with a page here →.

This page in Markdown → https://md.research.elamur.ai/free/en/authors/j-zico-kolter/