Research Fair Fall 2026


Blue Computer Science "CS@CU" logo with Columbia crown

Research Fair Fall 2026

The Fall 2026 Research Fair will be held on Thursday, September 10th, and Friday, September 11th in Carlton Commons. 12:00 to 14:00

This is an opportunity to meet faculty and Ph.D. Students working in areas of interest to you, and possibly work on these projects.

Please read these carefully! (You are less likely to be hired if you haven’t taken the time to read about the project.) There will be a couple of Zoom/Google Meet sessions available for students who cannot attend in person – see below for all details.

More research projects will be added up until the day of the fair, so keep checking back.


Faculty/Lab: Graphics Imaging & Light Measurement Lab (GILMLab) / Prof. Corey Toler-Franklin

Brief Research Project Description: https://coreytolerfranklin.com/gilmlab/ — Click here to apply: https://forms.gle/Rhwb88ULFdmCc4wn6. Available for Independent Study 1-3 Credits (COMS E6901, COMS E6902, COMS X3099, COMS W3998, COMS W4901).

AI for Computer Graphics: Physics-based Material Simulation for Real-time Rendering, Animation and Gaming. Seeking graduate students with experience in graphics and/or physics and machine learning for projects focused on physics-based photorealistic material simulation for computer graphics.

AI for Cancer Detection: Identifying Cancer Cells and Their Biomarker Expressions. Cell quantitation techniques are used in biomedical research to diagnose and treat cancer. Current quantitation methods are subjective and based mostly on visual impressions of stained tissue samples. This time-consuming process causes delays in therapy that reduce the effectiveness of treatments and add to patient distress. Our lab is developing computational algorithms that use deep learning to model changes in protein structure from multispectral observations of tissue. Once computed, the model can be applied to any tissue observation to detect a variety of protein markers without further spectral analysis. The deep learning model will be quantitatively evaluated on a learning dataset of cancer tumors.

AI for Neuroscience: Deep Learning for Diagnosing and Treating Neurological Disorders. Advances in biomedical research are based upon two foundations: preclinical studies using animal models and clinical trials with human subjects. However, translation from basic animal research to treatment of human conditions is not straightforward. Preclinical studies in animals may not replicate across labs, and a multitude of preclinical leads have failed in human clinical trials. Inspired by recent generative models for semi-supervised action recognition and probabilistic 3D human motion prediction, we are developing a system that learns animal behavior from unstructured video frames without labels or annotations. Our approach extends a generative model to incorporate adversarial inference, transformer-based self-attention modules, and zero-shot learning.

AI for Quantum Physics & Appearance Modeling: Quantum Level Optical Interactions in Complex Materials. The wavelength dependence of fluorescence is used in the physical sciences for material analysis and identification. However, fluorescent measurement techniques like mass spectrometry are expensive and often destructive. Empirical measurement systems effectively simulate material appearance but are time-consuming, requiring densely sampled measurements. Leveraging GPU processing and shared supercomputing resources, we develop deep learning models that incorporate principles from quantum mechanics theory to solve large-scale many-body problems in physics for non-invasive identification of complex proteinaceous materials.

AI for Multimodal Data & Document Analysis: Deciphering Findings from the Tulsa Race Massacre Death Investigation. The Tulsa Race Massacre (1921) destroyed a flourishing Black community and left up to 300 people dead. More than 1000 homes were burned and destroyed. Efforts are underway to locate the bodies of victims and reconstruct lost historical information for their families. Collaborating with the Tulsa forensics team, we are developing spectral imaging methods (on-site) for deciphering information on eroded materials (stone engravings, rusted metal, and deteriorated wood markings), novel multimodal transformer networks to associate recovered information on gravestones with death certificates and geographical information from public records, and graph networks for reconstructing social networks.

Required/preferred prerequisites and qualifications: COMS W3157, Python and/or C/C++. High priority to students with computer graphics and/or physics experience. Machine learning experience is a plus.


Faculty/Lab: Computer Graphics & User Interfaces Lab | Prof Steve Feiner

Brief Research Project Description: The Computer Graphics and User Interfaces Lab (Prof. Feiner, PI) does research in the design of 3D and 2D user interfaces, including eXtended Reality (XR)—encompassing augmented reality (AR) and virtual reality (VR)—and mobile and wearable systems, for people interacting individually and together, indoors and outdoors. We use a range of displays and devices to develop novel interaction and visualization techniques, which we evaluate in prototype applications. Multidisciplinary projects potentially involve working with faculty and students in other schools and departments, from dentistry and medicine to earth and environmental sciences.

Required/preferred prerequisites and qualifications: We’re looking for students who have done excellent work in one or more of the following courses or their equivalents elsewhere: COMS W4160 (Computer graphics), COMS W4170 (User interface design), COMS W4172 (3D user interfaces and augmented reality), and COMS E6173 (Topics in VR & AR), and who have software design and development expertise. Knowledge of robotics and vision is a plus for some projects. For those projects involving 3D user interfaces, we’re especially interested in students with Unity experience.


Faculty/Lab: Itsik Pe’er

Brief Research Project Description: Multiple projects across multiple levels, most involving representation learning, hyperbolic embeddings, and matryoshka learning, with applications in single cell genomics and microbiome analysis. Includes development of computational support for non-invasive prenatal genetic diagnostics using Hidden Markov Models and transformers. Available programming projects for sophomore–junior year students interested in women’s health, as well as graduate-level representation learning projects.

Required/preferred prerequisites and qualifications: Basic probability and statistics, being an independent programmer in Python, basic genetics.


Faculty/Lab: Ahmer Arif

Brief Research Project Description: This new project will look at the Pravda network—a coordinated pro-Kremlin operation that automatically republishes material from Russian media and Telegram channels across hundreds of news-like websites, languages, and countries—to explore how this can create the appearance of independent corroboration for AI assistants. Using CheckFirst’s open dataset, we could trace how individual stories are reproduced across the network and ask web-connected assistants about claims that vary in how widely they’re repeated and how much credible news coverage exists. We’d examine whether assistants cite, repeat, contextualize, or recognize the shared provenance of these stories, helping us understand how coordinated publishing operations might shape the evidence AI systems assemble and present to users.

Student collaborators will be expected to attend regular research meetings, help conceptualize the study, and help clean and analyze the CheckFirst dataset using a variety of methods (such as reading the stories or doing some basic clustering).

Required/preferred prerequisites and qualifications: An interest in qualitative and mixed-methods research and online disinformation. Sophomore and above. Lab site: http://ahmerarif.com


Faculty/Lab: Formal Methods and Reasoning Group (FoRG) | Prof. Mark Santolucito & MS Student Christian Scaff

Brief Research Project Description: Manhattan Reasoning is motivated by the rise of domain-specific architectures (DSAs). While domain-specific computing promises progress in scientific areas from artificial intelligence to chemistry, the barrier between software and hardware continues to prevent domain experts from developing novel computing architectures best fit for their problem space. We are actively working on closing that gap in two ways: (1) lowering the barrier to entry for FPGAs by providing cloud infrastructure backed by free and fast open-source toolchains, and (2) improving agentic reasoning capabilities around RTL, the design abstraction for digital circuits. We envision a cloud computing platform where domain experts with little knowledge of hardware design can experiment in an agentic manner with physical FPGAs in the loop. More information: https://docs.manhattanreasoning.com

We are actively looking for students passionate about software-hardware co-design to develop infrastructure for the Fall 2026 academic term, across three roles:

Position One: Developer Experience Engineer — CLI and Python SDK for the research tool (docs). Required: strong Python, Git/GitHub Actions. Preferred: CLI/Python packaging (PyPI, pip), REST/WebSocket APIs, Docker, developer documentation, experience building for LLM agents.

Position Two: Systems Engineer — cloud infrastructure handling the build process from RTL submission to bitstream generation, concurrent jobs, authentication, and resource states. Required: strong Python, Git/GitHub Actions. Preferred: FastAPI/Pydantic/Starlette, Redis/Postgres, Docker, AWS (ECS Fargate, S3), Tailscale, distributed systems (job queues, worker processes, resource pools).

See here (https://forms.gle/ueLDqv5mL1w8rkZF7) for application and timeline.


Faculty/Lab: Prof. Gil Zussman, Dr. Jhonatan Tavori, Alon Levin

Brief Research Project Description: Project #1 — Real-Time LLM Inference in Distributed and Wireless Mobile-Edge Networks: AI systems such as ChatGPT and other large language models (LLMs) require substantial computational resources and processing effort to generate responses, due to their architectures and parameter sizes. When deploying models for inference in resource-constrained edge environments, computation is often split across multiple edge devices, as no single device has sufficient memory or processing power to execute the entire model. This project explores adaptive quantization and compression techniques to make distributed AI inference faster and more efficient on edge devices, optimizing how intermediate data is shared across AI nodes to enable real-time AI applications.

Project #2 — Urban Spectrum Measurement: The wireless spectrum in urban environments, such as New York City, is constantly in use by communications providers, emergency services, commercial aviation, and other sources. This project runs a long-term, wideband, multi-sector measurement campaign using software-defined wireless infrastructure hosted on the COSMOS testbed to characterize how, when, and by whom spectrum is used on a typical day-to-day basis.

Required/preferred prerequisites and qualifications:
Project #1 — Strong programming skills in Python and experience with deep learning frameworks (PyTorch or TensorFlow); familiarity with LLM architectures and inference concepts (transformers, attention, token generation, KV caching); knowledge of distributed systems, edge computing, or wireless networks is an advantage; ability to read papers and implement new ideas; strong analytical and teamwork skills.
Project #2 — Strong C++ and Python skills required; comfortable in a Linux command-line environment with Git; data analysis skills with large datasets required; background in wireless communications and DSP preferred; familiarity with GNU Radio preferred but not necessary.
Contact: https://wimnet.ee.columbia.edu (j.tavori@columbia.edu, asl2228@columbia.edu)

Please apply through the WiMNet Lab Research Applications form: https://forms.gle/3vmKpxQTYQXcfcwE9


Faculty/Lab: Student: Alex Mathai, Prof. Baishakhi Ray and Prof. Junfeng Yang

Brief Research Project Description: 1. Self-cleaning LLMs — Training LLMs to predict clean and compact functional code, improving on techniques so LLMs predict as little AI slop as possible. This builds on this related paper.

2. Techniques to formally verify AI-generated code — LLM-generated patches for the security domain often contain subtle coding errors and flaws (e.g., off-by-1 index errors). We aim to develop Code-Verify-Repeat loops for agentic workflows and autonomous agents so that there are stronger guarantees for AI-generated code in security-sensitive domains.

Required/preferred prerequisites and qualifications: Research experience in AI for Software Engineering. PhD students. Contact: Alex Mathai (https://alexmathai.com)


Faculty/Lab: Julia Hirschberg

Brief Research Project Description: Project 1: Probing multilingual speech encoders’ representations of code-switched prosodic profiles. Project 2: Steering an instruction-tuned multilingual LLM backbone combined with a TTS component in order to generate naturalistic code-switched speech.

Required/preferred prerequisites and qualifications: Able to contribute at least two semesters of research, ideally three; strong background in machine learning and interpretability; interest in multilingual speech; strong independent coding ability; fluency in more than one language preferred.


Faculty/Lab: Zishen Wan

Brief Research Project Description: AI × Computing Lab (Wan Lab): Research in Computer Architecture, Systems, and Chips for AI. Our group develops new computing systems and hardware for emerging AI workloads and explores how AI agents can help design computing systems themselves. Undergraduate research projects may include: (1) efficient systems and hardware for physical/embodied AI, vision-language-action models, and world models; (2) systems optimization for LLM/VLM agents and multi-agent workloads; (3) GPU, edge, FPGA, and accelerator design for AI; and (4) AI/LLM agents for computer architecture, RTL generation and verification, compiler optimization, and chip design automation.

Students will work closely with the research team to read relevant papers, build or extend research infrastructure, implement and evaluate new ideas, run experiments on AI/system/hardware platforms, analyze results, and present findings, with opportunities to contribute to research publications. Lab: https://wan-research-group.github.io/

Required/preferred prerequisites and qualifications: Strong interest in computer systems, computer architecture, machine learning, robotics, or digital hardware/VLSI. Experience with one or more of Python/PyTorch, C/C++, CUDA/Triton, FPGA/RTL (Verilog/SystemVerilog), computer architecture, or ML/robotics frameworks is preferred but not required. Prior research experience is not required. Contact: Zishen Wan / zw3306@columbia.edu / https://zishenwan.github.io/


Faculty/Lab: Prof Venkat Venkatasubramanian and his PhD Students

Brief Research Project Description: Sign up: Link here.

1) RL for Emergent Collective Behavior

Statistical teleodynamics combines game theory and statistical thermodynamics to explain how large-scale order emerges from the local interactions of individual agents. In previous work, this framework captured the self-assembly and equilibrium of bird flocks: each bird acts to improve its utility, and flocking emerges naturally from those decisions. Using the same utility as the reward, we train RL agents that see only their nearby neighbors and must work out how to move on their own. If they rediscover flocking, we gain a more realistic picture of how birds actually behave, along with a method that extends to other collective systems where no clean equation exists — including groups of interacting LLM agents.

Requirements: Reinforcement Learning, Game Theory

2) Communication Topology and Control in Agent Swarms

We extend the same framework mentioned in Project #1 from physical space to information space. Prior work showed that a swarm’s communication topology largely determines whether it finds a global optimum. We rewrite the flocking utility so that each agent weighs the value of its neighbors’ information against the cost of information overload, and test whether the theory predicts which topologies search best. We then add random loss of agents and a coordinator that assigns tasks, and study when central control helps or hurts decentralized search. A recent incident in which ~1,200 LLM agents self-organized to collectively attack Hugging Face serves as the case study. The aim is to predict when agent collectives succeed at a shared goal, and which interventions most reliably stop them.

Requirements: LLM agent frameworks, Game Theory, Parameter Optimization

3) Neuro-Symbolic AI for Stochastic Financial Model Discovery

Stochastic differential equations (SDEs) describe systems where randomness drives behavior as much as the underlying trend. Conventional ML models can forecast prices, but offer no mechanistic explanation of what drives them. This project builds a neuro-symbolic framework to discover SDE models directly from financial time series.

Requirements: Stochastic Calculus, Deep Learning

Sanjeev (sn3130@columbia.edu)

4) Pattern Interpretability in Neural Networks: (1 Open Position)

It is known that each layer in a fully connected neural network contributes incrementally by reducing the complexly distributed data points into linearly separable points, allowing the last layer to simply draw the boundary for classification. Previous analysis by our group has shown that each layer of these fully trained Neural Network models exhibits a Log-Normal trend in its neuronal weights. 

Realizing that this weight distribution represents the perfectly trained ideal state, this project aims to accomplish the following:

  1. Use the recently found trends in neuronal weights as weight initialization to potentially reduce the training time by a huge factor, and have the ability to use less data to train.
  2. Understand the ability of a neural layer to reduce complexly shaped data into linearly separable data points. This can be compared with matrix rotation or operation. 
  3. The third aim of this project involves analysing the topological aspects of this lognormal initialization by understanding the Betti number to analyse the data resolving capabilities of these layers.
  4. Finally, the most interesting part is to build a game-theoretic framework that helps us understand how each layer in a bigger model helps dissolve data points incrementally. 

Paper: https://www.sciencedirect.com/science/article/pii/S0098135424003260

Prerequisites: Python, Statistics, Modeling and Reasoning ability, Experience or interest in internal study of Neural Networks

5)  Dynamics of tokens and the theory of LLMs: (1 Open Position)

By dividing the dynamics of LLMs into three major categories: training, planning, and inference, we analyse how a giant trained model like an LLM works. By analyzing the below-mentioned important questions and with some newly discovered dynamics of the tokens, we are building a game-theoretic framework that helps us understand how smaller agents like tokens build to form an emergent phenomenon or other knowledge. 

The important questions that we ask are the following:

  1. Does the LLM plan before prediction? (LLM planning)
  2. How does the LLM hallucinate; is it an imperfection in planning? (LLM inference)

6) Risk Analysis in Chemicals Using Reaction Pathways and Thermodynamic Properties (1 Open Position)

Chemical reactions are complex and often highly variable processes. Although it is difficult to predict the exact next step of a chemical reaction, this method aims to guide a downstream LLM by providing relevant domain-specific information. This involves identifying relevant databases and extracting their information into an ontology or knowledge graph. This will serve as a domain-information repository that adapts to the input and returns relevant information, similar to GraphRAG. (Example database: the DIPPR database for vapor–liquid equilibrium data.)

The downstream risks include deviations in the process resulting from changes in process conditions, process materials, and other relevant factors. The goal of this research is to create a pipeline that takes chemicals and their operating procedures as input and provides an LLM with relevant physical, thermodynamic, and kinetic properties in the most effective way possible for predicting hazardous reaction pathways.

Prerequisites: LLMs, Python, GraphRAG, Risk analysis

Collin (cjs2301@columbia.edu)

7) Neuro-Symbolic Systems for Analysis of Comparative Omics

There is an emerging picture of diseases as perturbations to complex biological networks. Leveraging large databases of semantic knowledge covering biochemical pathways, molecular functions, and protein structure, this project seeks to generate natural-language summaries of system-level differences between subjects inferred from high-throughput omics data.

Requirements: Causal Inference, Statistics, LLMs

8) AlphaFold as an Embedding Model

AlphaFold is a model trained to predict three-dimensional protein structure from amino acid sequences. In this process of mapping sequence to structure, high-dimensional vector representations of each amino acid in the sequence and amino acid pairs are iteratively generated. This project is concerned with identifying alternative tasks that can be performed using representation vectors, including, but not limited to, prediction of binding motifs, evolutionary relationships between folds, and identification of regions susceptible to conformational changes.

Requirements: ML, Computational Biology, Protein Structure

Collin & Vincent (2 to 4 Open Positions)

9) LLM as an emergent semantic field

LLMs are ubiquitous, yet we still lack a broadly accepted theory for how they represent knowledge, reason, generalize, and organize information internally. We propose that a trained LLM admits a coarse-grained description as an emergent semantic field. In this space, meanings compete with each other in a game-theoretic way. The resulting equilibrium is then tilted by external prompts and context. In this project, we aim to measure this field. More precisely, we want to:

  • define and measure the field of meanings for open-weight models (Pythia, Olmo)
  • measure how LLM answers vary when repeating the same prompt or instruction (replica) and compare it to random points

Requirements: Deep Learning, LLM, measure theory

Vincent (vv2435@columbia.edu)

10) Detection of insider trading on prediction markets

Prediction markets such as Kalshi and Polymarket have gained significant traction in the past few years. In the last year, concerns have grown about the existence of insider trading on these platforms. Insider traders undermine the fairness of the markets and impede the market from reaching the “wisdom of the crowd”. In this project, you will develop data science algorithms to detect, classify, and identify insider traders based on trading behavior.

Requirement: Data science


Faculty/Lab: Xia Zhou, with PhD student Ziang Ren

Brief Research Project Description: 3D scene-level reconstruction under challenging imaging conditions. Students should attend weekly meetings, perform data collection, design and run experiments, and replicate existing benchmarks.

Required/preferred prerequisites and qualifications: Computer vision, 3D Gaussian splatting, 3D modeling, deep learning. Computer Vision and Deep Learning coursework preferred.


Faculty/Lab: Prof Junfeng Yang, Prof Baishakhi Ray, with PhD student Hailie Mitchell

Brief Research Project Description: We are researching how to validate agent-generated patches, especially for security vulnerabilities where patch correctness is crucial. LLMs/agents are often evaluated on benchmarks that use validation testsuites to determine patch correctness — but this evaluation method is known to produce overfitting patches: patches that pass all validation tests but fail on other inputs or violate intended program behavior. As LLM capabilities improve, agents have also been shown to perform evaluation hacking, exploiting the evaluation procedure rather than addressing the underlying vulnerability. To develop and test stronger patch validation methods, we want to improve existing benchmarks (e.g., SEC-Bench, CyberGym-E2E). Students will help set up, run, and strengthen such benchmarks.

If interested, please fill out this application form: https://forms.gle/UfWR6x5n1aKjg65x7

Required/preferred prerequisites and qualifications: Strong debugging skills, experience in C/C++ required. Previous benchmarking experience is ideal, but not required.


Faculty/Lab: Computer-Enabled Abilities Laboratory (CEAL) | Prof. Brian A. Smith

Brief Research Project Description: CEAL will have three research assistant (RA) positions open for two different projects in Fall 2026. To express interest, complete this Google Form. Open-door lab meetings welcome drop-ins; join the CEAL Circle Google Group for calendar invites.

Project 1: Exploring Airflow as a New Game Interaction Modality — CEAL is collaborating with Sony Interactive Entertainment (PlayStation) to explore new gameplay and accessibility possibilities enabled by 5.5 Sense, a novel 360° airflow display (demo video). The project investigates how airflow can enhance spatial perception and navigation, including communicating directions, destinations, and hazards in accessibility contexts, as well as new game mechanics and immersion. Students will prototype ideas in Unity and help develop a game demo.

Project 2: Increasing Replayability in Educational Games — CEAL is developing an educational game to examine how to integrate educational content into game mechanics (research through design) and to propose a new “showroom” design methodology adapted for game design. Recruiting a gameplay programmer and a musician/sound designer.

Project 3: Audio-Haptic Artwork Exploration for Blind and Low-Vision Users — Developing an accessible image-exploration system that combines audio with electrotactile feedback. The student will help design, build, test, and document the hardware prototypes needed for reliable real-time haptic interaction.

Required/preferred prerequisites and qualifications:
Project 1: Interest in game design/development and prototyping; experience with Unity (or willingness to learn); strong communication and organizational skills; prior interest in accessibility, HCI, UI design, haptics, or novel interaction technologies.
Project 2 (Gameplay Programmer): Organized and detail-oriented; strong communication skills; Godot/GDScript/Git experience and a game dev portfolio are a plus; leadership ambition welcome.
Project 2 (Composer/Sound Designer): A portfolio of music composition/sound design; interest in video games; organized and detail-oriented; strong communication skills.
Project 3: Electronics and hardware prototyping; embedded systems/microcontroller programming; haptic/electrotactile/wearable-interface development; sensor integration and hand/finger tracking; circuit or PCB design; CAD/fabrication/rapid prototyping; Python, C, or C++; interest in accessibility and HCI; ~10 hrs/week. Prior electrotactile experience helpful but not required.
Lab: https://ceal.cs.columbia.edu/joinCEAL/


Faculty/Lab: Asaf Cidon and Junfeng Yang, with PhD student Haoda Wang

Brief Research Project Description: The collapsing cost of space launches has disrupted the way satellites are deployed, shifting the industry from a model of a few expensive fault-tolerant high-orbit satellites to arrays of commodity low-cost SmallSats in low-Earth orbit. However, satellite software hasn’t kept up with the hardware trends, and missions are still using the ad-hoc flight software infrastructure built for expensive one-off missions in high-altitude orbits, even as constellations of tens or hundreds of SmallSats come online. This motivates new approaches rooted in verification and language design. Students will develop and maintain novel software systems that will fly in real-world space missions.

Required/preferred prerequisites and qualifications: Experience with operating systems, compilers, or Verilog. Contact: Haoda Wang / haoda.wang@columbia.edu / h313.info


Faculty/Lab: Oculomics: The Eye + AI as a Window to Systemic Health, PI: Kaveri A. Thakoor, Artificial Intelligence for Vision Science (AI4VS) Lab

Brief Research Project Description: Recent findings support the ancient saying “the eyes are the window to the soul”; not only are eye gaze and pupil size linked to internal states of attention and intention, but vascular information in retinal fundus images has been found to carry evidence of cardiovascular disease risk, diabetes, jaundice, and even kidney function. Because the eye is the only organ in which one can directly view microvascular circulation, our eyes provide access to the internal state of the brain, heart, kidneys, and potentially other organs. This project employs advanced AI techniques — including continual learning, sparse autoencoders, and multimodal representation learning via joint embedding predictive architecture (JEPA) — to determine the extent to which the eye + AI can elucidate systemic changes in the body by uncovering connections between retinal images and imaging modalities/sensor data across cardiac CT/CTA, cardiac PET, blood glucose measurements, and MRI. We leverage datasets like AI-READI (2000+ patients with paired retinal imaging and lab/clinical data) to correlate retinal imaging with glycemic variability (eye-to-diabetes), portable fundus imaging from patients undergoing PET stress testing at Columbia to predict cardiac risk (eye-to-heart), and unpaired OCTA/MRI data with common dementia-state labels via self-supervised learning (eye-to-brain).

Required/preferred prerequisites and qualifications: Open to current MS students in CS or senior undergraduates in CS. Looking for students who are interested and self-motivated to learn through self-directed or guided literature review about state-of-the-art deep learning models (transformers, ViTs, diffusion models, foundation models, agentic AI frameworks). Prior PyTorch experience implementing deep learning models preferred. Previous knowledge or experience in representation learning, self-supervised learning, and/or continual learning is a plus.


Faculty/Lab: Kriste Krstovski

Brief Research Project Description: This project explores transformer-based architectures for identifying whether a patient has metastatic prostate cancer using unstructured text from electronic health records. We frame the task as a supervised binary text classification problem and compare different transformer models, representation strategies, and training paradigms to assess their impact on classification accuracy and robustness. An additional focus is on studying how prompt design and prompt-based strategies influence the reliability and performance of large language models on this classification task.

Required/preferred prerequisites and qualifications: Strong programming skills in Python and prior experience with deep learning methods for NLP, including hands-on experience training or fine-tuning transformer-based models. Familiarity with probabilistic modeling through coursework or research experience is expected; prior experience with latent variable models for text is a strong plus.


Faculty/Lab: Prof. Zhuo Zhang

Brief Research Project Description: This project aims to design and generate realistic reverse-engineering challenges for a benchmark that evaluates AI agents’ cybersecurity capabilities in binary-only settings. The challenges will cover representative reverse-engineering tasks and progressively incorporate program complexity, compiler transformations, and anti-analysis or obfuscation techniques. The resulting benchmark will study how effectively AI agents can understand binary programs, recover program semantics, identify security-relevant behaviors, and solve practical reverse-engineering problems without access to source code. The student will identify and curate codebases (at least 10K lines) not available in existing public benchmark datasets, modify them into deterministic, automatically verifiable reverse-engineering tasks, and design compiler- and binary-level obfuscation techniques to systematically increase difficulty and evaluate agent robustness.

Required/preferred prerequisites and qualifications: Strong interest in AI, systems, and cybersecurity. Experience with CTF competitions (especially reverse engineering or pwn challenges) is highly preferred. Familiarity with binary analysis, assembly, compilers, program analysis, and common obfuscation/anti-analysis techniques is valuable. Strong C/C++ skills and experience with tools such as Ghidra, IDA Pro, or Binary Ninja are preferred. Qualifications assessed via CV and interview. Contact: Zhuo Zhang / zz@cs.columbia.edu / https://zzhang.xyz/


Faculty/Lab: Richard Zemel

Brief Research Project Description: Benchmarking biologically plausible learning procedures (Hebbian learning, forward-forward algorithm, feedback alignment, etc.) in hypermodal settings. The work will involve conducting literature reviews, prototyping algorithms in JAX and/or PyTorch, and onboarding datasets.

Required/preferred prerequisites and qualifications: Machine learning, deep learning, and programming experience required; interest in neuroscience. Completion of at least one machine learning class (with Blei, Hsu, Zemel, etc.), experience implementing machine learning algorithms in JAX or PyTorch. toddmorrill.github.io


Faculty/Lab: Professor Kathleen McKeown, with PhD student Nicholas Deas

Brief Research Project Description: Professor McKeown’s lab is looking for undergraduate and master’s researchers to assist on several NLP projects:

1. Understanding Attitudes in Under-represented Dialects — Studying LLMs’ understanding of nuanced emotions and political perspectives in natural language, focusing in part on underrepresented dialects of English (e.g., African American Language) and other languages (currently including Persian, Indonesian, and Romanian). Possible projects include psychology/political-science-informed evaluations of LLMs’ understanding of attitudes, studies of dialect understanding and bias, and training-based approaches to improving model understanding.

2. Chatbot use in emotional support applications — LLM-powered chatbots are increasingly used for emotional support. This project uses a data-driven approach to understand how LLMs are used “in the wild” in these contexts, with a focus on the stories users tell chatbots about their personal/everyday lives.

3. Long-Horizon LLM RL — (A) LLM agents as long-context processors: improving accuracy, effective context length, and generalization of search/RAG agents through RL training. (B) Improving the efficiency of LLM RL training: methods for making RL more efficient for agentic tasks and/or LLM reasoning (transfer across model scales, on-policy distillation, trajectory-level credit assignment).

4. Multimodal and Interactive NLP — Multimodal NLP systems that generate and adapt language based on visual context, interaction history, user needs, and other input modalities: personalized/context-aware generation, multimodal reasoning, accessibility-focused applications, and understanding how model behavior changes across users and settings.

Students are also encouraged to apply for COMS 6975: Natural Language Generation and Summarization seminar: https://www.cs.columbia.edu/~kathy/coms6975.html

Required/preferred prerequisites and qualifications: We generally prefer that applicants have taken an NLP course at Columbia or their previous institution, as well as relevant ML courses depending on the topic. Interested students should be able to register for 3 course credits of supervised research for the Fall semester (~9 hrs/week). Contact: Kathleen McKeown, https://www.cs.columbia.edu/~kathy/


Faculty/Lab: Amogh Inamdar (PhD student) and Richard Zemel (PI)

Brief Research Project Description: Mathematical proofs are a frontier problem for LLMs. Learning hard proofs can be challenging for these models, as they involve a complex mix of concepts and long chains of reasoning. Post-training for proof solving in LLMs typically involves reinforcement learning on such data, with several works designing curricula over the data to induce learning. Most existing curricula are either fully domain-specific or model-specific. This project aims to develop a curriculum that combines both approaches: decomposing hard proofs into their constituent concepts, evaluating model proficiency on these concepts, and designing a teacher-driven curriculum that maximally develops the model’s proficiency on these concepts and on proof-solving as a whole.

Required/preferred prerequisites and qualifications: Proficiency with Python and ML tools (PyTorch, Hugging Face, vLLM, etc.); knowledge of basic CS math and ML concepts; experience training deep learning models for research/industry. Contact: Richard Zemel / zemel@cs.columbia.edu / https://www.cs.columbia.edu/~zemel/


Faculty/Lab: Ira Ceka (PhD student) and Baishaki Ray (PI)

Brief Research Project Description: Building small World Models (WMs) to simulate real-world infra systems where these WMs will be used as simulators to RL post-train larger coding agents for SRE tasks.

Required/preferred prerequisites and qualifications: Required: Understanding of reinforcement learning (RL) concepts, experience with coding agents, and familiarity with setting up machine learning training pipelines.
Preferred: An understanding of world models and SRE tasks, along with basic exposure to analyzing agentic trajectories.


Faculty/Lab: Yaniv Nemcovsky (PhD student) and Junfeng Yang (PI)

Brief Research Project Description: My research focuses on AI security and trustworthiness, helping ensure AI systems can’t be tricked, misused, or compromised. Current projects span adversarial robustness, real-world physical adversarial attacks and defenses, LLM safety alignment and jailbreak attacks, backdoor and unlearning detection, and semantic-based agentic routing and orchestration. Students will have the opportunity to explore and advance research on making AI systems safer and more reliable by uncovering and mitigating models’ failure modes.

Required/preferred prerequisites and qualifications: Strong foundation in machine learning and deep learning (coursework or projects). Experience with adversarial robustness or jailbreak attacks is preferred but not required.


Faculty/Lab: Christian Cianfarani (Postdcoc) and Junfeng Yang (PI)

Brief Research Project Description: Using theoretical methods to explore how collectives of individuals can leverage machine unlearning techniques to influence the behavior of AI systems.

Required/preferred prerequisites and qualifications: Required: Some background in learning theory and algorithms


Faculty/Lab: Judah Goldfeder (PhD student) and Hod Lipson (PI)

Brief Research Project Description: The Creative Machines Lab is looking for students to join a variety of projects across ML, AI, and Mechanical Engineering. For a list of projects, see here:
https://docs.google.com/document/d/1l1_YsDCfyp3Yli0I84fDVfmK_bmhonuN8KcMecmsneM/edit?tab=t.0#heading=h.1l1hafjy4l4q

To apply:
https://docs.google.com/forms/d/e/1FAIpQLSf1Iv-Ksd50Szh-uw7dTr0WX9Nk5l-AJCVrKPyPGEtn_yo_bA/viewform

 

Required/preferred prerequisites and qualifications: Research experience is prefered but not required

Faculty/Lab: Baishakhi Ray / ARiSE Lab, Student Member: In Keun Kim

Brief Research Project Description: Can an AI agent truly understand a large enterprise system—or is it simply searching faster than a human? We are building a rigorous benchmark to find out, using runnable legacy applications where the truth is scattered across code, databases, documentation, policies, tests, and system logs—and where those sources often disagree.

Research focus: reconstructing system knowledge (combining fragmented evidence into a reliable model while preserving conflicts between sources); improving agent reasoning (testing whether typed, provenance-aware knowledge graphs help agents predict outcomes and plan workflows better than retrieval or raw context); and measuring genuine understanding (designing hidden tasks and held-out evaluations that distinguish real system understanding from source lookup, copying, or memorization). The project brings together software engineering, AI agents, program analysis, knowledge representation, big-data systems, and benchmark design.

Required/preferred prerequisites and qualifications: Experience building an end-to-end software product, from design and implementation through testing; experience with data-intensive platforms or architectures such as data lakes and graph databases; passion for research and grit. If interested, email ik2619@columbia.edu with your CV/resume, including any prior software engineering or AI-related experience.

Session: Wednesday, 9/9/2026 at 11:00 AM

https://meet.google.com/ckn-ebhf-dya


Faculty/Lab: World Models (working under Nicolas Beltrán Vélez, Dave Blei’s PhD student)

Brief Research Project Description: World models correspond to probabilistic models of state-space transitions p(s’ | s, a). These models can be used to optimize at test-time via search, or to train agents on top of them, forming a fundamental class of algorithms for offline RL. We’re looking to work with students on two problems:

1. Efficient Inference-Time Optimization — The current standard for inference-time optimization consists of very simple algorithms such as the cross-entropy method, which is inefficient. Can we amortize this? Can the model learn intelligent proposals? This is an ambitious problem that will likely take more than a semester.

2. Small World Models — Video-based world models are generally quite large (1B+ params), but there has been significant progress on small flow models with good performance (e.g. minit2i-torch). The goal of this project is to build a video model useful for test-time planning that is as small as possible and can be trained on academic hardware in 2–3 days, enabling further study of questions like post-training RL on an academic budget.

Required/preferred prerequisites and qualifications: Should be able to code flow-model algorithms (flow-matching, diffusion models, sampling, etc.) and basic RL algorithms (policy gradients, etc.) without help; should be very comfortable with PyTorch, neural networks, transformers, and deep learning. The ideal mentee has good taste in using coding agents, is careful about reviewing their code, and has strong software engineering standards. Open to everyone, including freshmen, who can demonstrate knowledge in these areas.

Session: Thursday, 9/10/2026 at 2:00 PM

https://meet.google.com/pxe-tcgx-iyv (PIN: 996 186 499#)

Please apply here by Tuesday September 15: https://forms.gle/zwqMTLhHA2r7EaFTA.


Faculty/Lab: Prof Steve Feiner | Computer Graphics and User Interfaces Lab (Prof. Feiner, PI)

Brief Research Project Description: The Computer Graphics and User Interfaces Lab (Prof. Feiner, PI) does research in the design of 3D and 2D user interfaces, including eXtended Reality (XR)—encompassing augmented reality (AR) and virtual reality (VR)—and mobile and wearable systems, for people interacting individually and together, indoors and outdoors. We use a range of displays and devices to develop novel interaction and visualization techniques, which we evaluate in prototype applications. Multidisciplinary projects potentially involve working with faculty and students in other schools and departments, from dentistry and medicine to earth and environmental sciences.

Required/preferred prerequisites and qualifications: We’re looking for students who have done excellent work in one or more of the following courses or their equivalents elsewhere: COMS W4160 (Computer graphics), COMS W4170 (User interface design), COMS W4172 (3D user interfaces and augmented reality), and COMS E6173 (Topics in VR & AR), and who have software design and development expertise. Knowledge of robotics and vision is a plus for some projects. For those projects involving 3D user interfaces, we’re especially interested in students with Unity experience.

Session: Friday, 9/11/2026 at 14:30

https://columbiauniversity.zoom.us/j/99119389700?pwd=40Oi6xd2y46wlc2kZPv9XBVlcDylYy.1


Last updated: 09/10/2026