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

Brief Research Project Description:

https://coreytolerfranklin.com/gilmlab/
Click Here To Apply: https://forms.gle/HBenKuxio6TA7n7WA

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 super computing 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: High priority to students with computer graphics and/or physics experience. General requirements: Python and/or C/C++, Machine Learning experience

---

Faculty/Lab: Prof Hod Lipson and doctoral student Judah Goldfeder

Brief Research Project Description:

LLM Concept Analysis: Can we understand how LLMs think, and how their conception relates to human thought? Extending ideas from this paper: https://arxiv.org/pdf/2510.26025

Generating Auxiliary Tasks with Reinforcement Learning: Extending this paper: https://arxiv.org/abs/2510.22940

Self replicating NN: See if a NN can learn to output the value of its own weights (all of them) and at the same time also learn to perform some other task.

k-Nearest Neighbor Learning with Graph Neural Networks: Building off of this paper: https://www.mdpi.com/2227-7390/9/8/830

Required/preferred prerequisites and qualifications: We are looking for people with strong expertise in ML, high level coding abilities, and with some experience conducting research. https://docs.google.com/forms/d/e/1FAIpQLSduMnE4AkApxcYSWSWJ2_RBKFiGj9u-9ezBviWjPV9Cao6EiQ/viewform?usp=dialog

---

Faculty/Lab: Prof Asaf Cidon & Yuhong Zhong

Brief Research Project Description: Improving Memory Utilization in Modern Datacenters with CXL-aware Allocation
Large-scale datacenters suffer from low memory utilization even though memory is one of the most expensive server resources. Compute Express Link (CXL) is a new high-speed interconnect that lets multiple hosts access shared pools of memory, opening the door to far more efficient resource usage.
This project explores software techniques for intelligent memory allocation in CXL-enabled systems. You’ll help design and evaluate allocation policies that improve overall utilization by considering virtual machine (VM) history, dynamic VM migration, and cross-host load balancing. The work involves simulation and prototyping to study how different allocation strategies affect performance.
We’ll also study reliability challenges: when CXL links or memory devices fail, naive allocation policies can cause large disruptions. You’ll investigate how to make allocation resilient to failures and less sensitive to topology changes.
Students interested in software systems and distributed resource management will gain hands-on experience with modern memory technologies and research-grade system design.

Required/preferred prerequisites and qualifications:

* Have taken OS1 or equivalent
* Interested in software systems research
* Can contribute 15+ hours per week

---

Faculty/Lab: Prof Matthew Connelly | History Lab

Brief Research Project Description: The History Lab has more than 4 million declassified government documents. Part of our research focuses on defeating the “mosaic effect,” a vulnerability where an adversary pieces together innocuous, declassified records to reveal sensitive intelligence patterns, capabilities, or sources. This work aims to solve the critical challenge of sensitive information extraction and malicious user identification when interacting with Large Language Models (LLMs).We are seeking an intern to contribute to the development of two primary systems: the MIDLLMAT Filter, which performs “targeted unlearning” to redact documents that could reveal secrets, and the MIDLLMAT Sensor, which monitors user prompts in real-time to detect malicious intent.

Required/preferred prerequisites and qualifications: Advanced Python skills; knowledge of postgres and experience with LLMs

---

Faculty/Lab: Prof. Baishakhi Ray & Dr. Robin Ding

Brief Research Project Description: 

Red-teaming the Coding AI Agents
This project will red-team coding AI agents (e.g., Claude Code, Gemini CLI, Qwen-3-Coder) by systematically probing how they can be steered into unsafe behavior through the untrusted context it fetches and the agent’s own tool feedback loop. The goal is to systematically measure and analyze the failure modes of existing coding agents against prompt injection under developer environments. Ideally we also target producing actionable defenses and guardrails for tool use, safer prompting and retrieval patterns, and lightweight verification (static/security checks, patch risk scoring), along with an open benchmark and guidelines that help developers deploy coding agents more safely in real software workflows.

Required/preferred prerequisites and qualifications:

Requirement: If you are interested, please make sure to send your CV and research background to yrbding@cs.columbia.edu before you come to the research fair.

Preference: Hands-on engineering experiences with LLM agents and research experiences with AI security or adversarial machine learning are strongly preferred.

---

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

Brief Research Project Description:

INCREASING REPLAYABILITY IN EDUCATIONAL GAMES 

This project investigates how to increase replayability within a specific subgenre of educational games. This project is recruiting RAs for two different positions:

Position 1: User Researcher RA Representative Responsibilities:

• Conduct literature reviews  
• Conduct participant recruitment  
• Develop research instruments (e.g., interview guide, surveys, observation protocol)
• Run user studies and playtest sessions
• Data entry  
• Transcribe interview audio
• Analyze qualitative data (e.g., interview transcripts, observation notes, gameplay recordings)
• Run basic (i.e., introductory-level) statistical tests  
• Contribute to manuscript writing 

Desired Qualifications

• Has experience playing video games  
• Can analyze and describe the design of video games they have played  
• Is organized, detail-oriented, and accurate  
• Has experience with social science research or user research (courses count as prior experience)

Position 2: Gameplay Programmer RA Representative Responsibilities:

• Identify software design patterns used within similar games  
• Code the game mechanics based on the design documentation from the game designer  
• Review and provide feedback on code from other gameplay programmers  
• Write unit tests  
• Perform quality assurance testing  
• Communicate and discuss your code to technical and non-technical collaborators 

Desired Qualifications:

• Has experience playing video games  
• Can analyze and describe the design of video games they have played  
• Has experience with Unity  
• Has experience with C#  
• Has experience with Git

COMPUTATIONAL RATIONALITY FOR ACCESSIBLE VIDEO GAMES

This project investigates how computational rationality can be used to model player behavior and optimize assistive interfaces in video games, with a focus on improving accessibility for blind players and other players with disabilities. This project is a collaboration with Prof. Antti Oulasvirta and Dr. Thomas Langerak from Aalto University, who are world leaders in computational rationality. Computational rationality models users’ perception, cognition, and action selection in order to predict their behavior and optimize interactive systems for them.

The project builds on prior work from the CEAL lab on the Racing Auditory Display (RAD) (https://ceal.cs.columbia.edu/rad/), and moves beyond hand-crafted assistive cues toward automated generation of audio cues and assistive interfaces. The core goal is to use reinforcement learning to learn player models that represent cognitive capacity and perception, and to optimize auxiliary displays (e.g., audio cues) that maximize player performance in games.

Position: Machine Learning / Reinforcement Learning RA Representative Responsibilities:

• Implement reinforcement learning agents that model player behavior under perceptual and motor constraints
• Develop parametrizable gamer models, including (A) perception, (B) motor control, (C) internal representation, and (D) intrinsic and extrinsic reward models
• Integrate RL agents with games or simulators (e.g., Unity environments) that provide virtual sensors, reward signals, and training interfaces
• Model the input device, including (A) action spaces and (B) delays
• Model the auxiliary display as an output signal space
• Implement a bidirectional optimization process, including (A) an inner loop that trains the gamer model and (B) an outer loop that optimizes the auxiliary display set to maximize gamer reward
• Run computational experiments and analyze learning and performance outcomes
• Contribute to research publications and technical reports

Desired Qualifications:

• Strong technical background in machine learning
• Experience with reinforcement learning (coursework or project experience acceptable)
• Comfortable implementing and debugging RL systems and simulations
• Interest in accessibility or human-computer interaction (HCI)
• Interest in games, simulation environments, or interactive systems.

To express interest in working with us, please complete the following Google Form

Required/preferred prerequisites and qualifications: See each project above. We also have open-door lab meetings, open to everyone to drop in whenever they would like. They are a great place to meet everyone, ask questions, and get a better sense of our current projects. If you would like to attend any of these, add yourself to the CEAL Circle Google Group to start receiving calendar invites. The weekly time varies from term to term. You can remove yourself from the Google Group anytime to stop receiving calendar invites. The Google Group can be found at:
 https://groups.google.com/g/ceal-circle 

---

Faculty/Lab: ARCADE lab | Prof. Martha Kim & doctoral student Soyoon Park

Brief Research Project Description: We aim to optimize specialized heterogeneous hardware design for a given workload, such as CNN. Full cycle-level simulations of SoC(System-on-Chip) designs take too much time and are impractical to iterate over expansive design parameters. Therefore, exploration is limited, many depending on heuristics. Instead, we propose a fast SoC simulation to gauge relative performance, focusing on on-chip load/store data transfer abstractions. This tool will aid early-stage exploration of system-level SoC design choices.

Required/preferred prerequisites and qualifications:

Python scripting
Verilog/SystemVerilog
FPGA programming
ESP(Embedded Scalable Platforms)
Familiarity with CNN workloads

---

Faculty/Lab: Speech Lab | Prof. Julia Hirschberg & PhD student Yu-Wen Chen

Brief Research Project Description: Noise-Robust Speech Processing: Real-world speech processing systems are often degraded by environmental noise, including background sounds and reverberation from sources such as television audio. This project aims to investigate effective strategies for mitigating noise and improving the robustness of speech processing systems. The possible research directions include:
– Surveying state-of-the-art methods for noise-robust speech processing.
– Examining how noise reduction and speech enhancement techniques influence speaker traits and downstream speech processing tasks.
– Developing improved methods to further enhance the robustness of existing models.

Required/preferred prerequisites and qualifications: Demonstrated proficiency in Python, including experience with PyTorch and training machine learning models. Background knowledge in speech processing is a plus.

---

Faculty/Lab: Wireless and Mobile Networking Lab | Prof. Gill Zussman

Brief Research Project Description: List of projects:
– Security and Reliability of Edge–Cloud Systems for Video Analytics
– Real-Time LLM Inference in Distributed and Wireless Mobile-Edge Networks
– Joint Sensing and Communication with mmWave and Terahertz Wireless for 6G Networks
– Weather-Wireless Data Analyst and Integration Coordinator
– NQVL – Developing a New York Quantum Capable Internet Testbed
– INDIGO – Orchestration Across Multi-Operator, Multi-Vendor, 5G Networks During Civilian Disasters
– Experiments in Full-Duplex Physical Layer Security

Detailed descriptions and contact details appear HERE

Required/preferred prerequisites and qualifications: See Here 

To apply, please use – https://forms.gle/Km7HR4GoSRhSmGW36

---

Faculty/Lab: Prof. Xia Zhou, Prof Sal Stolfo, & PhD student Xiaofeng Yan

Brief Research Project Description: We’re exploring a unified direction for robust palm biometrics by creating the data we wish we had. The idea is to use diffusion-based generative modeling (e.g., Stable Diffusion + LoRA / identity-conditioned tokens) to synthesize new palm images that preserve a person’s identity while varying real-world nuisances like wet/dirty surfaces, illumination, specular highlights, and viewpoint. In parallel, we’ll study style/spectrum transfer methods (e.g., neural style transfer, frequency-domain spectrum mixing, domain adaptation) to “move” the appearance of a condition onto a clean palm image without changing identity, giving us controllable and interpretable augmentation.

Required/preferred prerequisites and qualifications: We’re looking for students who can help implement and evaluate these ideas end-to-end: solid Python skills, comfort working with image datasets, and basic machine learning/computer vision familiarity. Bonus points (not required) for interest in PyTorch, diffusion pipelines, and simple frequency analysis (FFT).

---

Faculty/Lab: Internet Real-Time Lab | Prof Henning Schulzrinne & PhD Student Sunny Fang

Brief Research Project Description: Federal grants such as the Broadband Equity Access and Deployment (BEAD) program aim to bring high-speed internet access (“broadband”) to unserved rural areas in the United States. To understand challenges faced by BEAD, a better understanding is needed to answer open questions as location eligibility inconsistencies, serviceability, and scalability of potential alternatives (e.g., satellite network). Selected students will analyze large datasets, formulate research questions, and test hypotheses related to the program’s efficacy.

Required/preferred prerequisites and qualifications: Statistics; Python; SQL preferred; Basic understanding of broadband technologies

---

Faculty/Lab: SEA Lab | Prof Xuhai (Orson) Xu, Dr. Blue Lin  Phd & student Millie Wu

Brief Research Project Description:

#1 Blue Lin: Beyond Single-Session Health AI: Designing and Evaluating Health Agents for Longitudinal Care

This project explores how AI health agents can move beyond single-session, task-oriented interactions to support longitudinal, multi-session care. Drawing on clinical and health informatics frameworks, it defines persistent-agent requirements and explores architectures ranging from stateless to fully stateful agents. The project focuses on learning evolving user representations and optimizing recommendations over behavioral trajectories. Evaluation combines scenario-based expert studies with system comparisons, with a longer-term goal of deploying and analyzing real-world longitudinal interaction data.

#2 Millie Wu: Sleep and Value-Aligned Mental Rehearsal for Student Well-Being and Self-Regulation

We are conducting a research project on sleep, mindfulness, and mental rehearsal to understand how digitally delivered cognitive training can support well-being, self-regulation, and daily goal pursuit in university students. The project combines an ongoing sleep-focused mindfulness study with a new value-based mental rehearsal intervention that helps individuals mentally rehearse meaningful daily actions, cope with obstacles, and reflect on progress. Using a multi-condition experimental design, the study compares sleep-based cognitive training, baseline planning, and value-aligned mental rehearsal to examine how nighttime recovery and daytime cognitive strategies jointly influence sleep quality, affect, self-efficacy, and overall wellness.

#3 Millie Wu: Personal Health SENSEI: Socially Embedded Neuro-Somatic Sensing for Educational Intervention

Personal Health SENSEI is a research project that develops socially intelligent personal health agents by integrating multimodal physiological sensing with longitudinal, expert-aligned reasoning. Using mindfulness training as a testbed, the system introduces a dual-graph neuro-symbolic architecture that combines a static domain knowledge graph with a dynamic personal knowledge graph to enable adaptive, explainable, and value-aware interventions over time. By grounding LLM-based agents in physiological signals and transparent reasoning structures, SENSEI aims to transform generic health chatbots into perceptive companions that support sustained learning, engagement, and well-being.

Required/preferred prerequisites and qualifications:

#1 Blue Lin: Beyond Single-Session Health AI: Designing and Evaluating Health Agents for Longitudinal Care
I am looking for 1–2 students with strong experience in at least one of the following areas: app development, machine/deep learning, or building/optimizing LLMs and fine-tuning. Ideal candidates are interested in both building and deploying systems as well as conducting research and writing papers.

#2 Millie Wu: Sleep and Value-Aligned Mental Rehearsal for Student Well-Being and Self-Regulation
We are seeking 2–3 students with experience in running and managing HCI and/or cognitive science studies, and/or in mobile app development (e.g., React Native, Firebase, FastAPI). Research assistants will support participant coordination, study execution, and data collection for a smartphone-based study, and may also contribute to app development, iteration, and exploratory data analysis, depending on background and interests.

#3 Millie Wu: Personal Health SENSEI: Socially Embedded Neuro-Somatic Sensing for Educational Intervention
We’re looking for 1-2 students with interest and experience in Knowledge graph, Multi-modal ML, and GraphRAG. RAs may help with building knowledge graphs, integrating multimodal systems, and supporting related modeling and evaluation tasks.

---

Faculty/Lab: Prof Henning Schulzrinne

Brief Research Project Description: (1) Analysis of high-speed internet deployment in rural areas in the United States; (2) Using smartphones to reduce injuries to pedestrians and bicyclists; (3) Reducing fraud and phishing by email, text and voice calls.

Required/preferred prerequisites and qualifications: Statistics; databases; networks (preferred)

---

Faculty/Lab: Prof Venkat Venkatasubramanian

Brief Research Project Description:

Naz (nt2513@columbia.edu)
Mechanistic Interpretability – Inner workings of LLMs
Despite their success, how large language models internally organize knowledge remains unclear. This project builds on prior work using sparse autoencoders (SAEs), geometric analyses, and probing methods to study monosemanticity in learned representations. The goal is to scale these analyses to larger LLMs on GPU clusters and extend them to multi-level SAE-based representations.

Requirements: experience with NLP, Machine Learning (PyTorch proficiency), familiary with LLMs/transformers, GPU experience is a plus.

AI for pharmaceutical discovery and manufacturing
SUSIE is our ontology-based tool for pharmaceutical information extraction, designed to convert unstructured text into knowledge graphs (see past work: https://doi.org/10.1016/j.compchemeng.2023.108446, https://doi.org/10.1016/j.compchemeng.2025.109318). Students joining the project will work on refining information extraction from pharmaceutical documents and on neural symbolic inference methods (eg: how to transparently infer new, domain-informed conclusions from existing knowledge, including plausible drug–target relationships and mechanistic hypotheses). Applications include drug safety labels and pharmaceutically relevant chemical reactions.
Requirements: experience with NLP, predictive algorithms.
Learning a shared drug-protein manifold for drug repurposing
This project studies internal representations from AlphaFold2 and related models to study how deep models organize biological knowledge internally. Students will extend contrastive learning frameworks to residue-level attention and leverage learned embeddings to infer drug targets, functional sites, and repurposing opportunities.
Requirement: experience with machine learning, preferred experience: representation learning, GPU usage.

Ricky (kk3764@columbia.edu)
1) Self-Learning Genetic Algorithm with Reinforcement Learning
Evolutionary model discovery frameworks generate large numbers of candidate ODE models to
explain time-series data. The proposed equations are often stiff and high-dimensional in
chemical and biological contexts, highlighting the need for intelligent and adaptive genetic
operators. To avoid unnecessary model evaluations and enable efficient navigation in large
symbolic search spaces, this project utilizes RL to inform structural refinement decisions.
– Requirement: RL, Combinatorial Optimization, Chemical Kinetics (optional)

2) SDE Model Discovery with Symbolic Regression
SDEs are used to model systems where randomness significantly affects behavior, bridging ODEs with real-world uncertainty. Several main applications include finance (stock prices) and biology (population dynamics). Our main goal is to extend an existing ODE-based symbolic regression framework to generate SDE model candidates and capture the underlying dynamics in the given time-series data.
Requirement: Extensive background in Python, Stochastic Calculus, Probability Theory

Vincent (vv2435@columbia.edu)
1) Reinforcement learning for competitive economic agents
Statistical teleodynamics is a conceptual framework that integrates principles from game theory and thermodynamics to explain emergent macroscopic phenomena from microscopic interactions. It explains how complex interacting systems can converge towards global equilibrium. The global economy resembles a globally stable system, whose multiple agents optimize their utility while being in competition with each other.
The modelling of competition justifies the use of RL to model and understand how an arbitrage-free equilibrium can be reached, depending on the level of competition. This project aims at building a model of competitive agents to study the resilience of the general equilibrium.
Requirements: reinforcement learning

2) Convergence of arbitrage in illiquid financial markets
Traditional financial models often assume perfect liquidity, symmetric information, and the absence of arbitrage. In practice, however, markets are defined by non-zero spreads and transient arbitrage windows.
This project investigates how information asymmetry triggers “liquidity avalanches.” We will analyze the transition from stable trading to rapid price movements, identifying how insiders capitalize on brief arbitrage opportunities created during these periods of market inefficiency.
Requirements: time series, financial modeling

Sanjeev (sn3130@columbia.edu)
1) Pattern Interpretability in Neural Networks
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: 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.
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.
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.
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 Network

2) Dynamics of tokens and the theory of LLMs
By dividing the dynamics of LLMs into three major categories: the training, planning and inference, we analyse how a giant trained model like 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:
Does LLM plan before prediction? (LLM planning)
How does the LLM hallucinate, is it imperfection in planning? (LLM inference)
How does the training looks like in the ambient feature space (LLM training)

3) LKM or Hybrid models
This project aims to engineer an architecture or framework for a transformer model to learn engineering. An LLM is good at natural language inference but really bad at engineering-specific language. This includes safety rules and reaction discovery. Understanding that LLMs’ predictions can be improved by explicitly providing knowledge, like RAG or prompting, we want to analyse techniques to improve the LLM implicitly by training or by a hybrid knowledge-driven method. We’re building a new architecture that works as a knowledge base that is backed up by fundamental laws and definitions. This architecture, called LKM, Large Knowledge Models, should be able to use 1st principal rules, including heat or energy balance.
Paper: https://aiche.onlinelibrary.wiley.com/doi/pdf/10.1002/aic.18661
Prerequisites: Python, Statistics, architecture design, math

4) Risk Analysis, Analytical and Statistical Methods for Engineering: (1-2 Open Positions)
This project aims to model the risk analysis techniques in engineering (HAZOP analysis in Industries) to understand and automate the mechanisms for risk. Systems like supply chains, chemical industries, or financial institutions all tend to fail at times. For a given system with multiple sub-systems, we are interested in the following:
Analysing failure events to automate these mechanisms with AI to discover new failures
Build a statistical model that predicts the probability of failure at any moment for different scenarios.
Based on these probabilities of the sub-system, can we build a system-level model hierarchically for improved analysis?
Paper: https://aiche.confex.com/aiche/2025/meetingapp.cgi/Paper/712964
Prerequisites: Python, Statistics, architecture design, math

Collin (cjs2301@columbia.edu)
1) 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 on biochemical pathways, molecular functions, and protein structure, this project is concerned with the generation of natural language summaries of system level differences between subjects and groups of subjects inferred from high-throughput, raw omics data.
Prerequisites: Causal Inference, Statistics, Large-Language Models

2) 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 sequence and amino acid pairs are iteratively generated. This project is concerned with identifying alternative tasks that can be performed using representation vectors included, but not limited to, prediction of binding motifs, evolutionary relationships between folds, and identification of regions susceptible to conformational changes.
Prerequisites: Computational Biology, Machine Learning, Protein Structure

3) Theory of Large Language Models
This project is concerned with the mathematical, organizing principles of Large Language Models. We will view LLMs from a variety of perspectives, including statistical mechanics, geometry, and Markov chains with memory. Akin to the “laws” of physics, we are seeking a framework of “laws” that describe the organization of knowledge inside LLMs at a variety of scales. This project will be fairly impressionistic, and will require creative thinking.
Prerequisites: NLP, Physics, Mathematics (Probability Theory, differential geometry)

Required/preferred prerequisites and qualifications: See each project and PhD student above.

---

Faculty/Lab: Prof. Zhuo Zhang

Brief Research Project Description:

Project 1: Benchmarking AI Agents for Reverse Engineering (Offensive Security) – Reverse engineering means figuring out what a program does when you only have the compiled binary (the machine-code version) and not the source code. This is common in malware analysis and vulnerability research, and it is a core skill in offensive security. Recently, AI agents powered by large language models have shown promise in security tasks, but we still do not have a good way to measure how well they handle realistic reverse-engineering work. In this project, you will help build a benchmark that tests AI agents on reverse-engineering tasks in a fair and repeatable way. The goal is to design meaningful tasks, run different agent setups, and summarize what the agents can do well, where they fail, and what kinds of tools or strategies help them most.

Project 2: Type Inference for Binaries from Modern Languages (Rust/Go/Swift, etc.) – A big part of reverse engineering is recovering “types”, meaning identifying what kind of data variables represent (e.g., integers, pointers, structs, strings, or objects). Knowing types makes low-level code much easier to understand, but most existing research and tools focus on C/C++ programs. In the real world, more and more software is written in modern languages like Rust, Go, and Swift, and these languages can produce binaries that behave differently and hide type information in different ways. In this project, you will study how type recovery changes across programming languages and why some languages are harder than others. The end goal is to build a fair evaluation approach: basically, a reliable way to test and compare type-recovery methods for each language, so we can understand what works, what does not, and what future tools should focus on.

Project 3: Next-Generation Debugging for Ethereum Smart Contracts – Smart contracts are programs that run on blockchains like Ethereum, and they often manage money or important assets. Even small bugs can cause real damage, so debugging is critical — but current smart contract debugging is still frustrating. Developers often struggle to understand what happened during a transaction, why it failed, and how the on-chain execution connects back to their original source code. In this project, you will explore why debugging remains painful and design a better debugging approach that makes failures and state changes easier to understand. You will also help evaluate the new design, including whether AI can assist by summarizing execution traces, pointing out likely root causes, or turning complex technical logs into explanations that developers can act on.

Project 4: Vision-Language-Model GUI Testing Agent for Android Apps – GUI testing checks whether an app behaves correctly through its user interface, and it is especially important for finding issues that only appear after certain button clicks, screens, or user flows. This is also useful for detecting risky or policy-violating behavior such as misleading prompts, suspicious permission requests, or hidden flows that are hard to catch with traditional scripted testing. Traditional GUI testing tools can be brittle because they do not truly “understand” what is on the screen.

 In this project, you will help build an Android testing agent that uses Vision-Language Models (VLMs) to look at the app screen and decide what to do next, more like a human tester. The goal is to make testing smarter and more effective, and to measure whether VLM-based testing can reach more app states and discover more meaningful issues than existing approaches.

Required/preferred prerequisites and qualifications:

# Projects 1 & 2: Reverse Engineering / Type Recovery (Binaries) – To be eligible for Projects 1 or 2, you must already have hands-on reverse engineering experience. A common way to show this is having taken COMS W4186 (Reverse Engineering) and performed strongly, or having equivalent practical experience (for example, solving CTF-style reversing challenges or doing malware/vulnerability analysis work). You should also be comfortable working in system programming languages such as C/C++, since a lot of reverse engineering work assumes you can read low-level program behavior and reason about memory, pointers, and calling conventions.

Having experience with modern languages such as Rust, Go, or Swift is a strong plus (especially for Project 2). Experience building or experimenting with AI/LLM agents (tool-using workflows, automation scripts, evaluation pipelines) is also highly appreciated. If you want to self-check your reverse engineering readiness, you can try this skills evaluation challenge: https://zzhang.xyz/challenges/level_0_welcome.html.

# Project 3: Ethereum Smart Contracts / Web3 Debugging – For Project 3, you must be genuinely comfortable with the Ethereum/Web3 stack: not just the buzzwords. In practice, this means you understand how smart contracts run on the EVM, how transactions change contract state, what gas is and why it matters, and how common development workflows work (writing/deploying/interacting with contracts). You should also be familiar with core concepts like events/logs, common failure modes (reverts, require checks), and the typical toolchain developers use to test and debug contracts.

It is a strong plus if you have prior experience writing TypeScript and building developer tools. Experience with MCP development is also appreciated.

If you want to gauge your Web3 fundamentals, you can try this knowledge evaluation challenge: https://zzhang.xyz/challenges/level_4_82457693012849571639.html.

# Project 4: VLM-Based Android GUI Testing Agent – For Project 4, Android emulator automation is a hard requirement. To join this project, you must be able to build a small command-line tool that can (1) start an Android emulator, (2) launch a specific app, and (3) take a screenshot reliably. This is the minimum bar because the research depends on stable, repeatable control of the emulator and the app UI. If you can go beyond the basics, such as tapping UI elements, typing text, scrolling, handling permission popups, or collecting simple logs, that is even better, but not required for the initial screening.

Experience related to Vision-Language Models (VLMs) is also a plus, and hands-on experience training or fine-tuning VLMs (or doing data collection/cleaning for VLM training) would be especially appreciated.

For the entry exercise, you will create a small GitHub repository containing your CLI tool and brief setup/run instructions, then share it with us so we can evaluate whether the automation is robust and reproducible.

---

Faculty/Lab: ARISE Lab | Prof. Baishakhi Ray & PhD Student Nikolaus Holzer

Brief Research Project Description: Project centered around coding agents, security, and scalability

Required/preferred prerequisites and qualifications: We are looking for students that have very strong engineering skills for ML research (Able to work with linux OS at a very deep level, pytorch, know how to use huggingface) Having taken classes like HPML and deep learning project classes is a plus.

---

Faculty/Lab: Prof. Kathy McKeown & 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, including 1) creating immersive experience for low-vision and blind people to explore art images, 2) analyzing and improving VLM’s ability to perform formal analysis of style, 3) understanding attitudes in under-represented dialects, and 4) update summarization with diffusion models.

Required/preferred prerequisites and qualifications: Preferably have taken NLP and relevant CS courses depending on specific project (ML, CV, etc.) or have relevant prior research experience.

_______________________________________________________________________

Faculty/Lab: The Computer Graphics and User Interfaces Lab | Prof. 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 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: head-worn, hand-held, and table-top, including Varjo XR-3, HoloLens 2, and Magic Leap 2, in addition to consumer headsets such as Meta Quest 3. Multidisciplinary projects potentially involve working with faculty and students in other schools and departments, from medicine and dentistry 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: Prof Junfeng Yang & PhD Students Hideaki Takahashi

Brief Research Project Description: Bug Detection and Verification of Zero-Knowledge Proof (ZKP) Systems
Zero-knowledge proofs (ZKPs) are an emerging cryptographic paradigm that has attracted significant attention in Web3, cryptocurrencies, privacy-preserving machine learning, secure identification, and related areas. However, even minor implementation errors can compromise the security guarantees of ZKP systems.
We are working on efficient techniques for bug detection and verification of ZKP systems, combining both system-level and theoretical perspectives.
Our official project page is available at https://zkfuzz.xyz/.

Required/preferred prerequisites and qualifications: Experience with software security, blockchain, cryptography, discrete mathematics, or Rust is a plus, but not required.

---

Faculty/Lab: Prof Ken Ross & PhD Student Junyoung Kim

Brief Research Project Description: In this project, we will be investigating how a fast, sparse matrix multiplication library previously developed by our group can be used to accelerate ML workloads. The role of the research student would be to configure existing, popular ML pipelines to use our novel matrix multiplication library, and run benchmarks to analyze performance. We will initially start by optimizing popular open-source ML operators that use popular sparse attention schemes (links below), and extend our scope to full ML pipelines later in the semester.
https://docs.pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html
https://www.tensorflow.org/api_docs/python/tfm/nlp/layers/BigBirdAttention
https://github.com/mit-han-lab/Block-Sparse-Attention

Required/preferred prerequisites and qualifications: Introduction to Databases, Machine Learning

---

Faculty/Lab: Prof Matthew McDermott, Greg Kondas

Brief Research Project Description: 1) Building Foundation Models for Medical Record Data: In this project, you’ll work on building foundation models capable of zero-shot inference about diverse questions in health via electronic health record data. This project will in particular aim to adapt state of the art techniques from natural language processing to the health AI field to build bigger, more powerful models using less data.

2) Building decentralized benchmarks for health data: In this project, you’ll leverage the MEDS framework and MEDS-DEV platform to establish a comprehensive benchmark of health-AI tasks, models, and datasets that can power an “ImageNet moment” for health AI

Required/preferred prerequisites and qualifications: 1) Required: Strong familiarity with AI/ML tools and methodologies, in particular Python, PyTorch, Pandas/Polars, HuggingFace, etc. Required: Some familiarity with large language models and how they are trained — e.g., the transformer architecture, next-token prediction (and why it is a sensible pre-training task), etc. Preferred: Familiarity (the more the better) with retrieval augmented _pre-training_ (not generation) methods, such as those explored in REALM, ATLAS, and RETRO.

2) Required: Strong familiarity with python. Some familiarity with prediction tasks of interest in AI for EHR data. Preferred: Familiarity with hydra, GitHub actions workflows, software development, test-driven-development, and MEDS.

Faculty/Lab: Dr. Chengzhi Mao & Prof Junfeng Yang

Brief Research Project Description: Training large language model to think better and faster.

We have projects on multimodal LLM reasoning and diffusion model. We work on both reinforcement learning and model pretraining.

Required/preferred prerequisites and qualifications: Know deep learning programming and have time to devote.

Time: 10:00:00 PM,1/22/2026

Link: https://columbiauniversity.zoom.us/j/94526958041 pwd=WAGEXcRMcadhi2kUTwDgpRPDUiUTGm.1

---

Faculty/Lab: The Computer Graphics and User Interfaces Lab | Prof. 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 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: head-worn, hand-held, and table-top, including Varjo XR-3, HoloLens 2, and Magic Leap 2, in addition to consumer headsets such as Meta Quest 3. Multidisciplinary projects potentially involve working with faculty and students in other schools and departments, from medicine and dentistry 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.

Time: 1/22/2026  4:00pm

Link: https://columbiauniversity.zoom.us/j/98567986066?pwd=CFjdwF5aR4vAE1HlxiMzLCB9W1zAPZ.1&jst=2

 

---

Faculty/Lab: Prof. Smaranda Muresan & PhD student Anubhav Jangra

Brief Research Project Description: To develop and evaluate an artificial intelligence (AI) model that identifies suspected drug overdose deaths using narrative reports from medicolegal investigation reports and describe ethical considerations related to the use of AI in overdose surveillance.

Required/preferred prerequisites and qualifications: Prior research/work experience in deep learning/machine learning, with a preference for experience in NLP and HCI.

Time: 1/22/2026 | 11:00 AM | https://meet.google.com/rmu-jjyi-gfm

---

Faculty/Lab: Prof. Eugene Wu, Prof Kostis Kaffes, & PhD Student Eliane Ang

Brief Research Project Description: Explore how database branching enables better LLM agent training and rollout

Required/preferred prerequisites and qualifications:

– Have experience developing LLM agents for data processing tasks
– Feel comfortable hacking database internals

Time: 01/21/2026 | 11:00 | Meeting ID: 808 436 5493 Passcode: columbia

---

Faculty/Lab: Prof Henning Schulzrinne 

Brief Research Project Description: Wi-Fi and BlueTooth for traffic safety: The project evaluates how well Wi-Fi and BlueTooth work for traffic safety applications in urban settings, despite interference. Devices will use built-in Wi-Fi functionality and GPS to measure their position, and then convey this information to nearby vehicles, bicyclists, and pedestrians. The goal is to reduce pedestrian accidents at intersections, “dooring” (i.e., opening car doors onto the path of a bicycle), and accidents at stop signs.

Required/preferred prerequisites and qualifications: CSEE4119, systems courses (e.g., Advanced Programming or Operating Systems)

Time: 1/22/2026 | 17:00 | https://columbiauniversity.zoom.us/j/98494288078?pwd=KbHD2KNHqf0Q8OZRlZOGy9oTcM5a22.1

---

Faculty/Lab: Prof. Sharon Di 

Brief Research Project Description: My Lab, DitecT, has a list of exciting projects on LLM, VLM, and generative model development, with applications to smart cities and selfdriving cars. 

This semester, we have exciting non-paid research opportunities available for master’s students or undergraduates for research credit registration, and I am hoping to kindly request your assistance in distributing this information to MS/UG. Here is the LINK for the application, which contains the project description.

Required/preferred prerequisites and qualifications: 

• Commitment: 15–20 hours per week
• Technical Stack: Strong Python and PyTorch coding skills and ability to manage large-scale experiments
• Domain Expertise: Solid knowledge of LLM, VLMFamiliar with generative models (e.g. VAEs, GANs, diffusion models)specifically in the context of interpreting synthetic or data-driven imagery
• Mindset: Creative problem-solving skills to bridge the gap between computer vision and urban computing

Brief Research Project Description: 

Required/preferred prerequisites and qualifications: 

Time: date | time | link

---

Faculty/Lab: Prof Daniel Bauer

Brief Research Project Description: Understanding Story-Level Analogy. Analogy is a defining feature of human cognition, allowing us to make sense of new stimuli and ideas by mapping them to past experiences. 

The goal of this project is to explore various techniques for analogy identification and possibly analogy generation at the narrative level. Following prior work you will initially explore one of three approaches:

• fine-tuning sentence encoders using contrastive learning. The learning objective would be to maximize the similarity between query narrative Q and analogous narrative A, while minimizing the similarity between Q and a non-analogous distractor narrative N.

• Prompting based approaches, using LLMs with chain-of-thought reasoning. One approach here might be to ask the model to first identify the different elements of a narrative (objects/participants, their attributes, and relations between them), and then ask it to judge analogy based on this information. We can also experiment with fine-tuning LLMs.

• An approach based on explicit graph-based representations of the narrative, such as in Abstract Meaning Representation (AMR). 

Additionally, we are interested in approaches to generating new analogies and potentially using such artificial data to boost training. Finally, we are especially interested in approaches that can explain analogies.

Full details Story_Analogy_Research_Fair_Proposal.

Required/preferred prerequisites and qualifications: You should have completed COMS 4705 Natural Language Processing with a good grade. You should have excellent Python programming skills and be comfortable working with NLP data sets. You should have some experience fine-tuning transformer models and/or working with LLMs.