Research Fair Fall 2025
The Fall 2025 Research Fair will be held on Thursday, September 4th, and Friday, September 5th, from 12:00 to 14:00 in Carleton Commons. 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 sessions and recordings available – see below for all details.
We will also have tables from GrOCS, CS Advising, and CS Careers. More research projects will be added up until the day of the fair, so keep checking back.
In Person
Faculty/Lab: Dr. Corey Toler-Franklin, Graphics Imaging & Light Measurement Lab (GILMLab)
Brief Research Project Description:
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 and transformer-based self-attention modules.
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: High priority to students with computer graphics and/or physics experience. General requirements: Python and/or C/C++, Machine Learning experience
Apply Here: https://forms.gle/Eny9mZm6HA6rek938
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 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 and social work.
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. For those projects involving 3D user interfaces, we’re especially interested in students with Unity experience.
Faculty/Lab: Prof Junfeng Yang & postdoc Penghui Li
Application form – https://forms.gle/SZXijv8XgEBeyaJo9
Brief Research Project Description: AI-Powered Cloud Security Research Project – Modern cloud applications consist of thousands of interconnected microservices, each with different security permissions. When these services communicate, they can accidentally create vulnerabilities that give attackers unauthorized access – but current security tools can’t detect these cross-service flaws. We’re building an intelligent system that uses advanced AI and program analysis solutions to automatically scan code across all services, identifying security vulnerabilities that traditional tools and human analysts would miss. Students will help design, build, and test these AI-powered security analysis tools.
Required/preferred prerequisites and qualifications: Strong Python programming skills and basic knowledge of security vulnerability concepts.
Faculty/Lab: Prof Julia Hirschberg & her PhD Students Yu-Wen Chen and Ziwei (Sara) Gong
Brief Research Project Description: 2 projects involving medical issues: one on identifying whether a patient needs to be re-admitted to the hospital from text and speech of their subsequent discussions with practitioners and one on multiple types of speech and text in voices with different medical issues (e.g. schizophrenia, anxiety, PTSD and many more).
Required/preferred prerequisites and qualifications: Some background in text and speech processing; LLMs
Faculty/Lab: Creative Machines Lab | Prof Hod Lipson, and his phd students Judah Goldfeder & Jiong Lin
Brief Research Project Description: All projects are described here.
Smart Buildings: Work on applying RL to optimize HVAC systems of commercial buildings. In partnership with Google. See: https://arxiv.org/pdf/2410.03756
Meta Learning: Develop algorithms that “learn how to learn”, where various parts of the learning process are themselves learned in a bi-level optimization.
AI For Biometrics: Develop AI algorithms for the next generation of biometric analysis. See https://www.science.org/doi/pdf/10.1126/sciadv.adi0329
AutoURDF: Unsupervised Robot Modeling from Point Cloud Videos
AutoURDF is a pipeline that automatically generates URDF (Unified Robot Description Format) files from time-series 3D point cloud data. It segments moving parts, infers the robot’s kinematic topology, and estimates joint parameters, without any ground-truth annotations or manual intervention. This makes AutoURDF a scalable and fully visual solution for automated robot modeling. Please see google doc above for full details!
Knolling: Design a robot that can look at a cluttered pile of Lego brick and sort them out neatly by color, shape and size. Use end-to-end ML
Foosball: Use reinforcement learning or other methods to train a robotic system to play on a foosball table. Start in simulation and continue in reality (physical system in development).
2D Shape vectorization: Go from a raster 2D shape (e.g. polygon) to a CSG tree of primitives (e.g. logic operations on simple shapes). If successful, apply in 3D.
Supervisory ML: Explore whether a neural network (NN) can learn to determine whether another observed neural network is confident in its answers, simply by looking at some of the observed NN internal states.
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.
Required/preferred prerequisites and qualifications: All applicants should have a strong CS background and experience with Deep learning, preferably in pytorch.
Faculty/Lab: MobileX lab | Prof Xia Zhou & Prof Salvatore Stolfo and their PhD Student Xiaofeng Yan
Brief Research Project Description: Palm vein biometrics leverage near-infrared (NIR) imaging to capture subdermal vascular patterns that are both unique and liveness-dependent, offering strong advantages over conventional traits. This project aims to support robust palm vein authentication by (1) collecting a high-variation palm vein video dataset under diverse surface, temperature, and pose conditions, and (2) developing generative models to synthesize identity-preserving palm vein videos, enabling scalable training for deep recognition models.
Required/preferred prerequisites and qualifications: – Proficiency in Python and PyTorch
– Experience with computer vision and deep learning
– Familiarity with biometrics or generative models is a plus
– Prior work with video processing or image synthesis preferred
Faculty/Lab: Prof. Junfeng Yang & Prof. Baishakhi Ray and their PhD Student Hailie Mitchell
Brief Research Project Description: We are studying LLM agents for automated program repair/debugging. Even with increasingly powerful LLMs, the performance of code agents on complex tasks remains dependent on what information, or *context*, the models have. When considering the task of code repair on repository-level projects, the entire codebase can be huge, too large to give to an LLM in a prompt. We are exploring how to identify and collect relevant context from a Python repository that enables coding agents to successfully resolve a reported bug. For a given buggy method, such context might include the class to which the buggy method belongs, code that invokes the buggy method, documentation describing intended behavior, etc. Research questions include: What code, documentation, or other repository context is necessary for an agent to understand the underlying cause of a given bug and resolve it? What strategies are the most effective for agents to retrieve such context? If agents obtain all relevant context, how much does their overall performance improve on the task of program repair?
Required/preferred prerequisites and qualifications: Proficiency in Python; strong programming/debugging skills; willingness to read research papers; (preferred) exposure to LLMs/agents.
Faculty/Lab: Prof Matthew Connelly & Dr. Raymond Hicks
Brief Research Project Description: History Lab has more than 4 million declassified government documents. We are looking for 2 students on two separate but related projects involving LLMs. The first involves fine-tuning an LLM by incorporating our documents with the aim of allowing users to ask questions about and receive answers from our data. The second project would include training a language model to correct OCR errors in our documents.
Required/preferred prerequisites and qualifications: Advanced knowledge of Python and postgres. Experience using LLMs
Faculty/Lab: Prof Kostis Kaffes and Prof Ethan Katz-Bassett, with their PhD students Meghna Pancholi, Carson Garland & Vahab Jabrayilov
Brief Research Project Description: Exercise games are a type of “serious game” that helps people move their bodies. Exercise games in extended reality (XR) encourage users to move around and interact in 3D space, a significant upgrade from interacting with a traditional 2D display. Such games have shown significant promise for physical rehabilitation; for example, XR games have helped people with Parkinson’s disease practice turning, obstacle avoidance, and problem solving during over-ground walking. A key goal is to make these games fun to play by including social interactions that boost enjoyment, leading to enhanced adherence, heightened engagement, and more rapid skill learning. Moreover, when the social interactions are enabled
between remote users, clinicians can guide and provide feedback to their patients from within the game environment as a form of tele-health, thereby increasing accessibility to rehabilitation interventions. The focus of this project is thus on exergames for multiple users with XR headsets located at the same or different sites. The project develops a richly adaptive XR platform and a framework for application traffic prioritization across layers, domains, and the public Internet. The project will pilot XR-based exergames across multiple sites, using NSF-funded community infrastructure such as the COSMOS 5G testbed and the PEERING BGP testbed. Users at each site will play the games, and performance will be assessed in terms of user experience and quality of service, with the aid of team members from the healthcare sciences, showing the rehabilitation community the transformative potential of network-supported XR.
Required/preferred prerequisites and qualifications: Basic networking knowledge or completing COMS 4119
Faculty/Lab: Prof Matthew McDermott
Brief Research Project Description: 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.
Required/preferred prerequisites and qualifications:
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.
Faculty/Lab: Prof Smaranda Muresan and her PhD student Anubhav Jangra
Brief Research Project Description: Title: Automatic Hint Generation.
Goal: Building capable hint generation systems towards improved self-study tutoring systems that emphasize on critical thinking, spanning several formal and informal logic subjects.
Student responsibilities: Towards accomplishing these goals, the students would get an experiencing developing reasoning benchmarks, automatic domain-specific evaluation metrics, and build UI/UX studies for human evaluations.
Required/preferred prerequisites and qualifications: Familiarity with deep learning, natural language processing. (preferred language python). A zeal to work in applied research and interdisciplinary problems.
Faculty/Lab: Prof Baishakhi Ray and her PhD Student Yangruibo (Robin) Ding
Brief Research Project Description: Generally, our lab focuses on fundamentally improving LLMs’ capability towards software engineering, addressing critical limitations in current LLM-based coding assistants through advanced symbolic reasoning and efficient multi-agent systems.
==Project-1: Training LLMs with advanced symbolic reasoning for program analysis==
In this project, we aim to (post-)training LLMs to reason about complex program semantics, memory management, and execution flows and equip them with advanced program analysis tools, static analyzers, and formal verification systems.
Our goal is to enable open-source LLMs with fundamental capabilities that targets to tackle software engineering challenges, such as large-scale programming and debugging, security analysis, performance optimization, and long-term maintenance.
==Project-2: Efficient multi-agent systems for collaborative software engineering==
We aim at building collaborative LLM agents to assist human developers for complicated software engineering tasks using small models (<32B parameters), making advanced agentic systems that are locally and privately deployable for companies and cheaply accessible to everyone.
Our goal is to leverage the insights from software engineering research and practice to design efficient multi-agent AI systems that can achieve GPT/Claude/Gemini-level performance in agentic software engineering tasks with 10x-100x fewer computational costs and latency.
Required/preferred prerequisites and qualifications: To facilitate the communication, please make sure to send an email to (yd2447@columbia.edu) before coming to the research fair. In the email, please attach your CV and a brief introduction about your background and why you are interested in our projects.
Solid software engineering skills and experiences in LLMs (pre-/post-training or agentic systems), program analysis, or security are strongly preferred.
Faculty/Lab: Prof Baishakhi Ray, & Prof Junfeng Yang and PhD students: Yangruibo (Robin) Ding, Jinjun Peng, Weiliang Zhao.
Brief Research Project Description: Code LLMs and Agents: Training and Evaluation of code LLMs on various coding/software engineering/security tasks including agentic tasks; the focus will be on the models rather than on pure agentic scaffold implementation
Required/preferred prerequisites and qualifications: Proficient Python programming; Hands-on model inference and training experience
Faculty/Lab: Prof. Asaf Cidon and his PhD student Tal Zussman
Brief Research Project Description: eBPF (Extended Berkeley Packet Filter) is a Linux technology that allows safely executing custom user code in the operating system kernel. eBPF has been used for customizing OS scheduling policies, accelerating network and database operations, and much more. We are developing novel applications of eBPF for memory management, virtualization, and scheduling, with the potential to accelerate use cases such as serverless computing, ML inference, and more. This project will involve extending eBPF’s capabilities in the Linux kernel and modifying production-level applications and workloads to take advantage of these new capabilities.
Required/preferred prerequisites and qualifications: Required: Proficiency in C or C++, experience with operating systems/Linux kernel programming
Preferred: Experience with eBPF
Nice to have: Experience with Rust
Faculty/Lab: Prof Xuhai (Orson) Xu | SEA Lab https://sea-lab.space/
Brief Research Project Description: At SEA Lab (https://sea-lab.space/), we design and build next-generation AI/ML and LLM systems—from single-agent to multi-agent architectures—that address real-world challenges in health and well-being.
Our projects focus on transforming personal data streams—from wearables, smartphones, home medical devices, and even electronic health records—into systems capable of complex reasoning, adaptive support, and intelligent interventions.
Our guiding vision is captured in our name: Sense, Empower, and Augment (SEA!) the health and well-being of individuals.
Opportunities for Students
If you’re interested in research at the intersection of HCI, AI, and health, we welcome you to express your interest by filling out our research opportunity form: https://forms.gle/zfHRzEqR9USm3sXV8
Required/preferred prerequisites and qualifications: Our most important requirement is passion and dedication: a genuine interest in doing high-quality research that makes real-world impact.
If you bring one or more of the following, that’s a big plus:
• Hands-on AI/ML research with time-series signals or real-time system development
• Hands-on HCI research in interaction design or user study design
• Publication experience in top AI/ML, HCI, or health venues
Faculty/Lab: Prof Asaf Cidon, Prof Junfeng Yang, and their PhD Student Haoda Wang
Brief Research Project Description: Satellite-backed services have become an essential component of everyday life, in areas such as navigation, Internet connectivity and imaging. The collapsing cost of launching to space has disrupted the way satellites are deployed, shifting the industry from a model of 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, wherein operators manually deploy software to each satellite individually.
This approach is woefully inadequate in the new emerging SmallSat operational model, where an operator needs to manage hundreds of “wimpy” satellites with varying hardware capabilities under intermittent communication. Furthermore, SmallSat operators increasingly “rent out” their infrastructure to third parties, and need to support the workloads of multiple different tenants on the same satellites, which raises the classic problems of isolation and security similar to cloud computing, but in the much more constrained hardware environment of space.
Students will help design a novel lightweight eBPF-based runtime for fleets of multi-tenant, heterogeneous and intermittently-connected satellites. The runtime will run on the FPrime flight software framework.
Required/preferred prerequisites and qualifications: Required: Experience with C or C++
Preferred: Taken operating systems – Previous experience with flight software or ground data systems
Faculty/Lab: Prof Silvia Sellán
Brief Research Project Description: Our new lab is looking for Master’s and undergraduate researchers to work on real-world problems that involve digital, three-dimensional geometry, including:
1. The Anti-Ghost-Gun certificate. Much like conventional 2D printers contain software that detects the intention to print counterfeit currency, we wish to design an algorithm that can be embedded in 3D printers to detect and stop the 3D printing of firearms. The first step in this exploratory project will be to build a baseline comparing existing geometric deep learning algorithm on this task.
2. Efficient swept volumes with continuation algorithms. A swept volume is the region of space covered by an object as it moves along a trajectory, a 3D analogue of the stroke of a brush in a two-dimensional canvas. Computing swept volumes is useful in applications from industrial design to manufacturing to 3D sculpting and digital art; however, doing so efficiently is only possible for a small set of trajectories. This project will seek to generalize the method in our prior work to multi-dimensional trajectories and robotic movements.
3. Geometry and Bouldering: Together with Prof. Steven Feiner, we want to build on Salzman et al.’s VHard Extended Reality Rock Climbing work and use geometric deep learning to predict the difficulties of bouldering routes and generate novel routes on existing climbing walls, as well as suggest sequences of hand poses to climbers in XR.
Required/preferred prerequisites and qualifications: Experience with one or several out of 3D machine learning, geometry processing, computer graphics and/or virtual and extended reality is preferred.
Faculty/Lab: Prof Mark Santolucito
Brief Research Project Description: Neurosymbolic Program Synthesis. Combine automated reasoning tools with ML to create cost-efficient reasoning models. Tackling problems like arc-agi (https://arcprize.org/arc-agi)
Required/preferred prerequisites and qualifications: Familiarity with automated reasoning tools like SMT solvers.
Faculty/Lab: Prof Kathy McKeown, and her PhD students Amith Ananthram, & Nick Deas, and Post Doc Milad Alshomary.
Brief Research Project Description:
Analyzing and improving VLM’s ability to perform formal analysis of style:
This project aims, on the one hand, to evaluate the ability of VLMs to analyze the style of art pieces (paintings, architecture, etc.) and whether these analyses are faithful to the visual features of the art piece. On the other hand, we will work on developing methods that improve VLM’s ability to correctly capture the visual features of art pieces and use that information to generate formal analyses of artwork as performed by art historians.
Understanding and Summarizing Demographic Perspectives:
This project aims to evaluate and improve LLMs’ ability to summarize perspectives (i.e., political and social attitudes) of various demographic groups (e.g., political party, race, geographic location) on contentious issues. Because identity and culture can influence language, the project will also involve improving LLMs’ understanding of attitudes expressed in the language of underrepresented groups.
Evaluating Impressions & Stereotypes in LLMs:
LLMs have been shown to be sensitive to characteristics of prompts such as politeness or emotional language and also tend to align with particular cultural perspectives over others. Expanding on this work, this project aims to evaluate LLMs’ sensitivity to prompts, biases, and cultural alignment drawing on both psychological theories of impressions and stereotypes and mechanistic interpretability approaches to understanding LLMs.
Those interested in any of these projects can submit their information here.
Required/preferred prerequisites and qualifications: Familiarity with deep learning and NLP. For the project working with VLMs, CV is a plus!
Faculty/Lab: Prof Baishakhi Ray and her PhD Student Yangruibo (Robin) Ding
Brief Research Project Description: Generally, our lab focuses on fundamentally improving LLMs’ capability towards software engineering, addressing critical limitations in current LLM-based coding assistants through advanced symbolic reasoning and efficient multi-agent systems.
==Project-1: Training LLMs with advanced symbolic reasoning for program analysis==
In this project, we aim to (post-)training LLMs to reason about complex program semantics, memory management, and execution flows and equip them with advanced program analysis tools, static analyzers, and formal verification systems.
Our goal is to enable open-source LLMs with fundamental capabilities that targets to tackle software engineering challenges, such as large-scale programming and debugging, security analysis, performance optimization, and long-term maintenance.
==Project-2: Efficient multi-agent systems for collaborative software engineering==
We aim at building collaborative LLM agents to assist human developers for complicated software engineering tasks using small models (<32B parameters), making advanced agentic systems that are locally and privately deployable for companies and cheaply accessible to everyone.
Our goal is to leverage the insights from software engineering research and practice to design efficient multi-agent AI systems that can achieve GPT/Claude/Gemini-level performance in agentic software engineering tasks with 10x-100x fewer computational costs and latency.
Required/preferred prerequisites and qualifications: To facilitate the communication, please make sure to send an email to (yd2447@columbia.edu) before coming to the research fair. In the email, please attach your CV and a brief introduction about your background and why you are interested in our projects.
Solid software engineering skills and experiences in LLMs (pre-/post-training or agentic systems), program analysis, or security are strongly preferred.
Faculty/Lab: Prof Julia Hirschberg & her PhD Students Yu-Wen Chen and Ziwei (Sara) Gong
Brief Research Project Description: Medical Speech Applications
These projects investigate speech and text from individuals with various medical conditions (e.g., schizophrenia, anxiety, PTSD, and others). The goal is to better understand how different disorders manifest in vocal and linguistic patterns, enabling more effective healthcare applications.
Speech Assessment for Pronunciation
We aim to evaluate the capabilities of current audio language models (ALMs) in speech pronunciation assessment. The study focuses on how model outputs vary under different prompting strategies, identifying both strengths and limitations.
Required/preferred prerequisites and qualifications: Proficient in Python programming, with knowledge of machine learning and a background in speech and text processing, including large language models (LLMs).
Faculty/Lab: Prof. Asaf Cidon and Prof. Ethan Katz-Bassett
Brief Research Project Description: Study the usage of LLM in email security with access to large scale and real data. Previous research featured in Forbes news: https://www.forbes.com/sites/daveywinder/2025/06/20/ai-is-behind-50-of-spam—and-now-its-hacking-your-accounts/
Required/preferred prerequisites and qualifications: Experience in NLP (at least the intro course) and data analysis (e.g. spark, clustering, topic modeling).
Faculty/Lab: Prof Junfeng Yang & his PhD Student Andreas Kellas
Brief Research Project Description: When you download and use an ML model, do you know how it was made? What data was used to train it, and what other models it was derived from? Would you want to know if the model was derived from a model with a security flaw, like a backdoor?
Today, models are reused and adapted from other models, because training models from scratch is very expensive. Model ecosystems, like Hugging Face, allow users to upload models, but do not require users to state how the model was created. This project aims to recover model lineage information that describes how a model was derived from base models. This information can be used to enable security analyses like identifying license and IP violations, the propagation of backdoors and vulnerabilities, and validating that the model is what it is claimed to be.
Required/preferred prerequisites and qualifications: Experience programming in Python;
Experience with using open-weights models from Hugging Face;
Experience adapting and modifying models to solve specific tasks.
Faculty/Lab: Prof Blei and his PhD Student Sweta Karlekar
Brief Research Project Description: Mechanistic interpretability aims to understand how neural networks represent concepts internally, and sparse autoencoders (SAEs) have become a central tool for this work. SAEs produce human-readable features of model activations, but their descriptions—typically generated by language models from examples of highly activating inputs—are often vague or inaccurate. Since these descriptions are used in downstream tasks such as auditing model behavior and identifying risky or useful features, improving their quality is critical. This project explores the use of evolutionary optimization techniques, inspired by FunSearch and the recent AlphaEvolve paper, to iteratively refine feature descriptions by evolving LLM-generated outputs. Candidate descriptions are evolved over multiple rounds, with the “interpretability score”—the degree to which a description predicts new high-activation inputs—serving as the evaluation metric. The goal is to assess whether such techniques can systematically improve SAE feature descriptions and thereby strengthen the reliability of mechanistic interpretability tools.
Required/preferred prerequisites and qualifications: Required:
– Proficiency in Python, especially regarding concurrency models and async programming
– Experience using PyTorch
– Familiarity with version control (Git)
– Completed coursework and projects in machine learning fundamentals (e.g., representation learning, optimization)
– Prior exposure to mechanistic interpretability concepts (through courses, papers, or tools like Gemma Scope or Neuronpedia)
Preferred:
– Experience working with large language models (LLMs) or sparse autoencoders (SAEs)
– Familiarity with mechanistic interpretability research literature
– Background in optimization methods (e.g., evolutionary algorithms, reinforcement learning)
– Prior involvement in a research project or open-source ML project
Faculty/Lab: PhD Student Elisavet Alvanaki, under the supervision of her Advisor Prof Luca Carloni
Brief Research Project Description: Leveraging Multimodal Checkpointing for Scalable AI Debugging:
The project involves agentic data generation from simulation of SoCs. Students would create an AI agent to autonomously navigate a Modelsim simulation through the GUI, and retrieve multimodal hardware data for each part of the workload execution.
Required/preferred prerequisites and qualifications: Basic hardware knowledge required
Python knowledge required
Verilog knowledge required
Experience with building agents preferred
Experience with ModelSim preferred
Experience with the ESP platform (https://www.esp.cs.columbia.edu/) preferred
Experience with AI (particularly Contrastive Learning) would be nice to have
Faculty/Lab: Mobile X Lab | Prof Xia Zhou and Prof Changxi, with their PhD Students
Hadleigh Schwartz
Brief Research Project Description: We have several projects in privacy and security, mostly focusing on proactive deepfake detection and anti-spoofing. We take a hardware and systems-oriented approach to these problems. Some examples of our current and planned work are:
1) Developing systems that watermark light and/or sound in an environment, to ensure all videos recorded at the scene contain provenance information
2) Developing a method for detecting when voice clones are being used to spoof a voice authentication system by injecting ultrasound probes during the interaction
Required/preferred prerequisites and qualifications: Preferred qualifications:
– Can dedicate at least 10 hours per week to the project
– Proficient in a programming language (preferably Python)
– Plus one or more of the below:
– Taken Computer Vision 1: First Principles, Computer Vision 2, or equivalent courses
– Experience with audio signal processing
– Experience developing and training models in PyTorch
– Experience with circuits and/or hardware prototyping
Ziang Ren
Brief Research Project Description: 3D vision for object counting
Required/preferred prerequisites and qualifications: Computer science students, if students took computer vision class
Ho Man Coleman Leung
Brief Research Project Description: We work on mobile computing systems that intersect with health, aiming to make sensing technologies more seamless and user-friendly
Project 1: Noninvasive Glucose Sensing with Polarized Light
Diabetes is a chronic disease that affects millions worldwide, yet most glucose monitoring methods are still invasive and inconvenient. Our lab is developing a new approach that exploits glucose’s optical activity to noninvasively measure blood glucose level. In this project, you will help design and carry out experiments to test this method, investigate how other substances may interfere with readings, and apply machine learning to extract reliable glucose values. Students with backgrounds in data analysis and signal processing are encouraged to apply, and having chemistry knowledge is a plus.
Project 2: Fabric-Based Physiological Sensing
Monitoring signals like ECG and EEG is crucial for health applications, but current sticky electrodes are uncomfortable and not suitable for long-term use. We are exploring fabric-based electrodes that can be seamlessly embedded into everyday objects such as pillows, bedding, or clothing to capture signals in natural settings. In this project, you will contribute to prototyping fabric-based sensors, collecting and analyzing physiological data, and developing algorithms that make long-term sensing both reliable and comfortable. Sewing or embroidery experience is a bonus.
Interested students can submit their information here: https://forms.gle/z98mxpEXgWxvYQtH8
Required/preferred prerequisites and qualifications: We are looking for self-motivated students with a strong interest in research and learning new things. Proficient in Python and with experience in data analysis or signal processing are preferred. Bonus skills include electrical circuit design, chemistry, and sewing/embroidery.
Faculty/Lab: Prof Venkat Venkatasubramanian and his PhD Students
Brief Research Project Description:
Google Sign up – Link Here
Ricky (kk3764 @columbia.edu)
From Data to Differential Equations: Interpretable Modeling
Develop hybrid models combining symbolic regression and ML to discover governing ODEs from data. Uses genetic algorithms with domain knowledge to uncover interpretable mechanisms in systems like chemistry, signaling, and electrochemistry.
Self-Organizing Dynamics in Active Matter with RL
Apply physics-informed reinforcement learning to model active matter agents whose local strategies give rise to macroscopic behaviors like pattern formation and phase transitions. Bridges statistical mechanics, game theory, and emergent dynamics.
Naz (nt2513 @columbia.edu)
Mechanistic Interpretability – Inner Workings of LLMs
Train sparse autoencoders (SAEs) on LLM representations to study “monosemantic” features and hierarchical organization of knowledge. Includes visualization tools for large-scale interpretability using GPU clusters.
Ontology-based Information Extraction (SUSIE)
Build SUSIE, an ontology-based tool that converts text into knowledge graphs, enhancing retrieval via COPE. Next steps: ontology population at scale, neural-symbolic inference, and safe scientific hypothesis generation.
Mechanistic Interpretability of AlphaFold2
Probe AlphaFold2’s internal embeddings to uncover laws of biological representation, geometric reasoning, and inference for drug discovery. Extend preliminary results to explore how transformer models encode biological knowledge.
Collin (cjs2301 @columbia.edu)
Causal Manipulation in AlphaFold
Test whether AlphaFold’s internal features causally influence protein structure predictions. Focus on mutations, folding dynamics, and mechanistic explanations for disease-related changes.
Multi-domain Folding and Ligand Binding in AlphaFold
Study how AlphaFold multimer represents domain-domain interactions and ligand binding. Goal: understand conformational shifts and biomolecular signaling at complex scales.
Comparative Analysis of Biomolecular AI Systems
Compare models like AlphaFold, RoseTTAFold, and Boltz-1, as well as function-level tools like DeepGo-SE. Explore synergies and integrate LLMs/ontologies into pipelines for protein science and drug design.
Modeling Microrobotic Dynamics (Sanjeev and Collin)
Model magnetic microrobots that self-assemble and coordinate under external fields. Use statistical/game-theoretic frameworks to study swarm behaviors like reconnaissance, transport, and collective pattern formation.
Sanjeev (sn3130 @columbia.edu)
Large Knowledge Models (LKM/Hybrid Models)
Develop architectures that embed domain laws (e.g., conservation principles) into LLMs for engineering tasks. Explore hybrid approaches that go beyond RAG/prompting to build Large Knowledge Models.
Dynamics of Tokens in LLMs
Study LLM embeddings and sparse autoencoders to analyze token-level knowledge representation. Apply statistical/game-theoretic frameworks to understand how tokens encode meaning in high-dimensional space.
Pattern Interpretability in Neural Networks
Investigate log-normal weight distributions in trained neural networks as an “ideal state.” Use them for initialization to reduce training time, improve generalization, and study topological data separability.
Statistical Risk Analysis
Automate risk analysis in chemical, financial, and supply-chain systems using probabilistic failure models. Risk is modeled as likelihood × severity to support predictive safety frameworks.
Required/preferred prerequisites and qualifications: Juniors, Seniors, or M.S. students with a background in Python programming and machine learning fundamentals. An interest in deep learning and curiosity about applying AI to scientific domains (e.g., drug discovery) are highly encouraged, though no prior experience in these fields is required. Students will collaborate in teams of 5–6 alongside PhD mentors.
Zoom
Faculty/Lab: Prof Shalmali Joshi
Brief Research Project Description: Our lab is building new foundation models for health and medicine data. We aim to improve their generalization, and robustness properties using principles from probabilistic modeling, causal inference, and/or reinforcement learning. The specific subtopic of your project will likely be multimodal deep learning integrating principles of causal inference.
Required/preferred prerequisites and qualifications: Strong ML foundations in one or more of deep learning, probabilistic and causal inference, reinforcement learning will make you suitable for collaborating with the group
Zoom: 9/4/2025 11:00
https://columbiacuimc.zoom.us/my/shalmalij
Faculty/Lab: Prof Junfeng Yang and his PhD student Chengzhi Mao
Brief Research Project Description: We work on understanding and enhancing LLM, VLLM, and generative models. We now explore RL, LLM thinking, and Multimodal generative models.
A few prior projects, first-authored and published by our previous Columbia students:
1. selfie for LLM: https://selfie.cs.columbia.edu/ ICML 2024
2. Interpreting Vision language model: https://arxiv.org/abs/2310.10591, ICLR 2024,
3. LLM safety: https://arxiv.org/pdf/2411.04223, NAACL 2024
4. Training LLM: https://arxiv.org/pdf/2408.04237? ACL 2025
Required/preferred prerequisites and qualifications: Pytorch, Machine Learning, willing to work hard and lead a research project
Zoom: 9/6/2025 15:00 (3pm)
https://us04web.zoom.us/j/79901748725?pwd=bSUVRzsJEkJqJG25fqgRKU2qLAZ4s3.1
Faculty/Lab: Computer Graphics & 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 and social work.
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. For those projects involving 3D user interfaces, we’re especially interested in students with Unity experience.
Zoom: 9/4/2025 4 PM
https://columbiauniversity.zoom.us/j/92276206657?pwd=GDldqQbIzklkjRgWDqwdJ5SdBBufqk.1
Faculty/Lab: SNL | Prof. Ethan Katz-Bassett and PhD students Loqman Salamatian & Shuyue Yu
Brief Research Project Description: Full details here.
Our group studies the Internet’s performance, topology, and security through large-scale measurement and data analysis. Current projects include (non-exhaustive list!):
-
HERMES: Detecting and Localizing Internet Performance Degradations from End-User Measurements – HERMES analyzes speed test and traceroute data from Measurement Lab (M-Lab) to identify and localize Internet slowdowns and disruptions as experienced by real users. By correlating performance metrics with network topology, HERMES produces a bottom-up view of Internet health and can pinpoint problems to specific network segments. Ongoing work includes quantifying packet loss as a detection signal, building interactive visualizations, detecting large-scale events, creating operator accountability scores, comparing IPv4/IPv6 performance, and enabling near real-time incident detection.
-
Filecoin: Measuring Decentralized Storage Node Performance – In collaboration with Protocol Labs, we analyze data from their live Bandwidth Measurement System (BMS), which measures the data retrieval rates of Filecoin storage nodes. Students working on this project would help improve the measurement methodology, detecting shared backends and node aliasing, designing scalable measurement strategies, comparing performance from different client locations, and developing better aggregation and estimation techniques.
-
MetAScritic: Inferring Missing AS-Level Links at the Metro Scale – Many interconnections between networks are invisible to public data sources, limiting our understanding of resilience and routing. MetAScritic uses a recommender system approach to predict unseen metro-level AS links and augments public data with targeted active measurements, producing more complete topologies for analysis of peering strategies, hijack resilience, and regional connectivity. Students working on this project would help design and train the inference models, plan and run targeted measurements, and analyze the resulting data to evaluate how the inferred links change our understanding of network structure and resilience.
-
Unsolicited Traffic Analysis – Unsolicited traffic refers to network communications that were not requested by the receiving system, such as scanning attempts, brute-force logins, malware propagation, or other potentially malicious probes. Honeypots capture unsolicited traffic by attracting inbound connections to decoy systems, while passive monitoring systems observe unsolicited flows as part of broader network traffic. Each method has different coverage and biases, so combining them can reveal a more complete threat landscape. Students working on this project would compare datasets from both sources, identify overlaps and differences in the types of malicious activity observed, and design techniques for integrating these complementary views to improve threat detection.
-
Detecting DNS-Based Exfiltration in Passive DNS Traces – Certain malware families exfiltrate stolen banking or credit card data via DNS queries. Students working on this project analyze passive DNS traces to detect suspicious patterns, such as unusual query lengths or random domains, and cross-reference findings with known threat reports.
Required/Preferred Prerequisites and Qualifications:
-
Preferred: Networking course (e.g., COMS 4119 or equivalent)
-
Preferred: Linear Algebra
-
Preferred: Strong interest in Internet measurement, networking, or security
Link: Friday the 5th at 4:00 PM via Google Meet Google Meet
Faculty/Lab: TBA
Brief Research Project Description: TBA
Required/preferred prerequisites and qualifications: TBC
Zoom: TBA
Tables
Thursday:
CS Advising!
Stop by the window booths for QQ (Quick Question) Hours and info sessions.
GrOCS!
The Graduate Organization of Computer Science (GrOCS) serves as the EGSC student group for graduate students in the Computer Science department. The organization aims to enhance the graduate experience and promote a collaborative, inclusive environment through events, workshops, and social initiatives. It also provides a platform for students to connect, share ideas, and build lasting relationships within the computer science community.
Friday:
CS Careers!
Come to learn more about CS Careers, and answer some quiz questions to win prizes.
Add to GMail