Adi Shnaidman

I’m an AI researcher and MSE student in Computer and Information Science at the University of Pennsylvania, where I conduct research under the supervision of Prof. Dan Roth. My work focuses on large language models, particularly safety, robustness, mechanistic interpretability, and inference-time control.

I study how internal representations relate to model behavior, how failures emerge under adversarial and real-world conditions, and how representation-level interventions can be used to evaluate and steer LLMs.

My recent work at DeepKeep has focused on adversarial robustness, model evaluation, and real-time detection and mitigation for LLMs and AI agents.

Philadelphia, PA · University of Pennsylvania

News

Joined the University of Pennsylvania as an MSE student in Computer and Information Science.
August 2026 · Philadelphia, PA
Activation Steering for Masked Diffusion Language Models was accepted to the ICLR 2026 ReALM-GEN Workshop.
April 2026 · Rio de Janeiro, Brazil

Research Interests

My research focuses on understanding and controlling the behavior of large language models through their internal representations.

I’m particularly interested in identifying representation-level signals associated with harmful, unreliable, or non-robust behavior, and in understanding how these signals emerge and evolve during inference.

A central direction in my work is inference-time control: using internal model representations to characterize behavioral states and develop principled interventions that steer model behavior without retraining.

Research interests:

LLM safety, robustness, and behavioral reliability
Mechanistic interpretability and representation analysis
Inference-time control and activation steering
Model evaluation under adversarial and distribution-shift conditions

Selected Research

Activation Steering for Masked Diffusion Language Models
A. Shnaidman, E. Feiglin, O. Yaari, E. Mentel, A. Levi, R. Lapid
ICLR 2026 ReALM-GEN Workshop

Introduces an inference-time activation steering method for masked diffusion language models, using activation-level interventions to steer model behavior during reverse diffusion.

Related & follow-up work

Gert Lek et al., An Optimal Transport View of Activation Steering in Masked Diffusion Models
Amit LeVi et al., You Had One Job: Per-Task Quantization Using LLMs’ Hidden Representations
Amit LeVi et al., Jailbreak Attack Initializations as Extractors of Compliance Directions
Hyeseon An et al., DLM-SWAI: Steering Diffusion Language Models Before They Unmask
Arth Singh, Re-Mask and Redirect: Exploiting Denoising Irreversibility in Diffusion Language Models
Ryandito Diandaru et al., TimpaTeks: Automatic In-place Text Sequence Modification via Diffusion Language Model Steering
Shuai Zhang et al., Low Perplexity is Repetition: A One-Dimensional Self-Conditioning Attractor in Continuous Diffusion LMs
Amit LeVi et al., Unsupervised Features Mining via Activation Geometry

Structure–Function Tradeoffs thumbnail (optional)

Structure–Function Tradeoffs in the Human Brain
Unpublished · 2024
A. Shnaidman, A. Meisel, O. Ventura, R. Schurr, Y. Hart

Applies Pareto Task Inference (ParTI) to high-dimensional structural MRI features from the Human Connectome Project to probe structure–function relationships.

BenchOverflow: Measuring Overflow in Large Language Models via Plain-Text Prompts
Erin Feiglin, Nir Hutnik, Raz Lapid
TMLR · 2025
Acknowledged contributor

Contributed to research on plain-text prompting strategies that induce excessive LLM generation without jailbreaks or policy circumvention.

Education

University of Pennsylvania

Master of Science in Engineering (MSE) · Computer and Information Science

2026 – 2028

Research under the supervision of Prof. Dan Roth.

The Hebrew University of Jerusalem

Bachelor of Science (B.Sc.) · Computer Science and Statistics · Data Science Track

Mentored by Prof. Matan Gavish, School of Computer Science and Engineering, and Prof. Asaf Weinstein, Department of Statistics and Data Science.

Professional Experience

Apr 2025 – Sep 2026

DeepKeep

Machine Learning Researcher

Research on LLM safety and robustness, with a focus on adversarial evaluation, model behavior under failure-inducing conditions, and inference-time detection and mitigation for LLMs and AI agents.
Oct 2023 – May 2025

ComDePri Lab, The Hebrew University of Jerusalem

Data Scientist (Jun 2024 – May 2025)

Developed statistical and machine-learning methods for modeling and reconstructing high-frequency human movement time series, including predictive modeling and data mining.

Capstone Researcher & Data Scientist (Oct 2023 – Oct 2024)

Conducted research on structure–function trade-offs in the human brain using dimensionality reduction, archetypal analysis, and Pareto Task Inference (ParTI).
Aug 2016 – Oct 2021

Large Enterprise Organization, Confidential

Principal Research & Data Lead (Dec 2020 – Oct 2021)

Led research and data-analysis efforts, translated complex findings into strategic recommendations, and established analytical standards.

Head of Research & Analytics (Feb 2019 – Dec 2020)

Led analytics teams and large-scale data analysis to support executive decision-making.

Data Research Analyst (Aug 2016 – Feb 2019)

Conducted quantitative research and data analysis to support organizational decision-making.

Selected Honors & Awards

Monday.com–HUJI Excellence Scholarship
Monday.com & The Hebrew University of Jerusalem · 2024

Selected among the top 30 students in the School of Computer Science & Engineering for outstanding academic and personal achievement.
Recognition for Significant Volunteer Contribution
The Hebrew University of Jerusalem · 2024

Selected among 100 students recognized by the university for significant volunteer contributions.

Academic Service

GroundLM Workshop at EMNLP 2026
Reviewer · 2026