Adi Shnaidman |
|
|
I’m an AI researcher and MSE student in Computer and Information Science at the University of Pennsylvania, where I work with Prof. Dan Roth. My research studies how language models and AI agents acquire, retain, and adapt behavior under changing contexts and constraints. My work spans inference-time model control, history-dependent behavior in multi-turn interactions, mechanistic analysis, and reliable adaptive agents, with a broader focus on making advanced AI systems more predictable, robust, and controllable. Philadelphia, PA · University of Pennsylvania |
|
News
- Invited Talk: Inference-Time Control of Diffusion Language Models at ML Collective's Deep Learning: Classics and Trends (DLCT) .
- Reset Is Not Recovery: Evaluating Recoverability from False Conversational Context via Sycophancy Hysteresis was accepted to the GroundLM Workshop at EMNLP 2026.
- Joined the University of Pennsylvania as an MSE student in Computer and Information Science.
- Activation Steering for Masked Diffusion Language Models was accepted to the ICLR 2026 ReALM-GEN Workshop.
Research
I study the reliability and controllability of language models and AI agents, focusing on how their behavior is shaped by internal representations, interaction history, and changing task and capability constraints. My research investigates how model behavior emerges, persists, and changes over time, and how targeted interventions and systematic evaluations can help understand and mitigate failure modes.
Current research themes:
- AI Agent Reliability and Safety
- Inference-time control and representation-level interventions
- Behavioral Evaluation and Robustness
- Language Model Behavioral Dynamics
Selected Research
|
Reset Is Not Recovery: Evaluating Recoverability from False Conversational Context via Sycophancy Hysteresis
Studies whether language models recover from false conversational context after user pressure is withdrawn, introducing sycophancy hysteresis and a hierarchy of recovery interventions. |
|
Activation Steering for Masked Diffusion Language Models
Introduces an inference-time activation steering method for masked diffusion language models, using activation-level interventions to steer model behavior during reverse diffusion. |
|
Structure–Function Tradeoffs in the Human Brain
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
Contributed to research on plain-text prompting strategies that induce excessive LLM generation without jailbreaks or policy circumvention. |
Education
Experience
-
Aug 2026 – PresentUniversity of PennsylvaniaGraduate ResearcherResearch with Prof. Dan Roth on reliable and trustworthy AI agents, including strategic capability dependence and adaptive agent behavior.
-
Apr 2026 – PresentOxford Robotics Institute, University of Oxford — PNPLResearch CollaboratorResearch on representation alignment for MEG-to-text generation, connecting neural representations with language-model representation spaces and studying conditioning and decoding behavior.
-
Apr 2025 – Sep 2026DeepKeepMachine Learning ResearcherResearch on LLM and AI-agent safety, including systematic adversarial evaluations of prompt injection, encoding-based attacks, and multi-turn attacks. Led three first-author research projects spanning inference-time model control, multi-turn behavioral persistence and recoverability, and mechanistic analysis of history-dependent model behavior.
-
Oct 2023 – May 2025ComDePri Lab, The Hebrew University of JerusalemData Scientist & Capstone ResearcherResearch across computational neuroscience and time-series modeling, including structure–function analysis and reconstruction of high-frequency human movement data.
-
Aug 2016 – Oct 2021Confidential OrganizationResearch & Analytics LeadershipLed research and analytics teams supporting high-stakes decision-making, progressing through research, analytics, and leadership roles.
Selected Honors & Awards
-
Monday.com–HUJI Excellence Scholarship
Selected among the top 30 students in the School of Computer Science & Engineering for outstanding academic and personal achievement.
-
Recognition for Significant Volunteer Contribution
Selected among 100 students recognized by the university for significant volunteer contributions.
Academic Service
- NeurIPS 2026 · InterpScience Workshop
- NeurIPS 2026 · VLM4RWD Workshop
- EMNLP 2026 · GroundLM Workshop