Skip to content
galwertPublic

About

Gal Wertheizer: ML safety and interpretability research

Resources

Stars

0 stars

Watchers

0 watching

Forks

Latest commit

 

History

5 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 

Repository files navigation

Gal Wertheizer

ML researcher and engineer. I work on the safety and interpretability of large models: adversarial robustness of LLMs, mechanistic interpretability of generative models, and learning on relational data.

Contact: wertheizer@campus.technion.ac.il · Models: huggingface.co/galwert

Featured: AnchorRep (NeurIPS 2026)

AnchorRep: a defense against cross-model jailbreak transfer. A lightweight LoRA adapter pushes the defended model's internal representations of harmful prompts away from a frozen anchor model, cutting cross-model adversarial transfer to under 2% on five open-weight LLMs without attack-specific training. Full training and evaluation code, configs, attack artifacts and audit logs, plus five trained adapters on Hugging Face.

Research projects

Repository What it is
sam3d-mechanistic-interpretability Sparse autoencoders, circuit discovery and activation steering inside SAM 3D Objects, a 1.2B image-to-3D generator. Finds a three-feature circuit controlling geometry and shows object identity is not linearly represented.
structml-relational-gnn-ablations Ablation study of relational GNNs on RelBench: edge directionality, heterogeneity, column encoders and depth.
backdoor-defense-quantization Do backdoor defenses survive quantization? Four defenses, three quantization schemes, both orderings, on DistilBERT.

Hackathon

sar-voidcrawler: navigation and perception stack for a search-and-rescue crawler robot. Synthetic LiDAR SLAM, goal-directed navigation with a 360 degree survivor scan, human-part LiDAR segmentation, and SAM2-based rubble analysis, all runnable on CPU without a robot.

Coursework (Technion)

Course Repository What's inside
236605 Structured Machine Learning 236605-structured-ml Relational feature extraction and node embeddings on RelBench. Final project: structml-relational-gnn-ablations
048954 Statistical Methods in Image Processing 048954-statistical-methods-image-processing Normalizing flows, energy-based models, score matching
236299 Introduction to NLP 236299-intro-to-nlp Text classification, slot-filling taggers, CKY parsing, semantic parsing for QA
236360 Theory of Compilation 236360-theory-of-compilation A compiler for a C-like language, from flex lexer to LLVM IR with backpatching
234123 Operating Systems 234123-operating-systems smash, a Linux shell in C++ with job control, pipes and signals
234218 Data Structures 1 234218-data-structures-1 AVL and rank-augmented trees, union-find, five wet assignments
234124 Introduction to Systems Programming 234124-intro-to-systems-programming C++ templates, iterators, exceptions, a turn-based game
234118 Computer Organization 234118-computer-organization x86-64 assembly with raw Linux syscalls
Computer architecture computer-architecture-simulators Cache hierarchy and branch predictor simulators

About

Gal Wertheizer: ML safety and interpretability research

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors