Skip to content
View galwert's full-sized avatar

Highlights

  • Pro

Block or report galwert

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
galwert/README.md

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.

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

Pinned Loading

  1. AnchorRep AnchorRep Public

    Defending LLMs against cross-model adversarial transfer via representation repulsion (NeurIPS 2026)

    Python 1

  2. 048954-statistical-methods-image-processing 048954-statistical-methods-image-processing Public

    Statistical Methods in Image Processing (EE 048954): normalizing flows, EBMs, score matching

    Jupyter Notebook

  3. backdoor-defense-quantization backdoor-defense-quantization Public

    Do backdoor defenses survive quantization? Fine-Pruning, ONION, STRIP, Neural Cleanse x PTQ/QAT/LAQ on DistilBERT

    Python

  4. sam3d-mechanistic-interpretability sam3d-mechanistic-interpretability Public

    SAE steering, circuit discovery and hallucination correction inside SAM 3D Objects (1.2B image-to-3D model)

    Python

  5. sar-voidcrawler sar-voidcrawler Public

    SAR VoidCrawler: navigation and perception stack for a search-and-rescue crawler robot (synthetic LiDAR SLAM, survivor scan, SAM2 rubble analysis). Hackathon 2026

    Python

  6. structml-relational-gnn-ablations structml-relational-gnn-ablations Public

    Ablation study of relational GNNs on RelBench: directionality, heterogeneity, encoders, depth (StructML 236605 final project)

    Python