Martin Vechev

Full Professor@ETHZ - ETH Zurich·瑞士·OpenReview

研究方向

safe and secure AI · AI for mathematics · LLM security · AI for programming languages and program analysis · AI and symbolic reasoning · combining logic and AI · trustworthy AI

Martin Vechev

The Open Proof Corpus: A Large-Scale Study of LLM-Generated Mathematical Proofs

Watch your steps: Dormant Adversarial Behaviors that Activate upon LLM Finetuning

LLM Fingerprinting via Semantically Conditioned Watermarks

Fewer Weights, More Problems: A Practical Attack on LLM Pruning

Constrained Decoding of Diffusion LLMs with Context-Free Grammars

BrokenMath: A Benchmark for Sycophancy in Theorem Proving with LLMs

Dual Randomized Smoothing: Beyond Global Noise Variance

Watermarking Diffusion Language Models

Expressiveness of Multi-Neuron Convex Relaxations in Neural Network Certification

ToolFuzz: Automated Agent Tool Testing

Pay Attention to the Triggers: Constructing Backdoors That Survive Distillation

Adaptive Generation of Bias-Eliciting Questions for LLMs

AutoBaxBench: Bootstrapping Code Security Benchmarking

Watermarking Autoregressive Image Generation

MixAT: Combining Continuous and Discrete Adversarial Training for LLMs

Black-Box Detection of Language Model Watermarks

MathConstruct: Challenging LLM Reasoning with Constructive Proofs

Polyrating: A Cost-Effective and Bias-Aware Rating System for LLM Evaluation

Ward: Provable RAG Dataset Inference via LLM Watermarks

Discovering Spoofing Attempts on Language Model Watermarks

A Unified Approach to Routing and Cascading for LLMs

Black-Box Adversarial Attacks on LLM-Based Code Completion

Mind the Gap: A Practical Attack on GGUF Quantization

Language Models are Advanced Anonymizers

Discovering Clues of Spoofed LM Watermarks

GRAIN: Exact Graph Reconstruction from Gradients

Average Certified Radius is a Poor Metric for Randomized Smoothing

Black-Box Adversarial Attacks on LLM-Based Code Completion

A Unified Approach to Routing and Cascading for LLMs

Automated Benchmark Generation for Repository-Level Coding Tasks

BaxBench: Can LLMs Generate Correct and Secure Backends?

CTBench: A Library and Benchmark for Certified Training

AlphaIntegrator: Transformer Action Search for Symbolic Integration Proofs

CTBench: A Library and Benchmark for Certified Training

Average Certified Radius is a Poor Metric for Randomized Smoothing

Evading Data Contamination Detection for Language Models is (too) Easy

Multi-Neuron Unleashes Expressivity of ReLU Networks Under Convex Relaxation

Gaussian Loss Smoothing Enables Certified Training with Tight Convex Relaxations