Search

Jailbreaking large-language models (LLMs) involves testing their robustness against adversarial prompts and evaluating their ability to withstand prompt attacks that could elicit unauthorized or malicious responses. In this paper, we present TurboFuzzLLM, a mutation-based fuzzing technique for efficiently finding a collection of effective jailbreaking templates that, when combined with harmful questions, can lead a target LLM to produce harmful responses through black-box access via user prompts. We describe the limitations of directly applying existing template-based attacking techniques in practice, and present functional and efficiency-focused upgrades we added to mutation-based fuzzing to generate effective jailbreaking templates automatically. TurboFuzzLLM achieves $\geq$ 95\% attack success rates (ASR) on public datasets for leading LLMs (including GPT-4o \& GPT-4 Turbo), shows impressive generalizability to unseen harmful questions, and helps in improving model defenses to prompt attacks., Comment: Accepted at NAACL 2025 industry track, 12 pages, 5 figures

The second workshop on the HEP Analysis Ecosystem took place 23-25 May 2022 at IJCLab in Orsay, to look at progress and continuing challenges in scaling up HEP analysis to meet the needs of HL-LHC and DUNE, as well as the very pressing needs of LHC Run 3 analysis. The workshop was themed around six particular topics, which were felt to capture key questions, opportunities and challenges. Each topic arranged a plenary session introduction, often with speakers summarising the state-of-the art and the next steps for analysis. This was then followed by parallel sessions, which were much more discussion focused, and where attendees could grapple with the challenges and propose solutions that could be tried. Where there was significant overlap between topics, a joint discussion between them was arranged. In the weeks following the workshop the session conveners wrote this document, which is a summary of the main discussions, the key points raised and the conclusions and outcomes. The document was circulated amongst the participants for comments before being finalised here.

Synthesizing multi-character interactions is a challenging task due to the complex and varied interactions between the characters. In particular, precise spatiotemporal alignment between characters is required in generating close interactions such as dancing and fighting. Existing work in generating multi-character interactions focuses on generating a single type of reactive motion for a given sequence which results in a lack of variety of the resultant motions. In this paper, we propose a novel way to create realistic human reactive motions which are not presented in the given dataset by mixing and matching different types of close interactions. We propose a Conditional Hierarchical Generative Adversarial Network with Multi-Hot Class Embedding to generate the Mix and Match reactive motions of the follower from a given motion sequence of the leader. Experiments are conducted on both noisy (depth-based) and high-quality (MoCap-based) interaction datasets. The quantitative and qualitative results show that our approach outperforms the state-of-the-art methods on the given datasets. We also provide an augmented dataset with realistic reactive motions to stimulate future research in this area. The code is available at https://github.com/Aman-Goel1/IMM, Comment: Accepted to SCA 2022 (will be published in CGF)

Lamport's celebrated Paxos consensus protocol is generally viewed as a complex hard-to-understand algorithm. Notwithstanding its complexity, in this paper, we take a step towards automatically proving the safety of Paxos by taking advantage of three structural features in its specification: spatial regularity in its unordered domains, temporal regularity in its totally-ordered domain, and its hierarchical composition. By carefully integrating these structural features in IC3PO, a novel model checking algorithm, we were able to infer an inductive invariant that identically matches the human-written one previously derived with significant manual effort using interactive theorem proving. While various attempts have been made to verify different versions of Paxos, to the best of our knowledge, this is the first demonstration of an automatically-inferred inductive invariant for Lamport's original Paxos specification. We note that these structural features are not specific to Paxos and that IC3PO can serve as an automatic general-purpose protocol verification tool., Comment: "to be published in Formal Methods in Computer-Aided Design (FMCAD) 2021, see https://fmcad.org/FMCAD21"

The COVID-19 pandemic has by-and-large prevented in-person meetings since March 2020. While the increasing deployment of effective vaccines around the world is a very positive development, the timeline and pathway to "normality" is uncertain and the "new normal" we will settle into is anyone's guess. Particle physics, like many other scientific fields, has more than a year of experience in holding virtual meetings, workshops, and conferences. A great deal of experimentation and innovation to explore how to execute these meetings effectively has occurred. Therefore, it is an appropriate time to take stock of what we as a community learned from running virtual meetings and discuss possible strategies for the future. Continuing to develop effective strategies for meetings with a virtual component is likely to be important for reducing the carbon footprint of our research activities, while also enabling greater diversity and inclusion for participation. This report summarizes a virtual two-day workshop on Virtual Meetings held May 5-6, 2021 which brought together experts from both inside and outside of high-energy physics to share their experiences and practices with organizing and executing virtual workshops, and to develop possible strategies for future meetings as we begin to emerge from the COVID-19 pandemic. This report outlines some of the practices and tools that have worked well which we hope will serve as a valuable resource for future virtual meeting organizers in all scientific fields., Comment: A report from the "Virtual Meetings" IRIS-HEP Blueprint Workshop: https://indico.cern.ch/event/1026363/

Proving that an unbounded distributed protocol satisfies a given safety property amounts to finding a quantified inductive invariant that implies the property for all possible instance sizes of the protocol. Existing methods for solving this problem can be described as search procedures for an invariant whose quantification prefix fits a particular template. We propose an alternative constructive approach that does not prescribe, a priori, a specific quantifier prefix. Instead, the required prefix is automatically inferred without any search by carefully analyzing the structural symmetries of the protocol. The key insight underlying this approach is that symmetry and quantification are closely related concepts that express protocol invariance under different re-arrangements of its components. We propose symmetric incremental induction, an extension of the finite-domain IC3/PDR algorithm, that automatically derives the required quantified inductive invariant by exploiting the connection between symmetry and quantification. While various attempts have been made to exploit symmetry in verification applications, to our knowledge, this is the first demonstration of a direct link between symmetry and quantification in the context of clause learning during incremental induction. We also describe a procedure to automatically find a minimal finite size, the cutoff, that yields a quantified invariant proving safety for any size. Our approach is implemented in IC3PO, a new verifier for distributed protocols that significantly outperforms the state-of-the-art, scales orders of magnitude faster, and robustly derives compact inductive invariants fully automatically.

Synthesis of long-term human motion skeleton sequences is essential to aid human-centric video generation with potential applications in Augmented Reality, 3D character animations, pedestrian trajectory prediction, etc. Long-term human motion synthesis is a challenging task due to multiple factors like, long-term temporal dependencies among poses, cyclic repetition across poses, bi-directional and multi-scale dependencies among poses, variable speed of actions, and a large as well as partially overlapping space of temporal pose variations across multiple class/types of human activities. This paper aims to address these challenges to synthesize a long-term (> 6000 ms) human motion trajectory across a large variety of human activity classes (>50). We propose a two-stage activity generation method to achieve this goal, where the first stage deals with learning the long-term global pose dependencies in activity sequences by learning to synthesize a sparse motion trajectory while the second stage addresses the generation of dense motion trajectories taking the output of the first stage. We demonstrate the superiority of the proposed method over SOTA methods using various quantitative evaluation metrics on publicly available datasets., Comment: Appearing in 2021 IEEE Winter Conference on Applications of Computer Vision (WACV)