Large Language Models (LLMs) have excelled in various tasks but are still vulnerable to jailbreaking attacks, where attackers create jailbreak prompts to mislead the model to produce harmful or offensive content. Current jailbreak methods either rely heavily on manually crafted templates, which pose challenges in scalability and adaptability, or struggle to generate semantically coherent prompts, making them easy to detect. Additionally, most existing approaches involve lengthy prompts, leading to higher query costs.In this paper, to remedy these challenges, we introduce a novel jailbreaking attack framework, which is an automated, black-box jailbreaking attack framework that adapts the black-box fuzz testing approach with a series of customized designs. Instead of relying on manually crafted templates, our method starts with an empty seed pool, removing the need to search for any related jailbreaking templates. We also develop three novel question-dependent mutation strategies using an LLM helper to generate prompts that maintain semantic coherence while significantly reducing their length. Additionally, we implement a two-level judge module to accurately detect genuine successful jailbreaks. We evaluated our method on 7 representative LLMs and compared it with 5 state-of-the-art jailbreaking attack strategies. For proprietary LLM APIs, such as GPT-3.5 turbo, GPT-4, and Gemini-Pro, our method achieves attack success rates of over 90%,80% and 74%, respectively, exceeding existing baselines by more than 60%. Additionally, our method can maintain high semantic coherence while significantly reducing the length of jailbreak prompts. When targeting GPT-4, our method can achieve over 78% attack success rate even with 100 tokens. Moreover, our method demonstrates transferability and is robust to state-of-the-art defenses. We will open-source our codes upon publication.

Robust therapeutic relationships between counselors and clients are fundamental to counseling effectiveness. The assessment of therapeutic alliance is well-established in traditional face-to-face therapy but may not directly translate to text-based settings. With millions of individuals seeking support through online text-based counseling, understanding the relationship in such contexts is crucial. In this paper, we present an automatic approach using large language models (LLMs) to understand the development of therapeutic alliance in text-based counseling. We adapt a theoretically grounded framework specifically to the context of online text-based counseling and develop comprehensive guidelines for characterizing the alliance. We collect a comprehensive counseling dataset and conduct multiple expert evaluations on a subset based on this framework. Our LLM-based approach, combined with guidelines and simultaneous extraction of supportive evidence underlying its predictions, demonstrates effectiveness in identifying the therapeutic alliance. Through further LLM-based evaluations on additional conversations, our findings underscore the challenges counselors face in cultivating strong online relationships with clients. Furthermore, we demonstrate the potential of LLM-based feedback mechanisms to enhance counselors' ability to build relationships, supported by a small-scale proof-of-concept.

With the increasing acquisition of datasets over time, we now have access to precise and varied descriptions of the world, capturing all sorts of phenomena. These datasets can be seen as empirical observations of unknown causal generative processes, which can commonly be described by Structural Causal Models (SCMs). Recovering these causal generative processes from observations poses formidable challenges, and often require to learn a specific generative model for each dataset. In this work, we propose to learn a \emph{single} model capable of inferring in a zero-shot manner the causal generative processes of datasets. Rather than learning a specific SCM for each dataset, we enable the Fixed-Point Approach (FiP) proposed in~\cite{scetbon2024fip}, to infer the generative SCMs conditionally on their empirical representations. More specifically, we propose to amortize the learning of a conditional version of FiP to infer generative SCMs from observations and causal structures on synthetically generated datasets. We show that our model is capable of predicting in zero-shot the true generative SCMs, and as a by-product, of (i) generating new dataset samples, and (ii) inferring intervened ones. Our experiments demonstrate that our amortized procedure achieves performances on par with SoTA methods trained specifically for each dataset on both in and out-of-distribution problems. To the best of our knowledge, this is the first time that SCMs are inferred in a zero-shot manner from observations, paving the way for a paradigmatic shift towards the assimilation of causal knowledge across datasets.

Disasters can result in the deaths of many, making quick response times vital. Large Language Models (LLMs) have emerged as valuable in the field. LLMs can be used to process vast amounts of textual information quickly providing situational context during a disaster. However, the question remains whether LLMs should be used for advice and decision making in a disaster. To evaluate the capabilities of LLMs in disaster response knowledge, we introduce a benchmark: DisasterQA created from six online sources. The benchmark covers a wide range of disaster response topics. We evaluated five LLMs each with four different prompting methods on our benchmark, measuring both accuracy and confidence levels through Logprobs. The results indicate that LLMs require improvement on disaster response knowledge. We hope that this benchmark pushes forth further development of LLMs in disaster response, ultimately enabling these models to work alongside. emergency managers in disasters.

Large language models (LLMs) hold significant promise in advancing network management and orchestration in 6G and beyond networks. However, existing LLMs are limited in domain-specific knowledge and their ability to handle multi-modal sensory data, which is critical for real-time situational awareness in dynamic wireless environments. This paper addresses this gap by introducing ENWAR, an ENvironment-aWARe retrieval augmented generation-empowered multi-modal LLM framework. ENWAR seamlessly integrates multi-modal sensory inputs to perceive, interpret, and cognitively process complex wireless environments to provide human-interpretable situational awareness. ENWAR is evaluated on the GPS, LiDAR, and camera modality combinations of DeepSense6G dataset with state-of-the-art LLMs such as Mistral-7b/8x7b and LLaMa3.1-8/70/405b. Compared to general and often superficial environmental descriptions of these vanilla LLMs, ENWAR delivers richer spatial analysis, accurately identifies positions, analyzes obstacles, and assesses line-of-sight between vehicles. Results show that ENWAR achieves key performance indicators of up to 70% relevancy, 55% context recall, 80% correctness, and 86% faithfulness, demonstrating its efficacy in multi-modal perception and interpretation.

Summary This paper presents a stacked semantics-guided attention (S2GA) model for improved zero-shot learning. The main idea of this paper is that important regions should contribute more to the prediction. To this end, the authors design an attention method to distribute different weights for different regions according to their relevance with class semantic features and integrate both the global visual features and the weighted region features into more semantics-relevant features to represent images. Strengths The method is well motivated. The presentation of the method is clear. Using stacked attention for zero-shot learning seems to be a new idea (I do not check exhaustively).

When investors poured 6.6 billion into OpenAI last week, they seemed largely unbothered by the latest drama, which recently saw the company's chief technology officer, Mira Murati, along with chief research officer Bob McCrew and Barret Zoph, a vice president of research, abruptly quit. And yet those three departures were just the latest in an ongoing exodus of key technical talent. Over the past few years, OpenAI has lost several researchers who played crucial roles in developing the algorithms, techniques, and infrastructure that helped make it the world leader in AI as well as a household name. Several other ex-OpenAI employees who spoke to WIRED said that an ongoing shift to a more commercial focus continues to be a source of friction. "People who like to do research are being forced to do product," says one former employee who works at a rival AI company but has friends at OpenAI. This person says some of their contacts at the firm have reached out in recent weeks to inquire about jobs.

It's too soon to say how the spate of deals between AI companies and publishers will shake out. OpenAI has already scored one clear win, though: Its web crawlers aren't getting blocked by top news outlets at the rate they once were. The generative AI boom sparked a gold rush for data--and a subsequent data-protection rush (for most news websites, anyway) in which publishers sought to block AI crawlers and prevent their work from becoming training data without consent. When Apple debuted a new AI agent this summer, for example, a slew of top news outlets swiftly opted out of Apple's web scraping using the Robots Exclusion Protocol, or robots.txt, the file that allows webmasters to control bots. There are so many new AI bots on the scene that it can feel like playing whack-a-mole to keep up.

Have you noticed certain words and phrases popping up everywhere lately? Phrases such as "delve into" and "navigate the landscape" seem to feature in everything from social media posts to news articles and academic publications. They may sound fancy, but their overuse can make a text feel monotonous and repetitive. This trend may be linked to the increasing use of generative artificial intelligence (AI) tools such as ChatGPT and other large language models (LLMs). These tools are designed to make writing easier by offering suggestions based on patterns in the text they were trained on.

In this scenario, we know what the test classes will be (although we do not have acess to data from them). At test time, however, the data may also come from training classes. This is, generally speaking, harder than classical zero-shot learning.