Replicating TEMPEST at Scale: Multi-Turn Adversarial Attacks Against Trillion-Parameter Frontier Models

QSTN has been introduced as an open-source Python framework designed to generate responses from questionnaire-style prompts, facilitating in-silico surveys and annotation tasks with large language models (LLMs). The framework allows for robust evaluation of questionnaire presentation and response generation methods, based on an extensive analysis of over 40 million survey responses.

A new study introduces RLHF-COV and DPO-COV algorithms designed to address critical issues in reinforcement learning from human feedback (RLHF) and direct preference optimization (DPO), specifically targeting corrupted preferences, reward overoptimization, and verbosity in large language models (LLMs). These algorithms promise to enhance the alignment of LLMs with human preferences in both offline and online settings.

A recent study published on arXiv explores the interpretability of machine translation models, particularly focusing on how gender bias manifests in translation choices. By utilizing contrastive explanations and saliency attribution, the research investigates the influence of context, specifically input tokens, on the gender inflection selected by translation models. This approach aims to uncover the origins of gender bias rather than merely measuring its presence.

A new study has introduced a soft inductive bias approach to enhance inappropriate utterance detection in conversational texts using large language models (LLMs), specifically focusing on Korean corpora. This method aims to define explicit reasoning perspectives to guide inference processes, thereby improving rational decision-making and reducing errors in detecting inappropriate remarks.

The evaluation of large language models (LLMs) is increasingly reliant on classifiers, either LLMs or human annotators, to assess desirable or undesirable behaviors. A recent study highlights that traditional metrics like Accuracy and F1 can be misleading due to class imbalances, advocating for the use of Youden's J statistic and Balanced Accuracy as more reliable alternatives for selecting evaluators.

The Biothreat Benchmark Generation Framework has introduced the Bacterial Biothreat Benchmark (B3) dataset, aimed at evaluating the biosecurity risks associated with frontier AI models, particularly large language models (LLMs). This framework employs web-based prompt generation, red teaming, and mining existing benchmark corpora to create over 7,000 potential benchmarks linked to the Task-Query Architecture.

The study investigates the short-context dominance hypothesis, suggesting that a small local prefix can often predict the next tokens in sequences. Using large language models, researchers found that 75-80% of sequences from long-context documents only require the last 96 tokens for accurate predictions, leading to the introduction of a new metric called Distributionally Aware MCL (DaMCL) to identify challenging long-context sequences.

The introduction of Omniguard presents a novel approach to AI safety moderation by enhancing the detection of harmful prompts across various languages and modalities, addressing the vulnerabilities of large language models (LLMs) to misuse. This method improves classification accuracy by 11.57% over existing baselines, marking a significant advancement in AI safety protocols.