PulseAugur / Brief
EN
LIVE 09:40:26

Brief

last 24h
[3/3] 221 sources

Multi-source AI news clustered, deduplicated, and scored 0–100 across authority, cluster strength, headline signal, and time decay.

  1. TOOL · arXiv cs.AI English(EN) ·

    Reasoning as an Attack Surface: Adaptive Evolutionary CoT Jailbreaks for LLMs

    Researchers have developed an adaptive evolutionary framework called AE-CoT to jailbreak large reasoning models (LRMs). This method rewrites harmful goals into mild prompts and decomposes them into reasoning fragments to create jailbreak candidates. The framework then uses evolutionary search with crossover and mutation strategies to expand candidate diversity, and an independent scoring model evaluates harmfulness to enhance destructive generations. Experiments show AE-CoT outperforms existing jailbreak methods across multiple models and datasets. AI

    IMPACT This research highlights new vulnerabilities in LLMs, potentially impacting their safe deployment and prompting further research into robust defense mechanisms.

  2. RESEARCH · arXiv cs.AI English(EN) · · [2 sources]

    Strategy-Induct: Task-Level Strategy Induction for Instruction Generation

    Researchers have developed Strategy-Induct, a novel framework for generating effective task-level instructions for Large Language Models (LLMs). This method derives instructions solely from example questions, bypassing the need for labeled answers, which can be costly to obtain. Strategy-Induct first prompts LLMs to generate reasoning strategies for each question, then uses these strategy-question pairs to induce a guiding task instruction. Experiments show this approach surpasses existing methods in question-only settings and suggests potential further improvements by combining LLMs with Large Reasoning Models. AI

    Strategy-Induct: Task-Level Strategy Induction for Instruction Generation

    IMPACT This new method for instruction generation could reduce the cost and effort required to fine-tune LLMs, potentially accelerating their adoption in new tasks.

  3. TOOL · Hugging Face Daily Papers English(EN) ·

    Attention-Guided Reward for Reinforcement Learning-based Jailbreak against Large Reasoning Models

    Researchers have developed a new jailbreak method specifically targeting Large Reasoning Models (LRMs), which are known for their step-by-step problem-solving abilities. The method leverages reinforcement learning and incorporates the models' attention patterns into the reward function, as studies show jailbreaks are more successful when attention is misdirected. This approach, enhanced with diverse persuasion strategies, significantly increases the attack success rate across various benchmarks and models. AI

    Attention-Guided Reward for Reinforcement Learning-based Jailbreak against Large Reasoning Models

    IMPACT This research highlights a new vulnerability in advanced reasoning models, potentially influencing future safety research and defense strategies.