In a realistic dialogue system, the input information from users is often subject to various types of input perturbations, which affects the slot-filling task. Although rule-based data augmentation methods have achieved satisfactory results, they fail to exhibit the desired generalization when faced with unknown noise disturbances. In this study, we address the challenges posed by input perturbations in slot filling by proposing Noise-BERT, a unified Perturbation-Robust Framework with Noise Alignment Pre-training. Our framework incorporates two Noise Alignment Pre-training tasks: Slot Masked Prediction and Sentence Noisiness Discrimination, aiming to guide the pre-trained language model in capturing accurate slot information and noise distribution. During fine-tuning, we employ a contrastive learning loss to enhance the semantic representation of entities and labels. Additionally, we introduce an adversarial attack training strategy to improve the model's robustness. Experimental results demonstrate the superiority of our proposed approach over state-of-the-art models, and further analysis confirms its effectiveness and generalization ability.

Recent works have argued that high-level semantic concepts are encoded "linearly" in the representation space of large language models. In this work, we study the origins of such linear representations. To that end, we introduce a simple latent variable model to abstract and formalize the concept dynamics of the next token prediction. We use this formalism to show that the next token prediction objective (softmax with cross-entropy) and the implicit bias of gradient descent together promote the linear representation of concepts. Experiments show that linear representations emerge when learning from data matching the latent variable model, confirming that this simple structure already suffices to yield linear representations. We additionally confirm some predictions of the theory using the LLaMA-2 large language model, giving evidence that the simplified model yields generalizable insights.

Fine-grained control over large language models (LLMs) remains a significant challenge, hindering their adaptability to diverse user needs. While Reinforcement Learning from Human Feedback (RLHF) shows promise in aligning LLMs, its reliance on scalar rewards often limits its ability to capture diverse user preferences in real-world applications. To address this limitation, we introduce the Directional Preference Alignment (DPA) framework. Unlike the scalar-reward RLHF, DPA incorporates multi-objective reward modeling to represent diverse preference profiles. Additionally, DPA models user preferences as directions (i.e., unit vectors) in the reward space to achieve user-dependent preference control. Our method involves training a multi-objective reward model and then fine-tuning the LLM with a preference-conditioned variant of Rejection Sampling Finetuning (RSF), an RLHF method adopted by Llama 2. This method enjoys a better performance trade-off across various reward objectives. In comparison with the scalar-reward RLHF, DPA offers users intuitive control over LLM generation: they can arithmetically specify their desired trade-offs (e.g., more helpfulness with less verbosity). We also validate the effectiveness of DPA with real-world alignment experiments on Mistral-7B. Our method provides straightforward arithmetic control over the trade-off between helpfulness and verbosity while maintaining competitive performance with strong baselines such as Direct Preference Optimization (DPO).

AI expert Marva Bailer explains how Will.i.am's app, FYI, powers his AI co-host for his radio show and why the platform has different capabilities than ChatGPT While artificial intelligence made headlines with ChatGPT, behind the scenes, the technology has quietly pervaded everyday life -- screening job resumes, rental apartment applications, and even determining medical care in some cases. While a number of AI systems have been found to discriminate, tipping the scales in favor of certain races, genders or incomes, there's scant government oversight. Lawmakers in at least seven states are taking big legislative swings to regulate bias in artificial intelligence, filling a void left by Congress' inaction. These proposals are some of the first steps in a decades-long discussion over balancing the benefits of this nebulous new technology with the widely documented risks. "AI does in fact affect every part of your life whether you know it or not," said Suresh Venkatasubramanian, a Brown University professor who co-authored the White House's Blueprint for an AI Bill of Rights.

The letter was signed by experts in AI research, policy, and law, including Stanford University's Percy Liang; Pulitzer Prize-winning journalist Julia Angwin; Renée DiResta from the Stanford Internet Observatory; Mozilla fellow Deb Raji, who has pioneered research into auditing AI models; ex-government official Marietje Schaake, a former member of European Parliament; and Brown University professor Suresh Venkatasubramanian, a former adviser to the White House Office of Science and Technology Policy.

If you'll recall, The Times sued both companies for using its published articles to train their GPT large language models (LLMs) without permission and compensation. In its filing, the company has accused The Times of pushing "doomsday futurology" by claiming that AI technologies pose a threat to independent journalism. It follows OpenAI's court filing from late February that's also seeking to dismiss some important elements on the case. Like OpenAI before it, Microsoft accused The Times of crafting "unrealistic prompts" in an effort to "coax the GPT-based tools" to spit out responses matching its content. It also compared the media organization's lawsuit to Hollywood studios' efforts to " stop a groundbreaking new technology:" The VCR. Instead of destroying Hollywood, Microsoft explained, the VCR helped the entertainment industry flourish by opening up revenue streams.

Recently we've seen some AI failures on a far bigger scale. In the latest (hilarious) gaffe, Google's Gemini refused to generate images of white people, especially white men. Instead, users were able to generate images of Black popes and female Nazi soldiers. Google had been trying to get the outputs of its model to be less biased, but this backfired, and the tech company soon found itself in the middle of the US culture wars, with conservative critics and Elon Musk accusing it of having a "woke" bias and not representing history accurately. Google apologized and paused the feature. In another now-famous incident, Microsoft's Bing chat told a New York Times reporter to leave his wife.

The battle between AI chatbots is more than a two-horse race. Anthropic, the company formed by several ex-OpenAI employees, claims its new Claude 3 language model outperforms ChatGPT and Google's Gemini in several key industry benchmarks. It even hit "near-human" levels on some tasks, the company wrote in a blog. There are three new chatbots under the Claude 3 umbrella, including Haiku, Sonnet, and Opus. Meanwhile, Opus is the largest and most powerful LLM and will be available with a 20 per month subscription via the "Claude Pro" service. It's also multi-modal, so it can work with both text and image inputs, unlike past versions.

Large Language Models (LLMs) are becoming a prominent generative AI tool, where the user enters a query and the LLM generates an answer. To reduce harm and misuse, efforts have been made to align these LLMs to human values using advanced training techniques such as Reinforcement Learning from Human Feedback (RLHF). However, recent studies have highlighted the vulnerability of LLMs to adversarial jailbreak attempts aiming at subverting the embedded safety guardrails. To address this challenge, this paper defines and investigates the Refusal Loss of LLMs and then proposes a method called Gradient Cuff to detect jailbreak attempts. Gradient Cuff exploits the unique properties observed in the refusal loss landscape, including functional values and its smoothness, to design an effective two-step detection strategy. Experimental results on two aligned LLMs (LLaMA-2-7B-Chat and Vicuna-7B-V1.5) and six types of jailbreak attacks (GCG, AutoDAN, PAIR, TAP, Base64, and LRL) show that Gradient Cuff can significantly improve the LLM's rejection capability for malicious jailbreak queries, while maintaining the model's performance for benign user queries by adjusting the detection threshold.

Embodied Instruction Following (EIF) is a crucial task in embodied learning, requiring agents to interact with their environment through egocentric observations to fulfill natural language instructions. Recent advancements have seen a surge in employing large language models (LLMs) within a framework-centric approach to enhance performance in embodied learning tasks, including EIF. Despite these efforts, there exists a lack of a unified understanding regarding the impact of various components-ranging from visual perception to action execution-on task performance. To address this gap, we introduce OPEx, a comprehensive framework that delineates the core components essential for solving embodied learning tasks: Observer, Planner, and Executor. Through extensive evaluations, we provide a deep analysis of how each component influences EIF task performance. Furthermore, we innovate within this space by deploying a multi-agent dialogue strategy on a TextWorld counterpart, further enhancing task performance. Our findings reveal that LLM-centric design markedly improves EIF outcomes, identify visual perception and low-level action execution as critical bottlenecks, and demonstrate that augmenting LLMs with a multi-agent framework further elevates performance.