There is a fast-growing literature on estimating optimal treatment rules directly by maximizing the expected outcome. In biomedical studies and operations applications, censored survival outcome is frequently observed, in which case the restricted mean survival time and survival probability are of great interest. In this paper, we propose two robust criteria for learning optimal treatment rules with censored survival outcomes; the former one targets at an optimal treatment rule maximizing the restricted mean survival time, where the restriction is specified by a given quantile such as median; the latter one targets at an optimal treatment rule maximizing buffered survival probabilities, where the predetermined threshold is adjusted to account the restricted mean survival time. We provide theoretical justifications for the proposed optimal treatment rules and develop a sampling-based difference-of-convex algorithm for learning them. In simulation studies, our estimators show improved performance compared to existing methods. We also demonstrate the proposed method using AIDS clinical trial data.

This paper introduces a new hyper-parameter for Retrieval-Augmented Generation (RAG) systems called Context Window Utilization. RAG systems enhance generative models by incorporating relevant information retrieved from external knowledge bases, improving the factual accuracy and contextual relevance of generated responses. The size of the text chunks retrieved and processed is a critical factor influencing RAG performance. This study aims to identify the optimal chunk size that maximizes answer generation quality. Through systematic experimentation, we analyze the effects of varying chunk sizes on the efficiency and effectiveness of RAG frameworks. Our findings reveal that an optimal chunk size balances the trade-off between providing sufficient context and minimizing irrelevant information. These insights are crucial for enhancing the design and implementation of RAG systems, underscoring the importance of selecting an appropriate chunk size to achieve superior performance.

A controversial bill that seeks to protect Californians from artificial intelligence-driven catastrophes has caused uproar in the tech industry. This week, the legislation passed a key committee but with amendments to make it more palatable to Silicon Valley. SB 1047, from state Sen. Scott Wiener (D-San Francisco), is set to go to the state Assembly floor later this month. If it passes the Legislature, Gov. Gavin Newsom will have to decide whether to sign or veto the groundbreaking legislation. The bill's backers say it will create guardrails to prevent rapidly advancing AI models from causing disastrous incidents, such as shutting down the power grid without warning.

San Francisco City Attorney David Chiu announced he intended to shut down 16 of the most popular AI "undressing" sites at a press conference on Thursday. The Verge reported that the City Attorney is accusing these sites of violating federal laws regarding revenge pornography, deepfake pornography and child pornography. Chiu's office also accused the sites of violating the state of California's unfair competition law because "the harm they cause to consumers greatly outweighs any benefits associated with those practices," according to the complaint for injunctive relief filed in a California superior court. The complaint focuses on a total of 50 defendants Chiu intends to prosecute for operating undressing websites. Some of the defendants' and websites' names were redacted but it also publicly identifies a few companies that operate "some of the world's most popular websites that offer to nudify images of women and girls" such as Sol Ecom located in Florida, Briver in New Mexico and the UK-based Itai Tech Ltd.

Reinforcement Learning from Human Feedback (RLHF) aims to align language models (LMs) with human values by training reward models (RMs) on binary preferences and using these RMs to fine-tune the base LMs. Despite its importance, the internal mechanisms of RLHF remain poorly understood. This paper introduces new metrics to evaluate the effectiveness of modeling and aligning human values, namely feature imprint, alignment resistance and alignment robustness. We categorize alignment datasets into target features (desired values) and spoiler features (undesired concepts). By regressing RM scores against these features, we quantify the extent to which RMs reward them - a metric we term feature imprint. We define alignment resistance as the proportion of the preference dataset where RMs fail to match human preferences, and we assess alignment robustness by analyzing RM responses to perturbed inputs. Our experiments, utilizing open-source components like the Anthropic/hh-rlhf preference dataset and OpenAssistant RMs, reveal significant imprints of target features and a notable sensitivity to spoiler features. We observed a 26% incidence of alignment resistance in portions of the dataset where LM-labelers disagreed with human preferences. Furthermore, we find that misalignment often arises from ambiguous entries within the alignment dataset. These findings underscore the importance of scrutinizing both RMs and alignment datasets for a deeper understanding of value alignment.

Sociotechnical requirements shape the governance of artificially intelligent (AI) systems. In an era where embodied AI technologies are rapidly reshaping various facets of contemporary society, their inherent dynamic adaptability presents a unique blend of opportunities and challenges. Traditional regulatory mechanisms, often designed for static -- or slower-paced -- technologies, find themselves at a crossroads when faced with the fluid and evolving nature of AI systems. Moreover, typical problems in AI, for example, the frequent opacity and unpredictability of the behaviour of the systems, add additional sociotechnical challenges. To address these interconnected issues, we introduce the concept of dynamic certification, an adaptive regulatory framework specifically crafted to keep pace with the continuous evolution of AI systems. The complexity of these challenges requires common progress in multiple domains: technical, socio-governmental, and regulatory. Our proposed transdisciplinary approach is designed to ensure the safe, ethical, and practical deployment of AI systems, aligning them bidirectionally with the real-world contexts in which they operate. By doing so, we aim to bridge the gap between rapid technological advancement and effective regulatory oversight, ensuring that AI systems not only achieve their intended goals but also adhere to ethical standards and societal values.

The impressive performance of Large Language Models (LLMs) has consistently surpassed numerous human-designed benchmarks, presenting new challenges in assessing the shortcomings of LLMs. Designing tasks and finding LLMs' limitations are becoming increasingly important. In this paper, we investigate the question of whether an LLM can discover its own limitations from the errors it makes. To this end, we propose a Self-Challenge evaluation framework with human-in-the-loop. Starting from seed instances that GPT-4 fails to answer, we prompt GPT-4 to summarize error patterns that can be used to generate new instances and incorporate human feedback on them to refine these patterns for generating more challenging data, iteratively. We end up with 8 diverse patterns, such as text manipulation and questions with assumptions. We then build a benchmark, SC-G4, consisting of 1,835 instances generated by GPT-4 using these patterns, with human-annotated gold responses. The SC-G4 serves as a challenging benchmark that allows for a detailed assessment of LLMs' abilities. Our results show that only 44.96\% of instances in SC-G4 can be answered correctly by GPT-4. Interestingly, our pilot study indicates that these error patterns also challenge other LLMs, such as Claude-3 and Llama-3, and cannot be fully resolved through fine-tuning. Our work takes the first step to demonstrate that LLMs can autonomously identify their inherent flaws and provide insights for future dynamic and automatic evaluation.

Federated Learning presents a way to revolutionize AI applications by eliminating the necessity for data sharing. Yet, research has shown that information can still be extracted during training, making additional privacy-preserving measures such as differential privacy imperative. To implement real-world federated learning applications, fairness, ranging from a fair distribution of performance to non-discriminative behaviour, must be considered. Particularly in high-risk applications (e.g. healthcare), avoiding the repetition of past discriminatory errors is paramount. As recent research has demonstrated an inherent tension between privacy and fairness, we conduct a multivocal literature review to examine the current methods to integrate privacy and fairness in federated learning. Our analyses illustrate that the relationship between privacy and fairness has been neglected, posing a critical risk for real-world applications. We highlight the need to explore the relationship between privacy, fairness, and performance, advocating for the creation of integrated federated learning frameworks.

The Mixture of Experts (MoE) framework has become a popular architecture for large language models due to its superior performance over dense models. However, training MoEs from scratch in a large-scale regime is prohibitively expensive. Existing methods mitigate this by pre-training multiple dense expert models independently and using them to initialize an MoE. This is done by using experts' feed-forward network (FFN) to initialize the MoE's experts while merging other parameters. However, this method limits the reuse of dense model parameters to only the FFN layers, thereby constraining the advantages when "upcycling" these models into MoEs. We propose BAM (Branch-Attend-Mix), a simple yet effective method that addresses this shortcoming. BAM makes full use of specialized dense models by not only using their FFN to initialize the MoE layers but also leveraging experts' attention parameters fully by initializing them into a soft-variant of Mixture of Attention (MoA) layers. We explore two methods for upcycling attention parameters: 1) initializing separate attention experts from dense models including all attention parameters for the best model performance; and 2) sharing key and value parameters across all experts to facilitate for better inference efficiency. To further improve efficiency, we adopt a parallel attention transformer architecture to MoEs, which allows the attention experts and FFN experts to be computed concurrently. Our experiments on seed models ranging from 590 million to 2 billion parameters demonstrate that BAM surpasses baselines in both perplexity and downstream task performance, within the same computational and data constraints.

As the US government weighs its options following a landmark "monopolist" ruling against Google last week, online publications increasingly face a bleak future. Bloomberg reports that their choice now boils down to allowing Google to use their published content to produce inline AI-generated search "answers" or losing visibility in the company's search engine. The crux of the problem lies in the Googlebot, the crawler that scours and indexes the live web to produce the results you see when you enter search terms. If publishers block Google from using their content for the AI-produced answers you now see littered at the top of many search results, they also lose the privilege of appearing in other Google search programs like snippets and Discover. Google uses a separate crawler for its Gemini (formerly Bard) chatbot, but its AI Overviews are generated using data from its main crawler.