Clair Obscur: Expedition 33 has been named game of the year in a record-breaking haul at this year's Game Awards. The French-developed role-playing game (RPG) cleaned up in nine of the 10 categories it was up for, with further wins in best narrative, best music and best performance. I think that's what I find really beautiful is that people really reacted strongly to all the passion we put into this game. Influencer Tallulah Metcalfe took part in the BBC Teen Summit in Bradford, and gave us her best tips. Dr Kaitlyn Regehr has this advice to help you stop looking at your phone and go on a digital detox.

But new biases are introduced due to the same LLM acting as both the evaluator and the evaluatee.

But new biases are introduced due to the same LLM acting as both the evaluator and the evaluatee.

Faithful generation in large language models (LLMs) is challenged by knowledge conflicts between parametric memory and external context. Existing contrastive decoding methods tuned specifically to handle conflict often lack adaptability and can degrade performance in low conflict settings. We introduce CoCoA (Confidence- and Context-Aware Adaptive Decoding), a novel token-level algorithm for principled conflict resolution and enhanced faithfulness. CoCoA resolves conflict by utilizing confidence-aware measures (entropy gap and contextual peakedness) and the generalized divergence between the parametric and contextual distributions. Crucially, CoCoA maintains strong performance even in low conflict settings. Extensive experiments across multiple LLMs on diverse Question Answering (QA), Summarization, and Long-Form Question Answering (LFQA) benchmarks demonstrate CoCoA's state-of-the-art performance over strong baselines like AdaCAD. It yields significant gains in QA accuracy, up to 9.2 points on average compared to the strong baseline AdaCAD, and improves factuality in summarization and LFQA by up to 2.5 points on average across key benchmarks. Additionally, it demonstrates superior sensitivity to conflict variations. CoCoA enables more informed, context-aware, and ultimately more faithful token generation.

Retrieval-augmented generation (RAG) enhances Large Language Models (LLMs) with relevant and up-to-date knowledge, improving their ability to answer knowledge-intensive questions. It has been shown to enhance both generation quality and trustworthiness. While numerous works have focused on improving retrieval, generation, and evaluation, the role of reward models in reinforcement learning for optimizing RAG remains underexplored. In this paper, we introduce \textbf{RAG-Reward}, a framework designed to develop reward models to enable \textit{hallucination-free, comprehensive, reliable, and efficient RAG}. We define four key metrics to assess generation quality and develop an automated benchmarking pipeline to evaluate the outputs of multiple LLMs across a variety of RAG scenarios. Using \textbf{RAG-Reward}, we train reward models and apply {reinforcement learning with human feedback (RLHF)} to improve LLMs' effectiveness in RAG. Experimental results demonstrate that our reward model achieves state-of-the-art performance in automatic benchmarking and aligns closely with human evaluations. Furthermore, the improved generation quality of the trained policy model highlights the feasibility and efficiency of using RLHF to enhance RAG outputs.

Self-evaluation using large language models (LLMs) has proven valuable not only in benchmarking but also methods like reward modeling, constitutional AI, and self-refinement. But new biases are introduced due to the same LLM acting as both the evaluator and the evaluatee. One such bias is self-preference, where an LLM evaluator scores its own outputs higher than others' while human annotators consider them of equal quality. But do LLMs actually recognize their own outputs when they give those texts higher scores, or is it just a coincidence? In this paper, we investigate if self-recognition capability contributes to self-preference. We discover that, out of the box, LLMs such as GPT-4 and Llama 2 have non-trivial accuracy at distinguishing themselves from other LLMs and humans. By fine-tuning LLMs, we discover a linear correlation between self-recognition capability and the strength of self-preference bias; using controlled experiments, we show that the causal explanation resists straightforward confounders. We discuss how self-recognition can interfere with unbiased evaluations and AI safety more generally.

Text summarization and simplification are among the most widely used applications of AI. However, models developed for such tasks are often prone to hallucination, which can result from training on unaligned data. One efficient approach to address this issue is Loss Truncation (LT) (Kang and Hashimoto, 2020), an approach to modify the standard log loss to adaptively remove noisy examples during training. However, we find that LT alone yields a considerable number of hallucinated entities on various datasets. We study the behavior of the underlying losses between factual and non-factual examples, to understand and refine the performance of LT. We demonstrate that LT's performance is limited when the underlying assumption that noisy targets have higher NLL loss is not satisfied, and find that word-level NLL among entities provides better signal for distinguishing factuality. We then leverage this to propose a fine-grained NLL loss and fine-grained data cleaning strategies, and observe improvements in hallucination reduction across some datasets. Our work is available at https://https://github.com/yale-nlp/fine-grained-lt.

Growing applications of large language models (LLMs) trained by a third party raise serious concerns on the security vulnerability of LLMs. It has been demonstrated that malicious actors can covertly exploit these vulnerabilities in LLMs through poisoning attacks aimed at generating undesirable outputs. While poisoning attacks have received significant attention in the image domain (e.g., object detection), and classification tasks, their implications for generative models, particularly in the realm of natural language generation (NLG) tasks, remain poorly understood. To bridge this gap, we perform a comprehensive exploration of various poisoning techniques to assess their effectiveness across a range of generative tasks. Furthermore, we introduce a range of metrics designed to quantify the success and stealthiness of poisoning attacks specifically tailored to NLG tasks. Through extensive experiments on multiple NLG tasks, LLMs and datasets, we show that it is possible to successfully poison an LLM during the fine-tuning stage using as little as 1% of the total tuning data samples. Our paper presents the first systematic approach to comprehend poisoning attacks targeting NLG tasks considering a wide range of triggers and attack settings. We hope our findings will assist the AI security community in devising appropriate defenses against such threats.

In-context learning, which offers substantial advantages over fine-tuning, is predominantly observed in decoder-only models, while encoder-decoder (i.e., seq2seq) models excel in methods that rely on weight updates. Recently, a few studies have demonstrated the feasibility of few-shot learning with seq2seq models; however, this has been limited to tasks that align well with the seq2seq architecture, such as summarization and translation. Inspired by these initial studies, we provide a first-ever extensive experiment comparing the in-context few-shot learning capabilities of decoder-only and encoder-decoder models on a broad range of tasks. Furthermore, we propose two methods to more effectively elicit in-context learning ability in seq2seq models: objective-aligned prompting and a fusion-based approach. Remarkably, our approach outperforms a decoder-only model that is six times larger and exhibits significant performance improvements compared to conventional seq2seq models across a variety of settings. We posit that, with the right configuration and prompt design, seq2seq models can be highly effective few-shot learners for a wide spectrum of applications.

Question-answering (QA) that comes naturally to humans is a critical component in seamless human-computer interaction. It has emerged as one of the most convenient and natural methods to interact with the web and is especially desirable in voice-controlled environments. Despite being one of the oldest research areas, the current QA system faces the critical challenge of handling multilingual queries. To build an Artificial Intelligent (AI) agent that can serve multilingual end users, a QA system is required to be language versatile and tailored to suit the multilingual environment. Recent advances in QA models have enabled surpassing human performance primarily due to the availability of a sizable amount of high-quality datasets. However, the majority of such annotated datasets are expensive to create and are only confined to the English language, making it challenging to acknowledge progress in foreign languages. Therefore, to measure a similar improvement in the multilingual QA system, it is necessary to invest in high-quality multilingual evaluation benchmarks. In this dissertation, we focus on advancing QA techniques for handling end-user queries in multilingual environments. This dissertation consists of two parts. In the first part, we explore multilingualism and a new dimension of multilingualism referred to as code-mixing. Second, we propose a technique to solve the task of multi-hop question generation by exploiting multiple documents. Experiments show our models achieve state-of-the-art performance on answer extraction, ranking, and generation tasks on multiple domains of MQA, VQA, and language generation. The proposed techniques are generic and can be widely used in various domains and languages to advance QA systems.