Recent works have shown that large language models (LLMs) work remarkably well on a wide range of tasks through in-context learning and optimization of in-context examples (ICE). However, most of these studies assume either a fixed or no instruction provided in the prompt, leading to the apparent consensus that the optimization of in-context examples is critical for better performance. We challenge this consensus for instruction-tuned LLMs by investigating the necessity of optimizing in-context examples when task-specific instructions are provided, and find that there are tasks for which various ways of optimizing in-context examples yield diminishing returns. We introduce a task-specific metric called \metriclong{} (\metric) that quantifies the learnability of tasks from a given instruction, and provides a heuristic that helps decide whether to optimize for instructions or ICE for any new task. On a wide range of tasks and a systematically created instruction set with gradually added details, we validate our hypothesis empirically by computing \metric with query-dependent bins of examples, comparing different instructions with ICE selection methods, and performing label perturbation experiments. We conclude that tasks can be divided into two broad classes based on the \metric metric, where the returns on ICE optimization follow predictable trends when instructions are provided in the prompt.

Large Language Models (LLMs) have shown promise as intelligent agents in interactive decision-making tasks. Traditional approaches often depend on meticulously designed prompts, high-quality examples, or additional reward models for in-context learning, supervised fine-tuning, or RLHF. Reinforcement learning (RL) presents a dynamic alternative for LLMs to overcome these dependencies by engaging directly with task-specific environments. Nonetheless, it faces significant hurdles: 1) instability stemming from the exponentially vast action space requiring exploration; 2) challenges in assigning token-level credit based on action-level reward signals, resulting in discord between maximizing rewards and accurately modeling corpus data. In response to these challenges, we introduce Entropy-Regularized Token-level Policy Optimization (ETPO), an entropy-augmented RL method tailored for optimizing LLMs at the token level. At the heart of ETPO is our novel per-token soft Bellman update, designed to harmonize the RL process with the principles of language modeling. This methodology decomposes the Q-function update from a coarse action-level view to a more granular token-level perspective, backed by theoretical proof of optimization consistency. Crucially, this decomposition renders linear time complexity in action exploration. We assess the effectiveness of ETPO within a simulated environment that models data science code generation as a series of multi-step interactive tasks; results show that ETPO achieves effective performance improvement on the CodeLlama-7B model and surpasses a variant PPO baseline inherited from RLHF. This underlines ETPO's potential as a robust method for refining the interactive decision-making capabilities of LLMs.

Language models influence the external world: they query APIs that read and write to web pages, generate content that shapes human behavior, and run system commands as autonomous agents. These interactions form feedback loops: LLM outputs affect the world, which in turn affect subsequent LLM outputs. In this work, we show that feedback loops can cause in-context reward hacking (ICRH), where the LLM at test-time optimizes a (potentially implicit) objective but creates negative side effects in the process. For example, consider an LLM agent posting tweets with the objective of maximizing Twitter engagement; the LLM may retrieve its previous tweets into the context window and make its subsequent tweets more controversial, increasing engagement but also toxicity. We identify and study two processes that lead to ICRH: output-refinement and policy-refinement. For these processes, evaluations on static datasets are insufficient--they miss the feedback effects and thus cannot capture the most harmful behavior. In response, we provide three recommendations for evaluation to capture more instances of ICRH. As AI development accelerates, the effects of feedback loops will proliferate, increasing the need to understand their role in shaping LLM behavior.

The advent and rapid evolution of Large Language Models (LLMs) represent a profound shift in the artificial intelligence landscape [1, 2]. These models, distinguished by their significant learning capabilities, have demonstrated exceptional aptitude in generating coherent and contextually relevant text[3]. This prowess has rendered them invaluable across diverse sectors, including finance, healthcare, and autonomous systems, revolutionizing conventional approaches to tasks in these domains [4-7]. Nevertheless, the advent of LLMs into various societal aspects has also heightened the scrutiny of their reliability, especially the integrity of their generated content [8]. Amidst their impressive feats, LLMs' consistency in delivering accurate and verifiable information remains a pertinent concern [9, 10]. Instances of models producing misleading information or showcasing unwarranted confidence in incorrect outputs have underscored the imperative for ensuring the veracity of LLM outputs, notably in critical sectors where precision and factual accuracy are non-negotiable [11]. The phenomenon of "hallucination," wherein models fabricate information, has catalyzed the urgency to amplify the truthfulness of LLMs, positioning it as a pivotal research focus with substantial implications on future model refinement and application [12].

Datasets are foundational to many breakthroughs in modern artificial intelligence. Many recent achievements in the space of natural language processing (NLP) can be attributed to the finetuning of pre-trained models on a diverse set of tasks that enables a large language model (LLM) to respond to instructions. Instruction fine-tuning (IFT) requires specifically constructed and annotated datasets. However, existing datasets are almost all in the English language. In this work, our primary goal is to bridge the language gap by building a human-curated instruction-following dataset spanning 65 languages. We worked with fluent speakers of languages from around the world to collect natural instances of instructions and completions. Furthermore, we create the most extensive multilingual collection to date, comprising 513 million instances through templating and translating existing datasets across 114 languages. In total, we contribute four key resources: we develop and open-source the Aya Annotation Platform, the Aya Dataset, the Aya Collection, and the Aya Evaluation Suite. The Aya initiative also serves as a valuable case study in participatory research, involving collaborators from 119 countries. We see this as a valuable framework for future research collaborations that aim to bridge gaps in resources.

We introduce FaBERT, a Persian BERT-base model pre-trained on the HmBlogs corpus, encompassing both informal and formal Persian texts. FaBERT is designed to excel in traditional Natural Language Understanding (NLU) tasks, addressing the intricacies of diverse sentence structures and linguistic styles prevalent in the Persian language. In our comprehensive evaluation of FaBERT on 12 datasets in various downstream tasks, encompassing Sentiment Analysis (SA), Named Entity Recognition (NER), Natural Language Inference (NLI), Question Answering (QA), and Question Paraphrasing (QP), it consistently demonstrated improved performance, all achieved within a compact model size. The findings highlight the importance of utilizing diverse and cleaned corpora, such as HmBlogs, to enhance the performance of language models like BERT in Persian Natural Language Processing (NLP) applications. FaBERT is openly accessible at https://huggingface.co/sbunlp/fabert

We study the problem of generating plans for given natural language planning task requests. On one hand, LLMs excel at natural language processing but do not perform well on planning. On the other hand, classical planning tools excel at planning tasks but require input in a structured language such as the Planning Domain Definition Language (PDDL). We leverage the strengths of both the techniques by using an LLM for generating the PDDL representation (task PDDL) of planning task requests followed by using a classical planner for computing a plan. Unlike previous approaches that use LLMs for generating task PDDLs directly, our approach comprises of (a) translate: using an LLM only for generating a logically interpretable intermediate representation of natural language task descriptions, (b) infer: deriving additional logically dependent information from the intermediate representation using a logic reasoner (currently, Answer Set Programming solver), and (c) compile: generating the target task PDDL from the base and inferred information. We observe that using an LLM to only output the intermediate representation significantly reduces LLM errors. Consequently, TIC approach achieves, for at least one LLM, high accuracy on task PDDL generation for all seven domains of our evaluation dataset.

We investigate the generalization boundaries of current Multimodal Large Language Models (MLLMs) via comprehensive evaluation under out-of-distribution scenarios and domain-specific tasks. We evaluate their zero-shot generalization across synthetic images, real-world distributional shifts, and specialized datasets like medical and molecular imagery. Empirical results indicate that MLLMs struggle with generalization beyond common training domains, limiting their direct application without adaptation. To understand the cause of unreliable performance, we analyze three hypotheses: semantic misinterpretation, visual feature extraction insufficiency, and mapping deficiency. Results identify mapping deficiency as the primary hurdle. To address this problem, we show that in-context learning (ICL) can significantly enhance MLLMs' generalization, opening new avenues for overcoming generalization barriers. We further explore the robustness of ICL under distribution shifts and show its vulnerability to domain shifts, label shifts, and spurious correlation shifts between in-context examples and test data.

The advancement of natural language processing has paved the way for automated scoring systems in various languages, such as German (e.g., German BERT [G-BERT]). Automatically scoring written responses to science questions in German is a complex task and challenging for standard G-BERT as they lack contextual knowledge in the science domain and may be unaligned with student writing styles. This paper developed a contextualized German Science Education BERT (G-SciEdBERT), an innovative large language model tailored for scoring German-written responses to science tasks. Using G-BERT, we pre-trained G-SciEdBERT on a corpus of 50K German written science responses with 5M tokens to the Programme for International Student Assessment (PISA) 2015. We fine-tuned G-SciEdBERT on 59 assessment items and examined the scoring accuracy. We then compared its performance with G-BERT. Our findings reveal a substantial improvement in scoring accuracy with G-SciEdBERT, demonstrating a 10% increase of quadratic weighted kappa compared to G-BERT (mean accuracy difference = 0.096, SD = 0.024). These insights underline the significance of specialized language models like G-SciEdBERT, which is trained to enhance the accuracy of automated scoring, offering a substantial contribution to the field of AI in education.

High-quality and consistent annotations are fundamental to the successful development of robust machine learning models. Traditional data annotation methods are resource-intensive and inefficient, often leading to a reliance on third-party annotators who are not the domain experts. Hard samples, which are usually the most informative for model training, tend to be difficult to label accurately and consistently without business context. These can arise unpredictably during the annotation process, requiring a variable number of iterations and rounds of feedback, leading to unforeseen expenses and time commitments to guarantee quality. We posit that more direct involvement of domain experts, using a human-in-the-loop system, can resolve many of these practical challenges. We propose a novel framework we call Video Annotator (VA) for annotating, managing, and iterating on video classification datasets. Our approach offers a new paradigm for an end-user-centered model development process, enhancing the efficiency, usability, and effectiveness of video classifiers. Uniquely, VA allows for a continuous annotation process, seamlessly integrating data collection and model training. We leverage the zero-shot capabilities of vision-language foundation models combined with active learning techniques, and demonstrate that VA enables the efficient creation of high-quality models. VA achieves a median 6.8 point improvement in Average Precision relative to the most competitive baseline across a wide-ranging assortment of tasks. We release a dataset with 153k labels across 56 video understanding tasks annotated by three professional video editors using VA, and also release code to replicate our experiments at: http://github.com/netflix/videoannotator.