Instruction Tuning (IT), the process of training large language models (LLMs) using instruction-response pairs, has emerged as the predominant method for transforming base pre-trained LLMs into open-domain conversational agents. While IT has achieved notable success and widespread adoption, its limitations and shortcomings remain underexplored. In this paper, through rigorous experiments and an in-depth analysis of the changes LLMs undergo through IT, we reveal various limitations of IT. In particular, we show that (1) IT fails to enhance knowledge or skills in LLMs. LoRA fine-tuning is limited to learning response initiation and style tokens, and full-parameter fine-tuning leads to knowledge degradation. (2) Copying response patterns from IT datasets derived from knowledgeable sources leads to a decline in response quality. (3) Full-parameter fine-tuning increases hallucination by inaccurately borrowing tokens from conceptually similar instances in the IT dataset for generating responses. (4) Popular methods to improve IT do not lead to performance improvements over a simple LoRA fine-tuned model. Our findings reveal that responses generated solely from pre-trained knowledge consistently outperform responses by models that learn any form of new knowledge from IT on open-source datasets. We hope the insights and challenges revealed inspire future work.

Hypertension, defined as blood pressure (BP) that is above normal, holds paramount significance in the realm of public health, as it serves as a critical precursor to various cardiovascular diseases (CVDs) and significantly contributes to elevated mortality rates worldwide. However, many existing BP measurement technologies and standards might be biased because they do not consider clinical outcomes, comorbidities, or demographic factors, making them inconclusive for diagnostic purposes. There is limited data-driven research focused on studying the variance in BP measurements across these variables. In this work, we employed GPT-35-turbo, a large language model (LLM), to automatically extract the mean and standard deviation values of BP for both males and females from a dataset comprising 25 million abstracts sourced from PubMed. 993 article abstracts met our predefined inclusion criteria (i.e., presence of references to blood pressure, units of blood pressure such as mmHg, and mention of biological sex). Based on the automatically-extracted information from these articles, we conducted an analysis of the variations of BP values across biological sex. Our results showed the viability of utilizing LLMs to study the BP variations across different demographic factors.

Chaining language model (LM) calls as composable modules is fueling a new way of programming, but ensuring LMs adhere to important constraints requires heuristic "prompt engineering". We introduce LM Assertions, a programming construct for expressing computational constraints that LMs should satisfy. We integrate our constructs into the recent DSPy programming model for LMs, and present new strategies that allow DSPy to compile programs with LM Assertions into more reliable and accurate systems. We also propose strategies to use assertions at inference time for automatic self-refinement with LMs. We report on four diverse case studies for text generation and find that LM Assertions improve not only compliance with imposed rules but also downstream task performance, passing constraints up to 164% more often and generating up to 37% more higher-quality responses. Our reference implementation of LM Assertions is integrated into DSPy at https://github.com/stanfordnlp/dspy

ChatGPT and other general large language models (LLMs) have achieved remarkable success, but they have also raised concerns about the misuse of AI-generated texts. Existing AI-generated text detection models, such as based on BERT and RoBERTa, are prone to in-domain over-fitting, leading to poor out-of-domain (OOD) detection performance. In this paper, we first collected Chinese text responses generated by human experts and 9 types of LLMs, for which to multiple domains questions, and further created a dataset that mixed human-written sentences and sentences polished by LLMs. We then proposed LLM-Detector, a novel method for both document-level and sentence-level text detection through Instruction Tuning of LLMs. Our method leverages the wealth of knowledge LLMs acquire during pre-training, enabling them to detect the text they generate. Instruction tuning aligns the model's responses with the user's expected text detection tasks. Experimental results show that previous methods struggle with sentence-level AI-generated text detection and OOD detection. In contrast, our proposed method not only significantly outperforms baseline methods in both sentence-level and document-level text detection but also demonstrates strong generalization capabilities. Furthermore, since LLM-Detector is trained based on open-source LLMs, it is easy to customize for deployment.

Recent developments in long-context large language models have attracted considerable attention. Yet, their real-world applications are often hindered by ineffective context information use. This work shows that structuring training data to increase semantic interdependence is an effective strategy for optimizing context utilization. To this end, we introduce Structured Packing for Long Context (SPLiCe), a method for creating training examples by using information retrieval methods to collate mutually relevant documents into a single training context. We empirically validate SPLiCe on large $3$B and $7$B models, showing perplexity improvements and better long-context utilization on downstream tasks. Remarkably, already relatively short fine-tuning with SPLiCe is enough to attain these benefits. Additionally, the comprehensive study of SPLiCe reveals intriguing transfer effects such as training on code data leading to perplexity improvements on text data.

Domain randomization is an effective computer vision technique for improving transferability of vision models across visually distinct domains exhibiting similar content. Existing approaches, however, rely extensively on tweaking complex and specialized simulation engines that are difficult to construct, subsequently affecting their feasibility and scalability. This paper introduces BehAVE, a video understanding framework that uniquely leverages the plethora of existing commercial video games for domain randomization, without requiring access to their simulation engines. Under BehAVE (1) the inherent rich visual diversity of video games acts as the source of randomization and (2) player behavior -- represented semantically via textual descriptions of actions -- guides the *alignment* of videos with similar content. We test BehAVE on 25 games of the first-person shooter (FPS) genre across various video and text foundation models and we report its robustness for domain randomization. BehAVE successfully aligns player behavioral patterns and is able to zero-shot transfer them to multiple unseen FPS games when trained on just one FPS game. In a more challenging setting, BehAVE manages to improve the zero-shot transferability of foundation models to unseen FPS games (up to 22%) even when trained on a game of a different genre (Minecraft). Code and dataset can be found at https://github.com/nrasajski/BehAVE.

Hallucinations in large language models (LLMs) are always seen as limitations. However, could they also be a source of creativity? This survey explores this possibility, suggesting that hallucinations may contribute to LLM application by fostering creativity. This survey begins with a review of the taxonomy of hallucinations and their negative impact on LLM reliability in critical applications. Then, through historical examples and recent relevant theories, the survey explores the potential creative benefits of hallucinations in LLMs. To elucidate the value and evaluation criteria of this connection, we delve into the definitions and assessment methods of creativity. Following the framework of divergent and convergent thinking phases, the survey systematically reviews the literature on transforming and harnessing hallucinations for creativity in LLMs. Finally, the survey discusses future research directions, emphasizing the need to further explore and refine the application of hallucinations in creative processes within LLMs.

We recently developed SLM, a joint speech and language model, which fuses a pretrained foundational speech model and a large language model (LLM), while preserving the in-context learning capability intrinsic to the pretrained LLM. In this paper, we apply SLM to speech dialog applications where the dialog states are inferred directly from the audio signal. Task-oriented dialogs often contain domain-specific entities, i.e., restaurants, hotels, train stations, and city names, which are difficult to recognize, however, critical for the downstream applications. Inspired by the RAG (retrieval-augmented generation) paradigm, we propose a retrieval augmented SLM (ReSLM) that overcomes this weakness. We first train a speech retriever to retrieve text entities mentioned in the audio. The retrieved entities are then added as text inputs to the underlying SLM to bias model predictions. We evaluated ReSLM on speech MultiWoz task (DSTC-11 challenge), and found that this retrieval augmentation boosts model performance, achieving joint goal accuracy (38.6% vs 32.7%), slot error rate (20.6% vs 24.8%) and ASR word error rate (5.5% vs 6.7%). While demonstrated on dialog state tracking, our approach is broadly applicable to other speech tasks requiring contextual information or domain-specific entities, such as contextual ASR with biasing capability.

Large language models (LLMs) have achieved widespread success on a variety of in-context few-shot tasks, but this success is typically evaluated via correctness rather than consistency. We argue that self-consistency is an important criteria for valid multi-step reasoning in tasks where the solution is composed of the answers to multiple sub-steps. We propose two types of self-consistency that are particularly important for multi-step reasoning -- hypothetical consistency (a model's ability to predict what its output would be in a hypothetical other context) and compositional consistency (consistency of a model's final outputs when intermediate sub-steps are replaced with the model's outputs for those steps). We demonstrate that multiple variants of the GPT-3/-4 models exhibit poor consistency rates across both types of consistency on a variety of tasks.

Learning from human preference has been considered key to aligning Large Language Models (LLMs) with human values. However, contrary to popular belief, our preliminary study reveals that reward models trained on human preference datasets tend to give higher scores to long off-topic responses than short on-topic ones. Motivated by this observation, we explore a preference-free approach utilizing `relevance' as a key objective for alignment. On our first attempt, we find that the relevance score obtained by a retriever alone is vulnerable to reward hacking, i.e., overoptimizing to undesired shortcuts, when we utilize the score as a reward for reinforcement learning. To mitigate it, we integrate effective inductive biases into the vanilla relevance to regularize each other, resulting in a mixture of reward functions: Regularized Relevance Reward ($R^3$). $R^3$ significantly improves performance on preference benchmarks by providing a robust reward signal. Notably, $R^3$ does not require any human preference datasets (i.e., preference-free), outperforming open-source reward models in improving human preference. Our analysis demonstrates that $R^3$ has advantages in elevating human preference while minimizing its side effects. Finally, we show the generalizability of $R^3$, consistently improving instruction-tuned models in various backbones and sizes without additional dataset cost. Our code is available at https://github.com/naver-ai/RRR.