Large language models (LLMs) often demonstrate inconsistencies with human preferences. Previous research typically gathered human preference data and then aligned the pre-trained models using reinforcement learning or instruction tuning, a.k.a. the finetuning step. In contrast, aligning frozen LLMs without requiring alignment data is more appealing. This work explores the potential of the latter setting. We discover that by integrating self-evaluation and rewind mechanisms, unaligned LLMs can directly produce responses consistent with human preferences via self-boosting. We introduce a novel inference method, Rewindable Auto-regressive INference (RAIN), that allows pre-trained LLMs to evaluate their own generation and use the evaluation results to guide rewind and generation for AI safety. Notably, RAIN operates without the need of extra data for model alignment and abstains from any training, gradient computation, or parameter updates. Experimental results evaluated by GPT-4 and humans demonstrate the effectiveness of RAIN: on the HH dataset, RAIN improves the harmlessness rate of LLaMA 30B from 82% of vanilla inference to 97%, while maintaining the helpfulness rate. On the TruthfulQA dataset, RAIN improves the truthfulness of the already-well-aligned LLaMA-2-chat 13B model by 5%.

Learning to predict masked tokens in a sequence has been shown to be a powerful pretraining objective for large language models. After training, such masked language models can provide distributions of tokens conditioned on bidirectional context. In this paper, we show that contrary to popular assumptions, such bidirectional conditionals often demonstrate considerable inconsistencies, i.e., they cannot be derived from a coherent joint distribution when considered together. We empirically quantify such inconsistencies in the simple scenario of bigram comparison for two common styles of masked language models: T5-style and BERT-style. For example, we show that T5 models often confuse their own preference regarding two similar bigrams. We show that inconsistencies exist ubiquitously in masked language models of diverse sizes and configurations, from RoBERTa-base to GLM-130B. As an initial attempt to address this issue during the inference phase, we propose Ensemble of Conditionals, a self-ensemble algorithm that jointly considers many inconsistent conditionals directly produced by the MLM to synthesize a distribution that is used as the model's final output. Such ensembling improves open-source SOTA results on LAMBADA.

I shall quantify the logical faults generated by ChatGPT when applied to reasoning tasks. For experiments, I use the 144 puzzles from the library \url{https://users.utcluj.ro/~agroza/puzzles/maloga}~\cite{groza:fol}. The library contains puzzles of various types, including arithmetic puzzles, logical equations, Sudoku-like puzzles, zebra-like puzzles, truth-telling puzzles, grid puzzles, strange numbers, or self-reference puzzles. The correct solutions for these puzzles were checked using the theorem prover Prover9~\cite{mccune2005release} and the finite models finder Mace4~\cite{mccune2003mace4} based on human-modelling in Equational First Order Logic. A first output of this study is the benchmark of 100 logical puzzles. For this dataset ChatGPT provided both correct answer and justification for 7\% only. %, while BARD for 5\%. Since the dataset seems challenging, the researchers are invited to test the dataset on more advanced or tuned models than ChatGPT3.5 with more crafted prompts. A second output is the classification of reasoning faults conveyed by ChatGPT. This classification forms a basis for a taxonomy of reasoning faults generated by large language models. I have identified 67 such logical faults, among which: inconsistencies, implication does not hold, unsupported claim, lack of commonsense, wrong justification. The 100 solutions generated by ChatGPT contain 698 logical faults. That is on average, 7 fallacies for each reasoning task. A third ouput is the annotated answers of the ChatGPT with the corresponding logical faults. Each wrong statement within the ChatGPT answer was manually annotated, aiming to quantify the amount of faulty text generated by the language model. On average, 26.03\% from the generated text was a logical fault.

Information retrieval involves selecting artifacts from a corpus that are most relevant to a given search query. The flavor of retrieval typically used in classical applications can be termed as homogeneous and relaxed, where queries and corpus elements are both natural language (NL) utterances (homogeneous) and the goal is to pick most relevant elements from the corpus in the Top-K, where K is large, such as 10, 25, 50 or even 100 (relaxed). Recently, retrieval is being used extensively in preparing prompts for large language models (LLMs) to enable LLMs to perform targeted tasks. These new applications of retrieval are often heterogeneous and strict -- the queries and the corpus contain different kinds of entities, such as NL and code, and there is a need for improving retrieval at Top-K for small values of K, such as K=1 or 3 or 5. Current dense retrieval techniques based on pretrained embeddings provide a general-purpose and powerful approach for retrieval, but they are oblivious to task-specific notions of similarity of heterogeneous artifacts. We introduce Adapted Dense Retrieval, a mechanism to transform embeddings to enable improved task-specific, heterogeneous and strict retrieval. Adapted Dense Retrieval works by learning a low-rank residual adaptation of the pretrained black-box embedding. We empirically validate our approach by showing improvements over the state-of-the-art general-purpose embeddings-based baseline.

Model alignment with human preferences is an essential step in making Large Language Models (LLMs) helpful and consistent with human values. It typically consists of supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF) stages. However, RLHF faces inherent limitations stemming from a complex training setup and its tendency to align the model with implicit values that end users cannot control at run-time. Moreover, reward models in RLHF stage commonly rely on single-dimensional feedback as opposed to explicit, multifaceted signals that indicate attributes such as helpfulness, humor, and toxicity. To address these limitations, we propose SteerLM, a supervised fine-tuning method that empowers end-users to control responses during inference. SteerLM conditions responses to conform to an explicitly defined multi-dimensional set of attributes, thereby empowering a steerable AI capable of generating helpful and high-quality responses while maintaining customizability. Experiments show that SteerLM trained on open source datasets generates responses that are preferred by human and automatic evaluators to many state-of-the-art baselines trained with RLHF while being much easier to train. Try SteerLM at https://huggingface.co/nvidia/SteerLM-llama2-13B

In this paper, we propose LAN-grasp, a novel approach towards more appropriate semantic grasping. We use foundation models to provide the robot with a deeper understanding of the objects, the right place to grasp an object, or even the parts to avoid. This allows our robot to grasp and utilize objects in a more meaningful and safe manner. We leverage the combination of a Large Language Model, a Vision Language Model, and a traditional grasp planner to generate grasps demonstrating a deeper semantic understanding of the objects. We first prompt the Large Language Model about which object part is appropriate for grasping. Next, the Vision Language Model identifies the corresponding part in the object image. Finally, we generate grasp proposals in the region proposed by the Vision Language Model. Building on foundation models provides us with a zero-shot grasp method that can handle a wide range of objects without the need for further training or fine-tuning. We evaluated our method in real-world experiments on a custom object data set. We present the results of a survey that asks the participants to choose an object part appropriate for grasping. The results show that the grasps generated by our method are consistently ranked higher by the participants than those generated by a conventional grasping planner and a recent semantic grasping approach.

In the mental health domain, Large Language Models (LLMs) offer promising new opportunities, though their inherent complexity and low controllability have raised questions about their suitability in clinical settings. We present MindfulDiary, a mobile journaling app incorporating an LLM to help psychiatric patients document daily experiences through conversation. Designed in collaboration with mental health professionals (MHPs), MindfulDiary takes a state-based approach to safely comply with the experts' guidelines while carrying on free-form conversations. Through a four-week field study involving 28 patients with major depressive disorder and five psychiatrists, we found that MindfulDiary supported patients in consistently enriching their daily records and helped psychiatrists better empathize with their patients through an understanding of their thoughts and daily contexts. Drawing on these findings, we discuss the implications of leveraging LLMs in the mental health domain, bridging the technical feasibility and their integration into clinical settings.

Large Language Models (LLMs) have emerged as dominant foundational models in modern NLP. However, the understanding of their prediction process and internal mechanisms, such as feed-forward networks and multi-head self-attention, remains largely unexplored. In this study, we probe LLMs from a human behavioral perspective, correlating values from LLMs with eye-tracking measures, which are widely recognized as meaningful indicators of reading patterns. Our findings reveal that LLMs exhibit a prediction pattern distinct from that of RNN-based LMs. Moreover, with the escalation of FFN layers, the capacity for memorization and linguistic knowledge encoding also surges until it peaks, subsequently pivoting to focus on comprehension capacity. The functions of self-attention are distributed across multiple heads. Lastly, we scrutinize the gate mechanisms, finding that they control the flow of information, with some gates promoting, while others eliminating information.

The extrapolation capability of Large Language Models (LLMs) based on Rotary Position Embedding Su et al. (2021) is currently a topic of considerable interest. In this work, we first observe that fine-tuning a RoPE-based LLM with either a smaller or larger base in pre-training context length could significantly enhance its extrapolation performance. After that, we propose Scaling Laws of RoPE-based Extrapolation, a unified framework from the periodic perspective, to describe the relationship between the extrapolation performance and base value as well as tuning context length. In this process, we also explain the origin of the RoPE-based extrapolation issue by critical dimension for extrapolation. Besides these observations and analyses, we achieve extrapolation up to 1 million context length within only 16K training length on LLaMA2 7B and 13B (Touvron et al., 2023b). Large Language Models (LLMs) have become the dominant architecture in a variety of natural language processing tasks(OpenAI, 2023; Touvron et al., 2023a;b), while Transformers (Vaswani et al., 2017) based on Rotary Position Embedding (RoPE) (Su et al., 2021) have become the dominant backbone in wide range of LLM design (Chowdhery et al., 2022; Nijkamp et al., 2022; Touvron et al., 2023a;b). While RoPE can theoretically represent sequences through trigonometric functions, as detailed in Appendix A, its performance drops when the input sequence or context length surpasses the training length(Press et al., 2021; Chen et al., 2023), seen in Figure 1. Concerning the extrapolation issue with RoPE, different works have provided various interpretations and corresponding solving attempts. These works could divided into two schools of thought. One limits the scope of self-attention (Ratner et al., 2022; Han et al., 2023) given the fact that selfattention computations in RoPE fail to keep stable beyond training context and exhibit attention RoPE fine-tuned with either a smaller or larger base on the original training length of 4K or a much longer context of 16K, could outperform other extrapolation strategies and extrapolate to 100K context length. The other aims to capture longer contexts by using smaller rotation angles and longer fine-tuning context (Chen et al., 2023; Peng et al., 2023). Currently, popular methods, such as Dynamic NTK (Local-LLaMA, 2023a) and Code LLaMA (Rozière et al., 2023), mainly come from the second approach. Both approaches adapt RoPE to longer contexts with a larger rotary base.

Zero-shot coordination (ZSC) is a new challenge focusing on generalizing learned coordination skills to unseen partners. Existing methods train the ego agent with partners from pre-trained or evolving populations. The agent's ZSC capability is typically evaluated with a few evaluation partners, including humans and agents, and reported by mean returns. Current evaluation methods for ZSC capability still need improvement in constructing diverse evaluation partners and comprehensively measuring ZSC capability. In this paper, we aim to create a reliable, comprehensive, and efficient evaluation method for ZSC capability. We formally define the ideal'diversity-complete' evaluation partners and propose the best response (BR) diversity, which is the population diversity of the BRs to the partners, to approximate the ideal evaluation partners. We propose an evaluation workflow including'diversity-complete' evaluation partners construction and a multidimensional metric, the Best Response Proximity (BR-Prox) metric. We re-evaluate strong ZSC methods in the Overcooked environment using the proposed evaluation workflow. Surprisingly, the results in some of the most used layouts fail to distinguish the performance of different ZSC methods. Moreover, the evaluated ZSC methods lack the ability to produce enough diverse and high-performing training partners. Our proposed evaluation workflow calls for a change in how we efficiently evaluate ZSC methods as a supplement to human evaluation. Zero-shot Coordination (ZSC) is a new challenge in training an agent named ego agent to have the capability to coordinate with unseen partners in cooperative AI (Hu et al., 2020).