Large language models (LLMs) are widely adopted in knowledge-intensive tasks and have achieved impressive performance thanks to their knowledge abilities. While LLMs have demonstrated outstanding performance on atomic or linear (multi-hop) QA tasks, whether they can reason in knowledge-rich scenarios with interweaving constraints remains an underexplored problem. In this work, we propose geometric reasoning over structured knowledge, where pieces of knowledge are connected in a graph structure and models need to fill in the missing information. Such geometric knowledge reasoning would require the ability to handle structured knowledge, reason with uncertainty, verify facts, and backtrack when an error occurs. We propose Knowledge Crosswords, a multi-blank QA dataset where each problem consists of a natural language question representing the geometric constraints of an incomplete entity network, where LLMs are tasked with working out the missing entities while meeting all factual constraints. Knowledge Crosswords contains 2,101 individual problems, covering various knowledge domains and further divided into three difficulty levels. We conduct extensive experiments to evaluate existing LLM prompting approaches on the Knowledge Crosswords benchmark. We additionally propose two new approaches, Staged Prompting and Verify-All, to augment LLMs' ability to backtrack and verify structured constraints. Our results demonstrate that while baseline approaches perform well on easier problems but struggle with hard ones, our proposed Verify-All outperforms other methods by a large margin and is more robust with hard problems. Further analysis reveals that LLMs' ability of geometric reasoning over structured knowledge is still far from robust or perfect, susceptible to confounders such as the order of options, certain structural patterns, assumption of existence of correct answer, and more.
Large language models (LLMs) often encounter knowledge conflicts, scenarios where discrepancy arises between the internal parametric knowledge of LLMs and non-parametric information provided in the prompt context. In this work we ask what are the desiderata for LLMs when a knowledge conflict arises and whether existing LLMs fulfill them. We posit that LLMs should 1) identify knowledge conflicts, 2) pinpoint conflicting information segments, and 3) provide distinct answers or viewpoints in conflicting scenarios. To this end, we introduce KNOWLEDGE CONFLICT, an evaluation framework for simulating contextual knowledge conflicts and quantitatively evaluating to what extent LLMs achieve these goals. KNOWLEDGE CONFLICT includes diverse and complex situations of knowledge conflict, knowledge from diverse entities and domains, two synthetic conflict creation methods, and settings with progressively increasing difficulty to reflect realistic knowledge conflicts. Extensive experiments with the KNOWLEDGE CONFLICT framework reveal that while LLMs perform well in identifying the existence of knowledge conflicts, they struggle to determine the specific conflicting knowledge and produce a response with distinct answers amidst conflicting information. To address these challenges, we propose new instruction-based approaches that augment LLMs to better achieve the three goals. Further analysis shows that abilities to tackle knowledge conflicts are greatly impacted by factors such as knowledge domain and prompt text, while generating robust responses to knowledge conflict scenarios remains an open research question.
In the current user-server interaction paradigm of prompted generation with large language models (LLM) on cloud, the server fully controls the generation process, which leaves zero options for users who want to keep the generated text to themselves. We propose LatticeGen, a cooperative framework in which the server still handles most of the computation while the user controls the sampling operation. The key idea is that the true generated sequence is mixed with noise tokens by the user and hidden in a noised lattice. Considering potential attacks from a hypothetically malicious server and how the user can defend against it, we propose the repeated beam-search attack and the mixing noise scheme. In our experiments we apply LatticeGen to protect both prompt and generation. It is shown that while the noised lattice degrades generation quality, LatticeGen successfully protects the true generation to a remarkable degree under strong attacks (more than 50% of the semantic remains hidden as measured by BERTScore).
In-context learning (ICL) improves language models' performance on a variety of NLP tasks by simply demonstrating a handful of examples at inference time. It is not well understood why ICL ability emerges, as the model has never been specifically trained on such demonstrations. Unlike prior work that explores implicit mechanisms behind ICL, we study ICL via investigating the pretraining data. Specifically, we first adapt an iterative, gradient-based approach to find a small subset of pretraining data that supports ICL. We observe that a continued pretraining on this small subset significantly improves the model's ICL ability, by up to 18%. We then compare the supportive subset constrastively with random subsets of pretraining data and discover: (1) The supportive pretraining data to ICL do not have a higher domain relevance to downstream tasks. (2) The supportive pretraining data have a higher mass of rarely occurring, long-tail tokens. (3) The supportive pretraining data are challenging examples where the information gain from long-range context is below average, indicating learning to incorporate difficult long-range context encourages ICL. Our work takes a first step towards understanding ICL via analyzing instance-level pretraining data. Our insights have a potential to enhance the ICL ability of language models by actively guiding the construction of pretraining data in the future.
Theory of Mind (ToM)$\unicode{x2014}$the ability to reason about the mental states of other people$\unicode{x2014}$is a key element of our social intelligence. Yet, despite their ever more impressive performance, large-scale neural language models still lack basic theory of mind capabilities out-of-the-box. We posit that simply scaling up models will not imbue them with theory of mind due to the inherently symbolic and implicit nature of the phenomenon, and instead investigate an alternative: can we design a decoding-time algorithm that enhances theory of mind of off-the-shelf neural language models without explicit supervision? We present SymbolicToM, a plug-and-play approach to reason about the belief states of multiple characters in reading comprehension tasks via explicit symbolic representation. More concretely, our approach tracks each entity's beliefs, their estimation of other entities' beliefs, and higher-order levels of reasoning, all through graphical representations, allowing for more precise and interpretable reasoning than previous approaches. Empirical results on the well-known ToMi benchmark (Le et al., 2019) demonstrate that SymbolicToM dramatically enhances off-the-shelf neural networks' theory of mind in a zero-shot setting while showing robust out-of-distribution performance compared to supervised baselines. Our work also reveals spurious patterns in existing theory of mind benchmarks, emphasizing the importance of out-of-distribution evaluation and methods that do not overfit a particular dataset.
Although much literature has established the presence of demographic bias in natural language processing (NLP) models, most work relies on curated bias metrics that may not be reflective of real-world applications. At the same time, practitioners are increasingly using algorithmic tools in high-stakes settings, with particular recent interest in NLP. In this work, we focus on one such setting: child protective services (CPS). CPS workers often write copious free-form text notes about families they are working with, and CPS agencies are actively seeking to deploy NLP models to leverage these data. Given well-established racial bias in this setting, we investigate possible ways deployed NLP is liable to increase racial disparities. We specifically examine word statistics within notes and algorithmic fairness in risk prediction, coreference resolution, and named entity recognition (NER). We document consistent algorithmic unfairness in NER models, possible algorithmic unfairness in coreference resolution models, and little evidence of exacerbated racial bias in risk prediction. While there is existing pronounced criticism of risk prediction, our results expose previously undocumented risks of racial bias in realistic information extraction systems, highlighting potential concerns in deploying them, even though they may appear more benign. Our work serves as a rare realistic examination of NLP algorithmic fairness in a potential deployed setting and a timely investigation of a specific risk associated with deploying NLP in CPS settings.
Despite remarkable advancements in few-shot generalization in natural language processing, most models are developed and evaluated primarily in English. To facilitate research on few-shot cross-lingual transfer, we introduce a new benchmark, called BUFFET, which unifies 15 diverse tasks across 54 languages in a sequence-to-sequence format and provides a fixed set of few-shot examples and instructions. BUFFET is designed to establish a rigorous and equitable evaluation framework for few-shot cross-lingual transfer across a broad range of tasks and languages. Using BUFFET, we perform thorough evaluations of state-of-the-art multilingual large language models with different transfer methods, namely in-context learning and fine-tuning. Our findings reveal significant room for improvement in few-shot in-context cross-lingual transfer. In particular, ChatGPT with in-context learning often performs worse than much smaller mT5-base models fine-tuned on English task data and few-shot in-language examples. Our analysis suggests various avenues for future research in few-shot cross-lingual transfer, such as improved pretraining, understanding, and future evaluations.
Diffusion-based language models (LMs) have been shown to be competent generative models that are easy to control at inference and are a promising alternative to autoregressive LMs. While autoregressive LMs have benefited immensely from scaling and instruction-based learning, existing studies on diffusion LMs have been conducted on a relatively smaller scale. Starting with a recently proposed diffusion model SSD-LM, in this work we explore methods to scale it from 0.4B to 13B parameters, proposing several techniques to improve its training and inference efficiency. We call the new model SSD-2. We further show that this model can be easily finetuned to follow instructions. Finally, leveraging diffusion models' capability at inference-time control, we show that SSD-2 facilitates novel ensembles with 100x smaller models that can be customized and deployed by individual users. We find that compared to autoregressive models, the collaboration between diffusion models is more effective, leading to higher-quality and more relevant model responses due to their ability to incorporate bi-directional contexts.
Language models (LMs) often struggle to pay enough attention to the input context, and generate texts that are unfaithful or contain hallucinations. To mitigate this issue, we present context-aware decoding (CAD), which follows a contrastive output distribution that amplifies the difference between the output probabilities when a model is used with and without context. Our experiments show that CAD, without additional training, significantly improves the faithfulness of different LM families, including OPT, GPT, LLaMA and FLAN-T5 for summarization tasks (e.g., 14.3% gain for LLaMA in factuality metrics). Furthermore, CAD is particularly effective in overriding a model's prior knowledge when it contradicts the provided context, leading to substantial improvements in tasks where resolving the knowledge conflict is essential.
Despite the major advances in NLP, significant disparities in NLP system performance across languages still exist. Arguably, these are due to uneven resource allocation and sub-optimal incentives to work on less resourced languages. To track and further incentivize the global development of equitable language technology, we introduce GlobalBench. Prior multilingual benchmarks are static and have focused on a limited number of tasks and languages. In contrast, GlobalBench is an ever-expanding collection that aims to dynamically track progress on all NLP datasets in all languages. Rather than solely measuring accuracy, GlobalBench also tracks the estimated per-speaker utility and equity of technology across all languages, providing a multi-faceted view of how language technology is serving people of the world. Furthermore, GlobalBench is designed to identify the most under-served languages, and rewards research efforts directed towards those languages. At present, the most under-served languages are the ones with a relatively high population, but nonetheless overlooked by composite multilingual benchmarks (like Punjabi, Portuguese, and Wu Chinese). Currently, GlobalBench covers 966 datasets in 190 languages, and has 1,128 system submissions spanning 62 languages.