Recent studies indicate that leveraging off-the-shelf or fine-tuned retrievers, capable of retrieving relevant in-context examples tailored to the input query, enhances few-shot in-context learning of English. However, adapting these methods to other languages, especially low-resource ones, poses challenges due to the scarcity of cross-lingual retrievers and annotated data. Thus, we introduce XAMPLER: Cross-Lingual Example Retrieval, a method tailored to tackle the challenge of cross-lingual in-context learning using only annotated English data. XAMPLER first trains a retriever based on Glot500, a multilingual small language model, using positive and negative English examples constructed from the predictions of a multilingual large language model, i.e., MaLA500. Leveraging the cross-lingual capacity of the retriever, it can directly retrieve English examples as few-shot examples for in-context learning of target languages. Experiments on the multilingual text classification benchmark SIB200 with 176 languages show that XAMPLER substantially improves the in-context learning performance across languages. Our code is available at \url{https://github.com/cisnlp/XAMPLER}.
In recent years, the rise of large language models (LLMs) has made it possible to directly achieve named entity recognition (NER) without any demonstration samples or only using a few samples through in-context learning (ICL). However, standard ICL only helps LLMs understand task instructions, format and input-label mapping, but neglects the particularity of the NER task itself. In this paper, we propose a new prompting framework P-ICL to better achieve NER with LLMs, in which some point entities are leveraged as the auxiliary information to recognize each entity type. With such significant information, the LLM can achieve entity classification more precisely. To obtain optimal point entities for prompting LLMs, we also proposed a point entity selection method based on K-Means clustering. Our extensive experiments on some representative NER benchmarks verify the effectiveness of our proposed strategies in P-ICL and point entity selection.
Attack knowledge graph construction seeks to convert textual cyber threat intelligence (CTI) reports into structured representations, portraying the evolutionary traces of cyber attacks. Even though previous research has proposed various methods to construct attack knowledge graphs, they generally suffer from limited generalization capability to diverse knowledge types as well as requirement of expertise in model design and tuning. Addressing these limitations, we seek to utilize Large Language Models (LLMs), which have achieved enormous success in a broad range of tasks given exceptional capabilities in both language understanding and zero-shot task fulfillment. Thus, we propose a fully automatic LLM-based framework to construct attack knowledge graphs named: AttacKG+. Our framework consists of four consecutive modules: rewriter, parser, identifier, and summarizer, each of which is implemented by instruction prompting and in-context learning empowered by LLMs. Furthermore, we upgrade the existing attack knowledge schema and propose a comprehensive version. We represent a cyber attack as a temporally unfolding event, each temporal step of which encapsulates three layers of representation, including behavior graph, MITRE TTP labels, and state summary. Extensive evaluation demonstrates that: 1) our formulation seamlessly satisfies the information needs in threat event analysis, 2) our construction framework is effective in faithfully and accurately extracting the information defined by AttacKG+, and 3) our attack graph directly benefits downstream security practices such as attack reconstruction. All the code and datasets will be released upon acceptance.
Numerous methods have been proposed to detect, estimate, and analyze properties of people in images, including the estimation of 3D pose, shape, contact, human-object interaction, emotion, and more. Each of these methods works in isolation instead of synergistically. Here we address this problem and build a language-driven human understanding system -- ChatHuman, which combines and integrates the skills of many different methods. To do so, we finetune a Large Language Model (LLM) to select and use a wide variety of existing tools in response to user inputs. In doing so, ChatHuman is able to combine information from multiple tools to solve problems more accurately than the individual tools themselves and to leverage tool output to improve its ability to reason about humans. The novel features of ChatHuman include leveraging academic publications to guide the application of 3D human-related tools, employing a retrieval-augmented generation model to generate in-context-learning examples for handling new tools, and discriminating and integrating tool results to enhance 3D human understanding. Our experiments show that ChatHuman outperforms existing models in both tool selection accuracy and performance across multiple 3D human-related tasks. ChatHuman is a step towards consolidating diverse methods for human analysis into a single, powerful, system for 3D human reasoning.
The progression to "Pervasive Augmented Reality" envisions easy access to multimodal information continuously. However, in many everyday scenarios, users are occupied physically, cognitively or socially. This may increase the friction to act upon the multimodal information that users encounter in the world. To reduce such friction, future interactive interfaces should intelligently provide quick access to digital actions based on users' context. To explore the range of possible digital actions, we conducted a diary study that required participants to capture and share the media that they intended to perform actions on (e.g., images or audio), along with their desired actions and other contextual information. Using this data, we generated a holistic design space of digital follow-up actions that could be performed in response to different types of multimodal sensory inputs. We then designed OmniActions, a pipeline powered by large language models (LLMs) that processes multimodal sensory inputs and predicts follow-up actions on the target information grounded in the derived design space. Using the empirical data collected in the diary study, we performed quantitative evaluations on three variations of LLM techniques (intent classification, in-context learning and finetuning) and identified the most effective technique for our task. Additionally, as an instantiation of the pipeline, we developed an interactive prototype and reported preliminary user feedback about how people perceive and react to the action predictions and its errors.
One of the key challenges in current Reinforcement Learning (RL)-based Automated Driving (AD) agents is achieving flexible, precise, and human-like behavior cost-effectively. This paper introduces an innovative approach that uses large language models (LLMs) to intuitively and effectively optimize RL reward functions in a human-centric way. We developed a framework where instructions and dynamic environment descriptions are input into the LLM. The LLM then utilizes this information to assist in generating rewards, thereby steering the behavior of RL agents towards patterns that more closely resemble human driving. The experimental results demonstrate that this approach not only makes RL agents more anthropomorphic but also achieves better performance. Additionally, various strategies for reward-proxy and reward-shaping are investigated, revealing the significant impact of prompt design on shaping an AD vehicle's behavior. These findings offer a promising direction for the development of more advanced, human-like automated driving systems. Our experimental data and source code can be found here
Large pretrained language models (LLMs) have shown surprising In-Context Learning (ICL) ability. An important application in deploying large language models is to augment LLMs with a private database for some specific task. The main problem with this promising commercial use is that LLMs have been shown to memorize their training data and their prompt data are vulnerable to membership inference attacks (MIA) and prompt leaking attacks. In order to deal with this problem, we treat LLMs as untrusted in privacy and propose a locally differentially private framework of in-context learning(LDP-ICL) in the settings where labels are sensitive. Considering the mechanisms of in-context learning in Transformers by gradient descent, we provide an analysis of the trade-off between privacy and utility in such LDP-ICL for classification. Moreover, we apply LDP-ICL to the discrete distribution estimation problem. In the end, we perform several experiments to demonstrate our analysis results.
Recently, Large Language Models (LLMs) have been demonstrated to possess impressive capabilities in a variety of domains and tasks. We investigate the issue of prompt design in the multi-turn text-to-SQL task and attempt to enhance the LLMs' reasoning capacity when generating SQL queries. In the conversational context, the current SQL query can be modified from the preceding SQL query with only a few operations due to the context dependency. We introduce our method called CoE-SQL which can prompt LLMs to generate the SQL query based on the previously generated SQL query with an edition chain. We also conduct extensive ablation studies to determine the optimal configuration of our approach. Our approach outperforms different in-context learning baselines stably and achieves state-of-the-art performances on two benchmarks SParC and CoSQL using LLMs, which is also competitive to the SOTA fine-tuned models.
With the deluge of information delivered by the daily news cycle, there is a growing need to effectively and efficiently summarize news feeds for quick consumption. We leverage large language models (LLMs), with their advanced learning and generative abilities as compared to conventional language models, to generate concise and coherent summaries for news articles from the XSum dataset. Our paper focuses on two key aspects of LLMs: Efficient in-context Learning (ELearn) and Parameter Efficient Fine-tuning (EFit). Under ELearn, we find that increasing the number of shots in prompts and utilizing simple templates generally improve the quality of summaries. We also find that utilizing relevant examples in few-shot learning for ELearn does not improve model performance. In addition, we studied EFit using different methods and demonstrate that fine-tuning the first layer of LLMs produces better outcomes as compared to fine-tuning other layers or utilizing LoRA. We also find that leveraging more relevant training samples using selective layers does not result in better performance. By combining ELearn and EFit, we create a new model (ELearnFit) that leverages the benefits of both few-shot learning and fine-tuning and produces superior performance to either model alone. We also use ELearnFit to highlight the trade-offs between prompting and fine-tuning, especially for situations where only a limited number of annotated samples are available. Ultimately, our research provides practical techniques to optimize news summarization during the prompting and fine-tuning stages and enhances the synthesis of news articles.
With recent advances in large language models (LLMs), this paper explores the potential of leveraging state-of-the-art LLMs,such as GPT-4, to transfer existing human-written properties (e.g.,those from Certora auditing reports) and automatically generate customized properties for unknown code. To this end, we embed existing properties into a vector database and retrieve a reference property for LLM-based in-context learning to generate a new property for a given code. While this basic process is relatively straightforward, ensuring that the generated properties are (i) compilable, (ii) appropriate, and (iii) verifiable presents challenges. To address (i), we use the compilation and static analysis feedback as an external oracle to guide LLMs in iteratively revising the generated properties. For (ii), we consider multiple dimensions of similarity to rank the properties and employ a weighted algorithm to identify the top-K properties as the final result. For (iii), we design a dedicated prover to formally verify the correctness of the generated properties. We have implemented these strategies into a novel LLM-based property generation tool called PropertyGPT. Our experiments show that PropertyGPT can generate comprehensive and high-quality properties, achieving an 80% recall compared to the ground truth. It successfully detected 26 CVEs/attack incidents out of 37 tested and also uncovered 12 zero-day vulnerabilities, leading to $8,256 in bug bounty rewards.
Large Language Models (LLMs) are trained on massive text corpora, which are encoded with diverse personality traits. This triggers an interesting goal of eliciting a desired personality trait from the LLM, and probing its behavioral preferences. Accordingly, we formalize the persona elicitation task, aiming to customize LLM behaviors to align with a target persona. We present Persona In-Context Learning (PICLe), a novel persona elicitation framework grounded in Bayesian inference. At the core, PICLe introduces a new ICL example selection criterion based on likelihood ratio, which is designed to optimally guide the model in eliciting a specific target persona. We demonstrate the effectiveness of PICLe through extensive comparisons against baseline methods across three contemporary LLMs. Code is available at https://github.com/deeplearning-wisc/picle.
The rapid adoption of large language models (LLMs) in recommender systems (RS) presents new challenges in understanding and evaluating their biases, which can result in unfairness or the amplification of stereotypes. Traditional fairness evaluations in RS primarily focus on collaborative filtering (CF) settings, which may not fully capture the complexities of LLMs, as these models often inherit biases from large, unregulated data. This paper proposes a normative framework to benchmark consumer fairness in LLM-powered recommender systems (RecLLMs). We critically examine how fairness norms in classical RS fall short in addressing the challenges posed by LLMs. We argue that this gap can lead to arbitrary conclusions about fairness, and we propose a more structured, formal approach to evaluate fairness in such systems. Our experiments on the MovieLens dataset on consumer fairness, using in-context learning (zero-shot vs. few-shot) reveal fairness deviations in age-based recommendations, particularly when additional contextual examples are introduced (ICL-2). Statistical significance tests confirm that these deviations are not random, highlighting the need for robust evaluation methods. While this work offers a preliminary discussion on a proposed normative framework, our hope is that it could provide a formal, principled approach for auditing and mitigating bias in RecLLMs. The code and dataset used for this work will be shared at "gihub-anonymized".
Digital assistants have become ubiquitous in e-commerce applications, following the recent advancements in Information Retrieval (IR), Natural Language Processing (NLP) and Generative Artificial Intelligence (AI). However, customers are often unsure or unaware of how to effectively converse with these assistants to meet their shopping needs. In this work, we emphasize the importance of providing customers a fast, easy to use, and natural way to interact with conversational shopping assistants. We propose a framework that employs Large Language Models (LLMs) to automatically generate contextual, useful, answerable, fluent and diverse questions about products, via in-context learning and supervised fine-tuning. Recommending these questions to customers as helpful suggestions or hints to both start and continue a conversation can result in a smoother and faster shopping experience with reduced conversation overhead and friction. We perform extensive offline evaluations, and discuss in detail about potential customer impact, and the type, length and latency of our generated product questions if incorporated into a real-world shopping assistant.
Recent advancements in Large Multimodal Models (LMMs) have attracted interest in their generalization capability with only a few samples in the prompt. This progress is particularly relevant to the medical domain, where the quality and sensitivity of data pose unique challenges for model training and application. However, the dependency on high-quality data for effective in-context learning raises questions about the feasibility of these models when encountering with the inevitable variations and errors inherent in real-world medical data. In this paper, we introduce MID-M, a novel framework that leverages the in-context learning capabilities of a general-domain Large Language Model (LLM) to process multimodal data via image descriptions. MID-M achieves a comparable or superior performance to task-specific fine-tuned LMMs and other general-domain ones, without the extensive domain-specific training or pre-training on multimodal data, with significantly fewer parameters. This highlights the potential of leveraging general-domain LLMs for domain-specific tasks and offers a sustainable and cost-effective alternative to traditional LMM developments. Moreover, the robustness of MID-M against data quality issues demonstrates its practical utility in real-world medical domain applications.
The last decade has seen blossoming research in deep learning theory attempting to answer, "Why does deep learning generalize?" A powerful shift in perspective precipitated this progress: the study of overparametrized models in the interpolation regime. In this paper, we argue that another perspective shift is due, since some of the desirable qualities of LLMs are not a consequence of good statistical generalization and require a separate theoretical explanation. Our core argument relies on the observation that AR probabilistic models are inherently non-identifiable: models zero or near-zero KL divergence apart -- thus, equivalent test loss -- can exhibit markedly different behaviors. We support our position with mathematical examples and empirical observations, illustrating why non-identifiability has practical relevance through three case studies: (1) the non-identifiability of zero-shot rule extrapolation; (2) the approximate non-identifiability of in-context learning; and (3) the non-identifiability of fine-tunability. We review promising research directions focusing on LLM-relevant generalization measures, transferability, and inductive biases.
Large Language Models (LLMs) are deep learning models designed to generate text based on textual input. Although researchers have been developing these models for more complex tasks such as code generation and general reasoning, few efforts have explored how LLMs can be applied to combinatorial problems. In this research, we investigate the potential of LLMs to solve the Travelling Salesman Problem (TSP). Utilizing GPT-3.5 Turbo, we conducted experiments employing various approaches, including zero-shot in-context learning, few-shot in-context learning, and chain-of-thoughts (CoT). Consequently, we fine-tuned GPT-3.5 Turbo to solve a specific problem size and tested it using a set of various instance sizes. The fine-tuned models demonstrated promising performance on problems identical in size to the training instances and generalized well to larger problems. Furthermore, to improve the performance of the fine-tuned model without incurring additional training costs, we adopted a self-ensemble approach to improve the quality of the solutions.
With the increasing ability of large language models (LLMs), in-context learning (ICL) has evolved as a new paradigm for natural language processing (NLP), where instead of fine-tuning the parameters of an LLM specific to a downstream task with labeled examples, a small number of such examples is appended to a prompt instruction for controlling the decoder's generation process. ICL, thus, is conceptually similar to a non-parametric approach, such as $k$-NN, where the prediction for each instance essentially depends on the local topology, i.e., on a localised set of similar instances and their labels (called few-shot examples). This suggests that a test instance in ICL is analogous to a query in IR, and similar examples in ICL retrieved from a training set relate to a set of documents retrieved from a collection in IR. While standard unsupervised ranking models can be used to retrieve these few-shot examples from a training set, the effectiveness of the examples can potentially be improved by re-defining the notion of relevance specific to its utility for the downstream task, i.e., considering an example to be relevant if including it in the prompt instruction leads to a correct prediction. With this task-specific notion of relevance, it is possible to train a supervised ranking model (e.g., a bi-encoder or cross-encoder), which potentially learns to optimally select the few-shot examples. We believe that the recent advances in neural rankers can potentially find a use case for this task of optimally choosing examples for more effective downstream ICL predictions.
Conventional recommendation systems (RSs) are typically optimized to enhance performance metrics uniformly across all training samples. This makes it hard for data-driven RSs to cater to a diverse set of users due to the varying properties of these users. The performance disparity among various populations can harm the model's robustness with respect to sub-populations. While recent works have shown promising results in adapting large language models (LLMs) for recommendation to address hard samples, long user queries from millions of users can degrade the performance of LLMs and elevate costs, processing times and inference latency. This challenges the practical applicability of LLMs for recommendations. To address this, we propose a hybrid task allocation framework that utilizes the capabilities of both LLMs and traditional RSs. By adopting a two-phase approach to improve robustness to sub-populations, we promote a strategic assignment of tasks for efficient and responsible adaptation of LLMs. Our strategy works by first identifying the weak and inactive users that receive a suboptimal ranking performance by RSs. Next, we use an in-context learning approach for such users, wherein each user interaction history is contextualized as a distinct ranking task and given to an LLM. We test our hybrid framework by incorporating various recommendation algorithms -- collaborative filtering and learning-to-rank recommendation models -- and two LLMs -- both open and close-sourced. Our results on three real-world datasets show a significant reduction in weak users and improved robustness of RSs to sub-populations $(\approx12\%)$ and overall performance without disproportionately escalating costs.
As model context lengths continue to increase, the number of demonstrations that can be provided in-context approaches the size of entire training datasets. We study the behavior of in-context learning (ICL) at this extreme scale on multiple datasets and models. We show that, for many datasets with large label spaces, performance continues to increase with hundreds or thousands of demonstrations. We contrast this with example retrieval and finetuning: example retrieval shows excellent performance at low context lengths but has diminished gains with more demonstrations; finetuning is more data hungry than ICL but can sometimes exceed long-context ICL performance with additional data. We use this ICL setting as a testbed to study several properties of both in-context learning and long-context models. We show that long-context ICL is less sensitive to random input shuffling than short-context ICL, that grouping of same-label examples can negatively impact performance, and that the performance boosts we see do not arise from cumulative gain from encoding many examples together. We conclude that although long-context ICL can be surprisingly effective, most of this gain comes from attending back to similar examples rather than task learning.
This paper presents a comprehensive exploration of relation extraction utilizing advanced language models, specifically Chain of Thought (CoT) and Graphical Reasoning (GRE) techniques. We demonstrate how leveraging in-context learning with GPT-3.5 can significantly enhance the extraction process, particularly through detailed example-based reasoning. Additionally, we introduce a novel graphical reasoning approach that dissects relation extraction into sequential sub-tasks, improving precision and adaptability in processing complex relational data. Our experiments, conducted on multiple datasets, including manually annotated data, show considerable improvements in performance metrics, underscoring the effectiveness of our methodologies.