A striking property of transformers is their ability to perform in-context learning (ICL), a machine learning framework in which the learner is presented with a novel context during inference implicitly through some data, and tasked with making a prediction in that context. As such, that learner must adapt to the context without additional training. We explore the role of softmax attention in an ICL setting where each context encodes a regression task. We show that an attention unit learns a window that it uses to implement a nearest-neighbors predictor adapted to the landscape of the pretraining tasks. Specifically, we show that this window widens with decreasing Lipschitzness and increasing label noise in the pretraining tasks. We also show that on low-rank, linear problems, the attention unit learns to project onto the appropriate subspace before inference. Further, we show that this adaptivity relies crucially on the softmax activation and thus cannot be replicated by the linear activation often studied in prior theoretical analyses.
Large language models have the ability to generate text that mimics patterns in their inputs. We introduce a simple Markov Chain sequence modeling task in order to study how this in-context learning (ICL) capability emerges. In our setting, each example is sampled from a Markov chain drawn from a prior distribution over Markov chains. Transformers trained on this task form \emph{statistical induction heads} which compute accurate next-token probabilities given the bigram statistics of the context. During the course of training, models pass through multiple phases: after an initial stage in which predictions are uniform, they learn to sub-optimally predict using in-context single-token statistics (unigrams); then, there is a rapid phase transition to the correct in-context bigram solution. We conduct an empirical and theoretical investigation of this multi-phase process, showing how successful learning results from the interaction between the transformer's layers, and uncovering evidence that the presence of the simpler unigram solution may delay formation of the final bigram solution. We examine how learning is affected by varying the prior distribution over Markov chains, and consider the generalization of our in-context learning of Markov chains (ICL-MC) task to $n$-grams for $n > 2$.
In this paper, we address the challenge of detecting hateful memes in the low-resource setting where only a few labeled examples are available. Our approach leverages the compositionality of Low-rank adaptation (LoRA), a widely used parameter-efficient tuning technique. We commence by fine-tuning large language models (LLMs) with LoRA on selected tasks pertinent to hateful meme detection, thereby generating a suite of LoRA modules. These modules are capable of essential reasoning skills for hateful meme detection. We then use the few available annotated samples to train a module composer, which assigns weights to the LoRA modules based on their relevance. The model's learnable parameters are directly proportional to the number of LoRA modules. This modularized network, underpinned by LLMs and augmented with LoRA modules, exhibits enhanced generalization in the context of hateful meme detection. Our evaluation spans three datasets designed for hateful meme detection in a few-shot learning context. The proposed method demonstrates superior performance to traditional in-context learning, which is also more computationally intensive during inference.We then use the few available annotated samples to train a module composer, which assigns weights to the LoRA modules based on their relevance. The model's learnable parameters are directly proportional to the number of LoRA modules. This modularized network, underpinned by LLMs and augmented with LoRA modules, exhibits enhanced generalization in the context of hateful meme detection. Our evaluation spans three datasets designed for hateful meme detection in a few-shot learning context. The proposed method demonstrates superior performance to traditional in-context learning, which is also more computationally intensive during inference.
In Large Visual Language Models (LVLMs), the efficacy of In-Context Learning (ICL) remains limited by challenges in cross-modal interactions and representation disparities. To overcome these challenges, we introduce a novel Visual In-Context Learning (VICL) method comprising Visual Demonstration Retrieval, Intent-Oriented Image Summarization, and Intent-Oriented Demonstration Composition. Our approach retrieves images via ''Retrieval & Rerank'' paradigm, summarises images with task intent and task-specific visual parsing, and composes language-based demonstrations that reduce token count and alleviate cross-modal interaction problem. Experimental evaluations on five visual reasoning datasets demonstrate the effectiveness of our method. Moreover, our extensive experiments leverage information flow analysis to elucidate the effectiveness of our method, and investigate the impact of length and position of demonstrations for LVLM. The use of in-context unlearning further shows promise in resetting specific model knowledge without retraining.
Memory Editing (ME) has emerged as an efficient method to modify erroneous facts or inject new facts into Large Language Models (LLMs). Two mainstream ME methods exist: parameter-modifying ME and parameter-preserving ME (integrating extra modules while preserving original parameters). Regrettably, previous studies on ME evaluation have two critical limitations: (i) evaluating LLMs with single edit only, neglecting the need for continuous editing, and (ii) evaluations focusing solely on basic factual triples, overlooking broader LLM capabilities like logical reasoning and reading understanding. This study addresses these limitations with contributions threefold: (i) We explore how ME affects a wide range of fundamental capabilities of LLMs under sequential editing. Experimental results reveal an intriguing phenomenon: Most parameter-modifying ME consistently degrade performance across all tasks after a few sequential edits. In contrast, parameter-preserving ME effectively maintains LLMs' fundamental capabilities but struggles to accurately recall edited knowledge presented in a different format. (ii) We extend our evaluation to different editing settings, such as layers to edit, model size, instruction tuning, etc. Experimental findings indicate several strategies that can potentially mitigate the adverse effects of ME. (iii) We further explain why parameter-modifying ME damages LLMs from three dimensions: parameter changes after editing, language modeling capability, and the in-context learning capability. Our in-depth study advocates more careful use of ME in real-world scenarios.
Despite the strong performance of large language models (LLMs) across a wide range of tasks, they still have reliability issues. Previous studies indicate that strong LLMs like GPT-4-turbo excel in evaluating the reliability of responses from LLMs, but face efficiency and local deployment issues. Thus, to enable weak LLMs to effectively assess the reliability of LLM responses, we propose a novel cross-query-comparison-based method called $\textit{Meta Ranking}$ (MR). Unlike previous few-shot methods that solely based on in-context learning capabilities in LLMs, MR assesses reliability by pairwisely ranking the target query-response pair with multiple reference query-response pairs. We found that MR is highly effective in error detection for LLM responses, where weak LLMs, such as Phi-2, could surpass strong baselines like GPT-3.5-turbo, requiring only five reference samples and significantly improving efficiency. We further demonstrate that MR can enhance strong LLMs' performance in two practical applications: model cascading and instruction tuning. In model cascading, we combine open- and closed-source LLMs to achieve performance comparable to GPT-4-turbo with lower costs. In instruction tuning, we use MR for iterative training data filtering, significantly reducing data processing time and enabling LLaMA-7B and Phi-2 to surpass Alpaca-13B with fewer training tokens. These results underscore the high potential of MR in both efficiency and effectiveness.
Over the past few years, the abilities of large language models (LLMs) have received extensive attention, which have performed exceptionally well in complicated scenarios such as logical reasoning and symbolic inference. A significant factor contributing to this progress is the benefit of in-context learning and few-shot prompting. However, the reasons behind the success of such models using contextual reasoning have not been fully explored. Do LLMs have understand logical rules to draw inferences, or do they ``guess'' the answers by learning a type of probabilistic mapping through context? This paper investigates the reasoning capabilities of LLMs on two logical reasoning datasets by using counterfactual methods to replace context text and modify logical concepts. Based on our analysis, it is found that LLMs do not truly understand logical rules; rather, in-context learning has simply enhanced the likelihood of these models arriving at the correct answers. If one alters certain words in the context text or changes the concepts of logical terms, the outputs of LLMs can be significantly disrupted, leading to counter-intuitive responses. This work provides critical insights into the limitations of LLMs, underscoring the need for more robust mechanisms to ensure reliable logical reasoning in LLMs.
Large Language Models (LLMs) exhibit strong In-Context Learning (ICL) capabilities when prompts with demonstrations are used. However, fine-tuning still remains crucial to further enhance their adaptability. Prompt-based fine-tuning proves to be an effective fine-tuning method in low-data scenarios, but high demands on computing resources limit its practicality. We address this issue by introducing a prompt-based parameter-efficient fine-tuning (PEFT) approach. GNNavi leverages insights into ICL's information flow dynamics, which indicates that label words act in prompts as anchors for information propagation. GNNavi employs a Graph Neural Network (GNN) layer to precisely guide the aggregation and distribution of information flow during the processing of prompts by hardwiring the desired information flow into the GNN. Our experiments on text classification tasks with GPT-2 and Llama2 show GNNavi surpasses standard prompt-based fine-tuning methods in few-shot settings by updating just 0.2% to 0.5% of parameters. We compare GNNavi with prevalent PEFT approaches, such as prefix tuning, LoRA and Adapter in terms of performance and efficiency. Our analysis reveals that GNNavi enhances information flow and ensures a clear aggregation process.
In-context learning of large-language models (LLMs) has achieved remarkable success in the field of natural language processing, while extensive case studies reveal that the single-step chain-of-thought prompting approach faces challenges such as attention diffusion and inadequate performance in complex tasks like text-to-SQL. To improve the contextual learning capabilities of LLMs in text-to-SQL, a workflow paradigm method is proposed, aiming to enhance the attention and problem-solving scope of LLMs through decomposition. Specifically, the information determination module for eliminating redundant information and the brand-new prompt structure based on problem classification greatly enhance the model's attention. Additionally, the inclusion of self-correction and active learning modules greatly expands the problem-solving scope of LLMs, hence improving the upper limit of LLM-based approaches. Extensive experiments conducted on three datasets demonstrate that our approach outperforms other methods by a significant margin. About 2-3 percentage point improvements compared to the existing baseline on the Spider Dev, Spider-Realistic, and Bird Dev datasets and new SOTA results on the Spider Test dataset are achieved. Our code is available on GitHub: \url{https://github.com/FlyingFeather/DEA-SQL}.
Currently, the in-context learning method based on large language models (LLMs) has become the mainstream of text-to-SQL research. Previous works have discussed how to select demonstrations related to the user question from a human-labeled demonstration pool. However, human labeling suffers from the limitations of insufficient diversity and high labeling overhead. Therefore, in this paper, we discuss how to measure and improve the diversity of the demonstrations for text-to-SQL. We present a metric to measure the diversity of the demonstrations and analyze the insufficient of the existing labeled data by experiments. Based on the above discovery, we propose fusing iteratively for demonstrations (Fused) to build a high-diversity demonstration pool through human-free multiple-iteration synthesis, improving diversity and lowering label cost. Our method achieves an average improvement of 3.2% and 5.0% with and without human labeling on several mainstream datasets, which proves the effectiveness of Fused.
Named Entity Recognition (NER) serves as a fundamental task in natural language understanding, bearing direct implications for web content analysis, search engines, and information retrieval systems. Fine-tuned NER models exhibit satisfactory performance on standard NER benchmarks. However, due to limited fine-tuning data and lack of knowledge, it performs poorly on unseen entity recognition. As a result, the usability and reliability of NER models in web-related applications are compromised. Instead, Large Language Models (LLMs) like GPT-4 possess extensive external knowledge, but research indicates that they lack specialty for NER tasks. Furthermore, non-public and large-scale weights make tuning LLMs difficult. To address these challenges, we propose a framework that combines small fine-tuned models with LLMs (LinkNER) and an uncertainty-based linking strategy called RDC that enables fine-tuned models to complement black-box LLMs, achieving better performance. We experiment with both standard NER test sets and noisy social media datasets. LinkNER enhances NER task performance, notably surpassing SOTA models in robustness tests. We also quantitatively analyze the influence of key components like uncertainty estimation methods, LLMs, and in-context learning on diverse NER tasks, offering specific web-related recommendations. Code is available at https://github.com/zhzhengit/LinkNER.
Demonstration ordering, which is an important strategy for in-context learning (ICL), can significantly affects the performance of large language models (LLMs). However, most of the current approaches of ordering require high computational costs to introduce the priori knowledge. In this paper, inspired by the human learning process, we propose a simple but effective demonstration ordering method for ICL, named the few-shot In-Context Curriculum Learning (ICCL). The ICCL implies gradually increasing the complexity of prompt demonstrations during the inference process. The difficulty can be assessed by human experts or LLMs-driven metrics, such as perplexity. Then we design extensive experiments to discuss the effectiveness of the ICCL at both corpus-level and instance-level. Moreover, we also investigate the formation mechanism of LLM's ICCL capability. Experimental results demonstrate that ICCL, developed during the instruction-tuning stage, is effective for representative open-source LLMs. To facilitate further research and applications by other scholars, we make the code publicly available.
Advancing the frontier of subquadratic architectures for Language Models (LMs) is crucial in the rapidly evolving field of natural language processing. Current innovations, including State Space Models, were initially celebrated for surpassing Transformer performance on language modeling tasks. However, these models have revealed deficiencies in essential In-Context Learning capabilities - a domain where the Transformer traditionally shines. The Based model emerged as a hybrid solution, blending a Linear Transformer with a kernel inspired by the Taylor expansion of exponential functions, augmented by convolutional networks. Mirroring the Transformer's in-context adeptness, it became a strong contender in the field. In our work, we present a singular, elegant alteration to the Based kernel that amplifies its In-Context Learning abilities evaluated with the Multi-Query Associative Recall task and overall language modeling process, as demonstrated on the Pile dataset.
Prompt-based learning is susceptible to intrinsic bias present in pre-trained language models (LMs), leading to sub-optimal performance in prompt-based zero/few-shot settings. In this work, we propose a null-input prompting method to calibrate intrinsic bias encoded in pre-trained LMs. Different from prior efforts that address intrinsic bias primarily for social fairness and often involve excessive computational cost, our objective is to explore enhancing LMs' performance in downstream zero/few-shot learning while emphasizing the efficiency of intrinsic bias calibration. Specifically, we leverage a diverse set of auto-selected null-meaning inputs generated from GPT-4 to probe intrinsic bias of pre-trained LMs. Utilizing the bias-reflected probability distribution, we formulate a distribution disparity loss for bias calibration, where we exclusively update bias parameters ($0.1\%$ of total parameters) of LMs towards equal probability distribution. Experimental results show that the calibration promotes an equitable starting point for LMs while preserving language modeling abilities. Across a wide range of datasets, including sentiment analysis and topic classification, our method significantly improves zero/few-shot learning performance of LMs for both in-context learning and prompt-based fine-tuning (on average $9\%$ and $2\%$, respectively).
We need to trust robots that use often opaque AI methods. They need to explain themselves to us, and we need to trust their explanation. In this regard, explainability plays a critical role in trustworthy autonomous decision-making to foster transparency and acceptance among end users, especially in complex autonomous driving. Recent advancements in Multi-Modal Large Language models (MLLMs) have shown promising potential in enhancing the explainability as a driving agent by producing control predictions along with natural language explanations. However, severe data scarcity due to expensive annotation costs and significant domain gaps between different datasets makes the development of a robust and generalisable system an extremely challenging task. Moreover, the prohibitively expensive training requirements of MLLM and the unsolved problem of catastrophic forgetting further limit their generalisability post-deployment. To address these challenges, we present RAG-Driver, a novel retrieval-augmented multi-modal large language model that leverages in-context learning for high-performance, explainable, and generalisable autonomous driving. By grounding in retrieved expert demonstration, we empirically validate that RAG-Driver achieves state-of-the-art performance in producing driving action explanations, justifications, and control signal prediction. More importantly, it exhibits exceptional zero-shot generalisation capabilities to unseen environments without further training endeavours.
Many existing theoretical analyses of in-context learning for natural language processing are based on latent variable models that leaves gaps between theory and practice. We aim to close these gaps by proposing a theoretical framework, the Pelican Soup Framework. In this framework, we introduce (1) the notion of a common sense knowledge base, (2) a general formalism for natural language classification tasks, and the notion of (3) meaning association. Under this framework, we can establish a $\mathcal{O}(1/T)$ loss bound for in-context learning, where $T$ is the number of example-label pairs in the demonstration. Compared with previous works, our bound reflects the effect of the choice of verbalizers and the effect of instruction tuning. An additional notion of \textit{atom concepts} makes our framework possible to explain the generalization to tasks unseen in the language model training data. Finally, we propose a toy setup, Calcutec, and a digit addition task that mimics types of distribution shifts a model needs to overcome to perform in-context learning. We also experiment with GPT2-Large on real-world NLP tasks. Our empirical results demonstrate the efficacy of our framework to explain in-context learning.
LLM hallucination, i.e. generating factually incorrect yet seemingly convincing answers, is currently a major threat to the trustworthiness and reliability of LLMs. The first step towards solving this complicated problem is to measure it. However, existing hallucination metrics require having a benchmark dataset with gold-standard answers, i.e. "best" or "correct" answers written by humans. Such requirements make hallucination measurement costly and prone to human errors. In this work, we propose Factualness Evaluations via Weighting LLMs (FEWL), an innovative hallucination metric that is specifically designed for the scenario when gold-standard answers are absent. FEWL leverages the answers from off-the-shelf LLMs that serve as a proxy of gold-standard answers. The key challenge is how to quantify the expertise of reference LLMs resourcefully. We show FEWL has certain theoretical guarantees and demonstrate empirically it gives more accurate hallucination measures than naively using reference LLMs. We also show how to leverage FEWL to reduce hallucination through both in-context learning and supervised fine-tuning. Extensive experiment results on Truthful-QA, CHALE, and HaluEval datasets demonstrate the effectiveness of FEWL.
Previous in-context learning (ICL) research has focused on tasks such as classification, machine translation, text2table, etc., while studies on whether ICL can improve human-like dialogue generation are scarce. Our work fills this gap by systematically investigating the ICL capabilities of large language models (LLMs) in persona-based dialogue generation, conducting extensive experiments on high-quality real human Chinese dialogue datasets. From experimental results, we draw three conclusions: 1) adjusting prompt instructions is the most direct, effective, and economical way to improve generation quality; 2) randomly retrieving demonstrations (demos) achieves the best results, possibly due to the greater diversity and the amount of effective information; counter-intuitively, retrieving demos with a context identical to the query performs the worst; 3) even when we destroy the multi-turn associations and single-turn semantics in the demos, increasing the number of demos still improves dialogue performance, proving that LLMs can learn from corrupted dialogue demos. Previous explanations of the ICL mechanism, such as $n$-gram induction head, cannot fully account for this phenomenon.
Selecting high-quality pre-training data is important for creating capable language models, but existing methods rely on simple heuristics. We introduce QuRating, a method for selecting pre-training data that can capture human intuitions about data quality. In this paper, we investigate four qualities - writing style, required expertise, facts & trivia, and educational value - and find that LLMs are able to discern these qualities, especially when making pairwise judgments of texts. We train a QuRater model to learn scalar ratings from pairwise judgments, and use it to annotate a 260B training corpus with quality ratings for each of the four criteria. In our experiments, we select 30B tokens according to the different quality ratings and train 1.3B-parameter language models on the selected data. We find that it is important to balance quality and diversity. When we sample using quality ratings as logits over documents, our models obtain lower perplexity and stronger in-context learning performance than baselines. Our best model is based on educational value and performs similarly to a model trained with uniform sampling for 50% more steps. Beyond data selection, we use the quality ratings to construct a training curriculum which improves performance without changing the training dataset. We extensively analyze the quality ratings and discuss their characteristics, biases, and wider implications.
Associative memory and probabilistic modeling are two fundamental topics in artificial intelligence. The first studies recurrent neural networks designed to denoise, complete and retrieve data, whereas the second studies learning and sampling from probability distributions. Based on the observation that associative memory's energy functions can be seen as probabilistic modeling's negative log likelihoods, we build a bridge between the two that enables useful flow of ideas in both directions. We showcase four examples: First, we propose new energy-based models that flexibly adapt their energy functions to new in-context datasets, an approach we term \textit{in-context learning of energy functions}. Second, we propose two new associative memory models: one that dynamically creates new memories as necessitated by the training data using Bayesian nonparametrics, and another that explicitly computes proportional memory assignments using the evidence lower bound. Third, using tools from associative memory, we analytically and numerically characterize the memory capacity of Gaussian kernel density estimators, a widespread tool in probababilistic modeling. Fourth, we study a widespread implementation choice in transformers -- normalization followed by self attention -- to show it performs clustering on the hypersphere. Altogether, this work urges further exchange of useful ideas between these two continents of artificial intelligence.